test(signing): align key storage contract with Gitea

The optional XBPS publisher signs repository metadata after package signing. Remove the runner key only after publication and release asset upload.

.gitea/workflows/release.yml

@@ -159,7 +159,6 @@ jobs:
             gpg --batch --yes --delete-secret-keys "${FINGERPRINT}"
             gpg --batch --yes --delete-keys "${FINGERPRINT}"
           fi
-          rm -f ~/.ssh/id_xbps
 
       - name: Determine version
         id: version
@@ -327,9 +326,11 @@ jobs:
           # Add XBPS artifacts if they exist
           XBPS_FILE="fenris-${VERSION}_1.x86_64.xbps"
           if [ -f "${XBPS_FILE}" ]; then
-            ARTIFACTS+=("${XBPS_FILE}")
             if [ -f "${XBPS_FILE}.sig2" ]; then
+              ARTIFACTS+=("${XBPS_FILE}")
               ARTIFACTS+=("${XBPS_FILE}.sig2")
+            else
+              echo "::warning::Unsigned XBPS artifact omitted from release assets"
             fi
           fi
           for FILE in "${ARTIFACTS[@]}"; do
@@ -343,3 +344,7 @@ jobs:
                 "https://git.bongbetic.com/api/v1/repos/xavierk/Fenris/releases/${RELEASE_ID}/assets"
             fi
           done
+
+      - name: Remove XBPS signing key
+        if: always()
+        run: rm -f ~/.ssh/id_xbps

AGENTS.md

@@ -1,3 +1,7 @@
+## Session style
+
+After the first user message in each session, load the global `caveman` skill and activate `ultra` mode. Keep `ultra` mode active until the user changes or stops it.
+
 ## Agent skills
 
 ## Commit messages

CHANGELOG.md

@@ -9,6 +9,19 @@ backfill releases from before this changelog.
 
 ## [Unreleased]
 
+## [0.6.0] - 2026-09-29
+
+### Added
+
+- Preserve local-day read and write evidence separately, including ambiguous midnight-spanning volume as shared evidence instead of assigning it to either day.
+- Keep valid observations pending when publication fails, and recover them without exposing partial history or counting activity twice.
+- Preserve live-point inspection and historical activity selection through graph refresh.
+
+### Changed
+
+- Prune old detail only after durable summaries and required boundary evidence are published.
+- Require a complete local day to fit within one monitoring period before it can open the endurance outlook.
+
 ## [0.5.0] - 2026-09-19
 
 ### Changed

CONTEXT.md

@@ -28,6 +28,10 @@ _Avoid_: Daemon uptime, calibration window
 The single SQLite database at `/var/lib/fenris/observations.db` that persists the observation history, monitoring periods, hour observations, day aggregates, and endurance baseline.
 _Avoid_: Data directory, history.jsonl, the database (generic)
 
+**Pending publication**:
+The condition where valid acquired observations are retained for recovery but their dependent evidence has not yet been published consistently. Those observations are not part of the reader-visible observation history until publication succeeds; readers retain the last consistent evidence with the pending condition made explicit.
+_Avoid_: Successful collection, fresh published evidence
+
 **Store fault**:
 The condition where the observation store is present but cannot be read or trusted — unreadable, corrupt, or written by a newer Fenris — degrading every view that depends on it rather than crashing or guessing.
 _Avoid_: Database error, corruption, broken data
@@ -40,6 +44,18 @@ _Avoid_: Hourly record, hourly.jsonl entry
 One row per UTC day derived from hour observations; the grain at which usage-habit evidence is judged.
 _Avoid_: Daily summary, daily stats
 
+**Local-day evidence**:
+Measured read and write activity attributable to a local calendar day with its recorded timezone and midnight boundaries. Known volumes remain incomplete when gaps or shared local-day evidence prevent an exact total; they are not estimates of the missing activity.
+_Avoid_: Estimated daily total, localised UTC day aggregate
+
+**Shared local-day evidence**:
+Measured read or write volume spanning a local midnight that cannot honestly be allocated to either adjacent day. It is retained once, separately from either day's known volume, rather than prorated or counted in both days.
+_Avoid_: Missing bytes, estimated midnight split
+
+**Activity selection**:
+The read/write measurement, date and point being inspected in live or historical drive activity. Live activity follows the newest point until deliberate inspection pins a point; historical selection survives refresh, while an expired live pin is explained before following resumes.
+_Avoid_: Projection window, current usage habit
+
 **Usage-history window**:
 An exact consecutive span of UTC calendar days ending today, shown from day aggregates; a day without trustworthy evidence remains an explicit gap rather than disappearing or being estimated.
 _Avoid_: Available records, dataset range

README.md

@@ -22,7 +22,7 @@ The Gitea instance Debian registry signs metadata with its own key. Verify the
 instance key fingerprint (TOFU hardening):
 
 ```text
-Fingerprint: <print after first release — paste beside the curl one-liner>
+Fingerprint: F937E81D2FB0736B15BC611884BEBD586DFAC010
 ```
 
 Add the instance key and repository:
@@ -263,6 +263,12 @@ measured read/write volumes, not transfer speed; missing evidence breaks the
 trace. `?` marks a gap, `~` a partial total, and `u` unallocated daily volume.
 Use the selected-point readout for exact values and evidence state.
 
+Local-day totals use measured intervals inside each recorded local date. An
+interval that crosses local midnight appears once as shared evidence and stays
+outside both known totals. The dashboard labels current totals as so far and
+shows incomplete or unavailable dates without treating them as zero. Older
+UTC-only summaries cannot establish exact local-day totals.
+
 - `v` cycles Live / Day / History; the tabs are also clickable.
 - `←` / `→` inspect points; `w` switches read/write volume in every view.
 - `[` / `]` browse dates, `g` enters a date, and `t` returns to today/live.

docs/adr/0010-local-day-activity-history.md

@@ -0,0 +1,11 @@
+# 10. Preserve local-day activity history with its recorded timezone
+
+Status: Accepted; legacy-history migration and repair implemented in issue #99.
+
+Fenris will retain local-day read/write summaries and the boundary evidence needed to interpret them in the existing observation store before three-minute detail expires after 14 days. Each historical summary retains its recorded timezone and day boundaries; this extends [ADR 0001](0001-observation-store-sqlite.md) while preserving the UTC hour/day evidence used for endurance projections. Collection derives new measured intervals. Ordered schema migration and explicit repair rebuild legacy summaries from surviving evidence; TUI and CLI readers remain read-only.
+
+UTC hourly summaries alone cannot recover a local day whose midnight falls inside a UTC hour. Preserving labelled local summaries trades arbitrary future timezone reinterpretation for bounded detailed-history retention; retaining all fine-grained intervals indefinitely or estimating a split from UTC summaries would violate the agreed retention or evidence semantics.
+
+Only surviving evidence may be used to derive older local summaries. Dates without enough evidence remain incomplete or unavailable; a measured interval crossing midnight is retained once as shared boundary evidence, never prorated or counted in full on both days. A later timezone change does not silently rewrite historical day boundaries.
+
+The agreed presentation and validation requirements are in the [live drive activity specification](../spec/live-drive-activity.md).

docs/adr/0011-preserve-unpublished-observations.md

@@ -0,0 +1,20 @@
+# 11. Preserve valid unpublished observations without exposing partial history
+
+Status: Accepted; implemented on `main` for the planned v0.6.0 release (commit `e22b994`).
+
+A non-invariant derivation failure must not discard valid acquired observations or report a successful collection: retain the observations as pending publication in the existing observation store and retry through the normal scheduled collection path. Readers continue to see the last consistent published evidence, with an explicit pending-publication explanation, rather than combining newly acquired counters with older derived totals. This trades recovery bookkeeping for preservation of measured evidence and consistent read views; neither discarding every valid acquisition on derivation failure nor exposing partially derived history satisfies both requirements.
+
+## Constraints
+
+- The publication distinction applies to every dependent reader, including activity, freshness, controller-segment interpretation and projection inputs, not just the graph. ADR 0001's freshest-sample signal means the freshest published sample; an unpublished observation must not make published evidence appear fresh.
+- Pending-work bookkeeping qualifies ADR 0005's recovery-without-new-state wording: it records unfinished evidence publication, not a new health grade or escalation mechanism. Last-collection outcomes still come from the existing native monitoring and logging paths.
+- Invariant violations retain [ADR 0005](0005-failure-detection-and-recovery.md)'s write-nothing rule; retaining recoverable work is not permission to persist invalid observations. Actual store faults retain their existing refusal and degradation behavior.
+- Recovery and retention remain collector-owned. Repeated recovery must not duplicate measured volume, and required source evidence cannot be pruned before trustworthy derived evidence is durable.
+- This extends [ADR 0001](0001-observation-store-sqlite.md)'s consistent read model and [ADR 0010](0010-local-day-activity-history.md)'s preservation rule without another observation store, sampler, background process or retry cadence. Pending-work metadata is not a persisted projection or a replacement for journalled collection outcomes.
+- Pending observations use private staging separate from published samples and derived evidence inside the same observation store. This makes exclusion from ordinary reader and projection queries structural, rather than requiring each query to remember a publication filter; the exact table layout remains an implementation detail.
+
+## Bounded pending work
+
+Pending publication has a fixed initial admission capacity of 6,720 observations, equivalent to 14 days at the default three-minute cadence. This is a count limit, not an expiry rule: older pending observations are never discarded merely to admit newer ones.
+
+The collection module attempts recovery and checks capacity before invoking acquisition. If capacity remains exhausted, the collection is unsuccessful and visibly explains why no new observation was acquired; normal scheduled recovery attempts continue. This deliberately accepts a gap in new observations rather than unlimited pending growth or loss of already retained evidence. It is not a deliberate disable, changes no monitoring intent, and never permits fabricated activity across the gap.

docs/install/signing-key-ceremony.md

@@ -12,7 +12,7 @@ and destruction.
 | UID | `Fenris Packaging <packaging@bongbetic.com>` |
 | Expiry | 2 years from creation |
 | Hierarchy | Single key — no master/subkey split (single maintainer, manual builds) |
-| Private key storage | Password manager only |
+| Private key storage | Gitea repository Actions secret `GPG_PRIVATE_KEY` |
 | Public key storage | `packaging/keys/fenris-packaging.asc` in-repo, release notes, docs |
 | Keyservers | Never — TOFU-over-TLS via raw URL |
 
@@ -37,71 +37,72 @@ gpg --armor --export packaging@bongbetic.com > packaging/keys/fenris-packaging.a
 gpg --fingerprint packaging@bongbetic.com
 ```
 
-Save the **private key** to the password manager immediately:
+Provision the **private key** as the repository Actions secret `GPG_PRIVATE_KEY`.
+Run the export on the trusted key-generation machine, then enter its output in
+the Gitea repository's Actions secret settings. Do not save it in the checkout,
+logs, or a runner directory. The release workflow checks its fingerprint
+against the committed public key before signing.
 
 ```bash
 gpg --armor --export-secret-keys packaging@bongbetic.com
 ```
 
-Then **delete the private key from the local keyring** — it must never persist
-on any build host:
+After provisioning the secret, delete the private key from the key-generation
+keyring:
 
 ```bash
-gpg --delete-secret-keys packaging@bongbetic.com
-gpg --delete-keys packaging@bongbetic.com
+gpg --batch --yes --delete-secret-keys packaging@bongbetic.com
+gpg --batch --yes --delete-keys packaging@bongbetic.com
 ```
 
 The committed `fenris-packaging.asc` must contain the real public key (replace
 the placeholder comments).
 
+## XBPS signing key
+
+XBPS uses a separate RSA 3072 key. Its private half is stored as the Gitea
+repository Actions secret `XBPS_SIGNING_KEY`. The corresponding public key is
+published at
+`https://git.bongbetic.com/xavierk/Fenris-xbps/raw/branch/stable/keys/fenris-xbps-signing.pub`,
+with fingerprint `SHA256:AvPMRlKMikPg75u0iKr8AUkxlfU/Ad4k/S4o2M9W4/w`.
+The secret must match that public key.
+
+The release workflow writes the key to `~/.ssh/id_xbps` to sign the XBPS
+package. A requested XBPS publication also uses the key to sign repository
+metadata. A final `always()` cleanup removes the runner copy after publication
+and release asset upload, including when an earlier step fails.
+
 ## Per-release signing flow
 
-Each release performs: **import → sign → delete**. The private key is never
-stored on disk longer than the release takes.
+Each tagged release performs: **import → verify → sign → delete** on the
+repository-scoped Gitea Actions runner. The Gitea secret remains configured;
+the runner's keyring copy is removed after the job.
 
-### Step 1: Import the private key
+### Step 1: Push the release tag
 
-Retrieve the private key from the password manager and import it:
+After updating the version and dated changelog section, push the matching tag:
 
 ```bash
-gpg --import /tmp/packaging-key-private.asc
-rm /f /tmp/packaging-key-private.asc   # Shred if possible
+git push origin v<version>
 ```
 
-### Step 2: Build and sign packages
+### Step 2: Build, verify, and sign packages
 
-The Makefile target `make release` handles signing automatically when the
-key is in the keyring:
+The release workflow imports `GPG_PRIVATE_KEY`, checks it against
+`packaging/keys/fenris-packaging.asc`, builds packages, signs the RPM and
+clearsigned checksum manifest, validates both, and publishes the release. The
+workflow imports `XBPS_SIGNING_KEY` separately and signs the XBPS package.
+XBPS publication is optional and also signs repository metadata; it requires
+host acceptance and explicit selection during workflow dispatch.
 
-```bash
-make release    # builds, signs RPM, clearsigns SHA256SUMS, prints upload steps
-```
+### Step 3: Verify runner cleanup
 
-Under the hood:
+The workflow's `always()` cleanup removes the GPG key from the runner's keyring
+and deletes `~/.ssh/id_xbps`, including after a failed job. Confirm no signing
+key remains on the runner after the release job.
 
-1. `rpmsign --addsign` signs the RPM payload with the packaging key
-   (invoked by `make sign-rpm`).
-2. `sha256sum` generates the checksum manifest.
-3. `gpg --clearsign` produces `SHA256SUMS.asc` with the packaging key.
-
-### Step 3: Delete the private key
-
-Immediately after signing:
-
-```bash
-gpg --delete-secret-keys packaging@bongbetic.com
-gpg --delete-keys packaging@bongbetic.com
-```
-
-Verify the key is gone:
-
-```bash
-gpg --list-keys packaging@bongbetic.com
-# Should produce: gpg: keyblock resource ...: No such file or directory
-```
-
-The entire import → sign → delete cycle should take minutes. The private key
-must never be left in any keyring between releases.
+The Gitea Actions secret remains the approved signing source. Do not copy it to
+the runner or repository outside the workflow.
 
 ## Key rotation (outline)
 

docs/spec/live-drive-activity.md

@@ -0,0 +1,117 @@
+## Problem Statement
+
+Fenris users cannot readily understand how much data their monitored drive reads and writes each day, inspect recent activity as it arrives, or select historical dates from the keyboard. Existing graph shortcuts depend on focus, date navigation is limited, collection and screen refresh default to five minutes, and the graph reads hourly/daily summaries rather than three-minute activity. Source inspection also identified derivation and rendering paths that can leave displayed history incomplete or stale despite newly stored readings.
+
+Users want daily data volume as the primary measurement because writes contribute to drive endurance. They also want an understandable end-of-life outlook even with limited history, without presenting an unsupported hardware-failure prediction or a misleading numerical confidence score.
+
+## Solution
+
+Show both local-day read and write totals, updating from background collection targeted every three minutes. Open on a live last-three-hours graph of written data volume, provide a read/write toggle, and support selected-day and longer-history views. Make date navigation discoverable through visible keyboard hints and clickable equivalents.
+
+Retain three-minute detail for 14 days, followed by durable hourly/daily summaries. Preserve the timezone and boundaries of historical local-day summaries, identify incomplete evidence, and never invent missing activity or divide midnight-spanning measurements by assumption.
+
+Show the usage-adjusted theoretical lifespan after one full local calendar day of observations, provided an applicable endurance baseline and usable write rate exist. Explain that this estimates remaining write endurance if observed habits continue, not a physical failure date. Present categorical projection confidence with contributing facts, including limited-history reasons.
+
+## User Stories
+
+1. As a drive owner, I want to see the amount of data written each local day, so that I can understand the activity that contributes to write endurance.
+2. As a drive owner, I want to see the amount of data read each local day, so that I can understand the drive's broader activity.
+3. As a drive owner, I want read and write totals shown separately, so that reads are not mistaken for writes consuming the endurance allowance.
+4. As a TUI user, I want data-volume units rather than transfer speed as the primary graph measurement, so that the graph answers how much data was transferred.
+5. As a TUI user, I want today's totals labelled as totals so far, so that a partial day is not presented as a completed day's usage.
+6. As a TUI user, I want the opening graph to show written volume over the last three hours, so that recent activity is visible immediately.
+7. As a TUI user, I want to switch the graph between writes and reads while both daily totals remain visible, so that I can inspect either measurement without losing context.
+8. As a TUI user, I want each new activity point to include transfers between readings, so that brief bursts between collection runs contribute to the graph.
+9. As a TUI user, I want new measurements approximately every three minutes, so that the recent-activity view stays useful while I work.
+10. As a user who closes the TUI, I want collection to continue in the background, so that reopening it shows activity gathered while I was away.
+11. As a user who reboots, I want observation history and the established monitoring lifecycle preserved, so that restarting does not erase activity or confuse monitoring state.
+12. As a keyboard user, I want `[` and `]` to select the previous and next day, so that browsing nearby dates takes one action.
+13. As a keyboard user, I want `g` to open a date field, so that I can jump directly to a historical date.
+14. As a keyboard user, I want `t` to return to today and the live view, so that I can leave historical browsing immediately.
+15. As a keyboard user, I want arrows to inspect graph points, so that I can read the volume, time, and evidence state behind a point.
+16. As a new TUI user, I want visible shortcut hints that work from normal launch, so that I do not have to discover an invisible graph-focus prerequisite.
+17. As a mouse user, I want clickable equivalents for navigation and the read/write toggle, so that I can use the same features without memorizing shortcuts.
+18. As a user entering a date, I want typing and cancelling to remain within the date-entry workflow, so that graph or monitoring shortcuts do not fire accidentally.
+19. As a user browsing history, I want background refresh to preserve my selected date and measurement, so that new activity does not interrupt inspection.
+20. As a user inspecting daily totals, I want to switch between recent detail, a selected day, and longer history, so that I can understand both short bursts and daily habits.
+21. As a user outside UTC, I want calendar navigation and timestamps expressed in labelled local time, so that the displayed day matches my calendar.
+22. As a user in a timezone with a half-hour offset, I want local-day totals based on the correct midnight boundary, so that relabelled UTC totals do not misrepresent my day.
+23. As a user experiencing a daylight-saving transition, I want real local-day boundaries respected, so that a 23-hour or 25-hour day remains understandable.
+24. As a user who changes system timezone, I want historical summaries to retain their recorded timezone and boundaries, so that old totals do not silently change meaning.
+25. As a user reviewing recent history, I want three-minute detail available for 14 days, so that I can investigate recent usage.
+26. As a user reviewing older history, I want durable hourly/daily summaries after detailed readings expire, so that long-term activity remains available.
+27. As a user with legacy history, I want dates that cannot be reconstructed at local-day precision labelled incomplete or unavailable, so that old summaries are not presented with invented precision.
+28. As a user whose readings straddle midnight, I want the measured volume preserved once and its uncertain day allocation explained, so that it is neither lost nor counted twice.
+29. As a new user with only one reading, I want an awaiting-another-reading state, so that a missing interval is not shown as zero activity.
+30. As a user with a measured zero-activity interval, I want zero distinguished from missing evidence, so that a quiet drive is not confused with a collection failure.
+31. As a user with missed collection runs, I want gaps, actual timestamps, and freshness facts, so that the graph does not imply measurements that never occurred.
+32. As a user who deliberately pauses monitoring, I want paused time excluded according to monitoring-period rules, so that the pause does not distort the observed usage habit.
+33. As a user whose controller resets or changes, I want counter and identity boundaries respected, so that unrelated readings do not create invalid activity or lifespan estimates.
+34. As a user with a store fault, I want a clear explanation instead of guessed history, so that I know which information cannot be trusted.
+35. As a drive owner, I want an estimate of the time until remaining write endurance is consumed, so that I can plan around my observed usage habit.
+36. As a new user, I want that estimate withheld until one full local calendar day has been observed, so that a few minutes of activity do not immediately produce a lifespan number.
+37. As a user who starts monitoring at noon, I want the partial first day excluded from the full-day gate, so that 24 hours since launch is not mistaken for a complete calendar day.
+38. As a user with one complete day but limited history, I want an estimate when the other inputs support it, so that I do not have to wait for high confidence before seeing an outlook.
+39. As a user with limited evidence, I want a confidence category and concrete reasons, so that I understand why an estimate may change without being shown an unjustified percentage.
+40. As a user without an applicable endurance baseline or usable write rate, I want the missing prerequisite explained, so that unavailable evidence is not disguised as merely low confidence.
+41. As a drive owner, I want the endurance estimate distinguished from a physical failure date, so that I do not treat a write-endurance projection as a hardware guarantee.
+42. As a CLI user, I want status and the TUI to agree on forecast eligibility and confidence, so that choosing a different interface does not change the facts.
+43. As a user on a narrow terminal, I want daily totals, dates, evidence labels, and controls to remain accessible, so that terminal size does not hide the information I need.
+44. As a user of existing themes and reduced-motion settings, I want those features preserved when navigation shortcuts change, so that better date controls do not remove accessibility or preferences.
+45. As a user upgrading Fenris, I want migration, repair, refresh, and pruning to preserve measured evidence without duplication, so that the new graph can be trusted across restarts and upgrades.
+
+## Implementation Decisions
+
+- Extend the existing collector, observation store, derivation, shared status/projection, service scheduling, and TUI responsibilities. The privileged collector remains the single device-acquisition and store-write owner; the TUI remains an unprivileged reader. Do not add another sampler, service, export layer, or generic plotting abstraction.
+- Set the default background collection target to three minutes on both systemd and runit. Update shared cadence/freshness semantics, scheduler definitions, user documentation, and relevant acceptance contracts together. Use actual observation timestamps: delayed runs must not be represented as perfectly spaced measurements or synthetic catch-up points.
+- Derive interval read/write volumes from compatible cumulative counters. A first reading is an anchor, not an interval. Preserve controller-segment and monitoring-period boundaries, measured zero, unknown time, deliberate disables, and unsupported evidence. Do not compute deltas across a controller identity or counter discontinuity.
+- Make successful collection publish consistent data for dependent read views, using the existing transactional and migration patterns. Ensure raw samples, affected summaries, and retained boundary evidence cannot disagree through partial publication. Validate source-level derivation gaps before implementing fixes; this spec does not claim a runtime diagnosis.
+- Keep daily read and write totals separate. The write graph is the default; a read/write toggle changes only the displayed measurement. Each live point represents measured interval volume, not instantaneous speed. Daily totals remain visible independently of graph mode, and graph selection must not redefine the projection's evidence window.
+- Open on the latest three hours. Offer selected-day and longer-history views at the precision supported by retained evidence. Selected-point readouts expose time, labelled timezone, data-volume units, and incomplete/gap state. Do not imply that a partial current day or an incompletely allocated day is a known full-day total.
+- Use `[` / `]` for previous/next day, `g` for date entry, `t` for today/live, and arrows for point inspection, with visible hints and clickable equivalents. Date entry must handle invalid or unavailable dates visibly and must not leak keystrokes into unrelated actions. Cancelling returns to the prior selection; refresh preserves historical selection until the user changes it.
+- Resolve the existing `t` theme binding in favor of the accepted today/live action. Preserve theme selection through a discoverable non-conflicting control and update help consistently. Preserve pause, resume, collect-now, disclosures, quit, and reduced-motion behavior; quitting still does not pause monitoring. No exact replacement theme shortcut was selected in the interview.
+- Preserve UTC timestamps and the established UTC hour/day evidence used by endurance calculations. Local-day activity is a separate presentation aggregate using the actual local midnight boundaries, including non-whole-hour offsets and daylight-saving transitions. Timestamp relabelling alone is insufficient.
+- Extend the existing versioned observation store to retain local-day read/write summaries with their controller identity/segment context, recorded timezone, local date, UTC boundaries, evidence state, and the information needed to explain unallocated boundary volume. These are required semantics, not prescribed table names or a finalized column layout. The collector derives durable summaries before source evidence is eligible for pruning.
+- Retain three-minute detail for 14 days and hourly/daily summaries indefinitely. Keep the boundary evidence and anchors necessary for honest derivation. Do not introduce unbounded retention of all fine-grained intervals solely to support arbitrary historical timezone reinterpretation, and do not delete existing durable evidence merely to simplify migration.
+- Preserve the timezone and boundaries recorded for historical summaries when the system timezone changes. New observations follow the current local timezone without rewriting old days; labels must make historical timezone context explicit. Do not blend differently bounded summaries into an apparently exact daily total.
+- Allocate a measured interval to a local day only when evidence supports that allocation. Preserve an ambiguous midnight-spanning delta once as shared/unallocated boundary evidence; neither prorate it nor copy its full value into both days. If only coarse legacy UTC summaries survive, show them at their actual precision and mark unreconstructable local-day totals incomplete/unavailable. Repeated derivation and repair must be idempotent.
+- Use the existing usage-adjusted theoretical lifespan model and categorical projection confidence. Forecast remaining write endurance from an applicable endurance baseline and a usable observed write rate. Lifetime-written counters consume the endurance allowance; observation-window deltas determine the usage rate. Read volume is not included in endurance consumption.
+- Add the agreed complete-observation-day gate to the shared projection contract. A completed local midnight-to-midnight day within a monitoring period with usable evidence is required; a partial first day, deliberately disabled span, or day with no usable evidence cannot independently qualify. Monday-noon setup can first qualify at Wednesday 00:00 after observing Tuesday. Real daylight-saving day boundaries apply; do not replace this with a fixed 24-hour timer.
+- Before that gate, explain that Fenris is waiting for a full local observation day. Afterward, allow an estimate with Limited confidence when the existing baseline, rate, identity, and evidence-validity rules permit it; preserve the longer warm-up and Supported-confidence requirements. Missing evidence is still evaluated under the existing validity/coverage rules: completing a date does not turn unknown measurements into known data.
+- Show confidence as a category plus contributing facts, never a numeric confidence percentage. Explain missing baseline, unusable/zero rate, stale evidence, and controller changes through their applicable states. Do not fabricate a baseline, suppress a missing prerequisite behind a generic confidence warning, or label the output as a predicted hardware-failure date. Preserve the existing baseline setup path and zero-rate explanation.
+- Preserve visible freshness, last-collection outcome, paused state, and store-fault behavior. On constrained terminals retain the existing textual fallback/reflow approach so both daily totals, the selected date, evidence state, confidence, and navigation remain accessible; colour alone must not carry meaning.
+- Migrate through the existing ordered, versioned, transactional store mechanism. Preserve observation history, refuse unknown newer schemas, establish new durable evidence before pruning, and keep readers consistent while the collector writes. Migration cannot manufacture local precision from historical data that lacks it.
+- This design amends ADR 0003's five-minute default on both native service backends under ADR 0008, adds a first-full-local-day eligibility gate to ADR 0002 while retaining its forecasting and confidence model, and extends ADR 0001 with the local-history decision recorded as ADR 0010. ADR 0005's no-fabrication/store-fault contract remains authoritative. Existing historical design documents are not evidence that the current runtime already implements these amendments.
+
+## Testing Decisions
+
+- Prefer one primary acceptance path: controlled acquisition fixtures and an injected clock enter the existing public collection operation; it writes a real temporary observation store; the normal read path supplies the TUI and CLI. Assert resulting displayed volumes, evidence states, selections, forecast eligibility, and public outcomes. Do not mock the internal derivation or query results under test. A good test would continue passing after an internal refactor and fail when a user's observed result becomes incorrect.
+- Reuse existing collector tracer and collector-history tracer fixtures, including SMART JSON, a temporary sysfs tree, the injected clock, and a real SQLite store. Compose these with the existing status reader, projection fixtures, and headless TUI driver rather than creating a new production testing interface. Fix defects that this path reproduces; do not make tests reproduce incorrect implementation arithmetic.
+- Exercise normal launch, previous/next date, direct date entry, invalid input, cancel, today/live, arrow inspection, read/write switching, refresh during historical browsing, and hourly/day navigation through user actions. Assert visible graph/readout results, not only private selection indices or row counts. In particular, do not manually force graph focus to make the advertised shortcuts work. Existing headless graph, help, theme, motion, and constrained-terminal tests provide prior art.
+- Drive known successive read/write counter changes through collection and verify exact interval volumes and aggregate totals. Cover first reading, first compatible pair, measured zero, repeated collection, repeated read/refresh, restart, same-hour accumulation, and a day-boundary crossing. Preserve each measured byte once: attributable totals plus separately retained unallocated evidence must account for the measured delta without duplication.
+- Use the same public path for local midnight, Asia/Kolkata's half-hour boundary, a 23-hour local day, a 25-hour local day, repeated local clock labels, and a system-timezone change. Verify historical labels and boundaries remain stable and ambiguous intervals are not silently divided. Add direct derivation cases only when they prove an invariant that the primary path cannot isolate clearly.
+- Advance controlled time beyond the 14-day detail window, invoke normal pruning, restart readers, and verify recent detail expiry with durable summary/boundary preservation. Use existing pruning, repair, legacy-migration, and store-migration tests for focused persistence cases that the main path cannot prove, including interruption, reruns, concurrent readers, unknown newer schema, and inability to reconstruct legacy local-day precision.
+- Test the forecast gate at concrete local times: Monday noon setup; Tuesday noon with 24 elapsed hours but no complete observed calendar day; and Wednesday 00:00 after a usable Tuesday. Include partial/disabled/unknown days and daylight-saving dates. After eligibility, verify Limited confidence with reasons and a correct estimate where inputs support one; no baseline, invalid counters, zero rate, stale evidence, and segment changes must preserve the relevant unavailable/degraded behavior. Reuse existing projection and status-parity tests.
+- Independently calculate an expected endurance result from known lifetime-written counters, a known applicable baseline, and known observed-window deltas. This guards against confusing lifetime consumption with writes inside the rate window. Rendering, browsing another day, and switching to reads must not change that accounting.
+- Add focused checks at the existing native scheduler boundary for the three-minute defaults, serialization, failures, and freshness calculations. Reuse systemd/runit control and packaging tests; use only necessary controlled platform probes to verify actual scheduled/background behavior. No live-system scheduling, pause, install, or reboot action is authorized merely by writing this spec. Report any unavailable platform verification during implementation.
+- Verify at 80×24 and below that visible totals, date entry, navigation, confidence, and evidence states remain accessible and that theme/reduced-motion controls still work. Avoid pixel-perfect snapshots and assertions about private widget layout when visible text and action results express the contract.
+- The affected modules are collection/derivation, observation-store migration and retention, shared projection/status, native service scheduling, and the TUI. Most acceptance coverage should remain at the single collection-to-visible-result path; targeted persistence and scheduler checks are supporting boundaries, not a new layer of test-only architecture.
+
+## Out of Scope
+
+- Adding SATA/HDD acquisition, another monitored device, or multi-drive management; the selected existing NVMe drive remains the source.
+- Reporting occupied/free filesystem capacity, making transfer speed the primary measurement, or treating read bytes as write-endurance consumption.
+- Forecasting today's final data volume or future daily activity; the selected forecast is the usage-adjusted theoretical lifespan.
+- Predicting a physical hardware-failure date, adding numeric confidence percentages, fabricating endurance baselines, or introducing a new forecasting algorithm when the existing model fits.
+- Interpolating missing samples, zero-filling unknown activity, prorating ambiguous midnight intervals, or retroactively rewriting historical timezone boundaries.
+- Requiring all historical dates to retain three-minute detail forever or to support arbitrary later timezone regrouping.
+- A new web UI, plotting service, generic graph framework, notifications, telemetry, theme redesign, or unrelated dashboard rework.
+- Package publication, release creation, installation, host-monitoring changes, or runtime diagnosis of the user's deployed observation store as part of this specification task.
+
+## Further Notes
+
+- This spec synthesizes the completed design conversation. The user accepted daily volume, a live three-hour opening view, the specified keyboard date workflow, local dates, 14-day detail retention, a writes-first graph with read toggle and both totals, an endurance forecast, categorical confidence, one full local observation day before forecasting, and historically labelled timezone boundaries with honest incomplete evidence.
+- Related prior issue: [Implement Fenris TUI polish and hourly history](https://git.bongbetic.com/xavierk/Fenris/issues/72). This newer accepted design supersedes its conflicting requirements for the opening graph, writes-only presentation, the `t` theme shortcut, unrestricted historical timezone regrouping, indefinite fine-grained interval retention, and the earlier projection-eligibility presentation. Preserve unrelated requirements for identity, status, themes, motion, accessibility, collector publication, evidence conservation, and CLI parity. Do not treat the older issue as a reason to undo these newly agreed choices.
+- ADR 0010 records the deliberate trade-off: retain labelled local summaries and necessary boundary evidence before detailed readings expire, rather than retaining every fine-grained interval for arbitrary future timezone reinterpretation. The design adds durable local activity evidence while retaining UTC projection evidence and the existing collector/reader ownership boundary.
+- Source inspection identified normal collection paths that do not consistently rebuild daily totals, existing-hour updates that omit read deltas, cross-day deltas added to both dates, and drill-down rendering overwritten by a loading placeholder. These are concrete investigation/regression targets, not claims of executed reproductions or deployed-store corruption.
+- The user confirmed the existing collector-to-store-to-TUI/CLI test boundary, with focused migration/pruning and scheduler checks where needed. Publication is an implementation handoff; no application behavior or test result is claimed solely by creating this issue.

docs/spec/release-packaging.md

@@ -44,7 +44,8 @@
 - **RPM payload: signed.** rpmsign with the dedicated packaging key, invoked by `make sign-rpm` after the package is built. This is required, not optional: it is the only working dnf-native verification path.
 - **deb: unsigned.** apt never verifies payload signatures; trust = instance-signed `InRelease` (signed-by keyring) + TLS + Acquire-By-Hash. Manual-download integrity is covered by SHA256SUMS.
 - **SHA256SUMS: clearsigned** with the packaging key — the trust anchor for manually downloaded release assets, independent of TLS.
-- **Packaging key:** single dedicated key, RSA 3072, UID `Fenris Packaging <packaging@bongbetic.com>`, 2-year expiry, no master/subkey hierarchy (single maintainer, manual builds). Private key lives in the password manager only; each release does import → sign → delete — nothing permanent on any build host. The full ceremony is documented in `docs/install/signing-key-ceremony.md`.
+- **Packaging key:** single dedicated key, RSA 3072, UID `Fenris Packaging <packaging@bongbetic.com>`, 2-year expiry, no master/subkey hierarchy. The private key is stored as the repository Actions secret `GPG_PRIVATE_KEY`. The release workflow imports it on the self-hosted runner, verifies it against the in-repo public key, signs the RPM and SHA256SUMS, then deletes the runner's keyring copy in an `always()` cleanup step. The full ceremony is documented in `docs/install/signing-key-ceremony.md`.
+- **XBPS key:** separate RSA 3072 key stored as the repository Actions secret `XBPS_SIGNING_KEY`; its public key and fingerprint are published in `Fenris-xbps`. The release workflow uses it for the XBPS package and, when publication is explicitly requested, the repository index. A final `always()` cleanup deletes its runner copy after publication and release asset upload.
 - **Public key publication:** in-repo `packaging/keys/fenris-packaging.asc` (raw URL doubles as the `.repo` gpgkey target), release notes, docs page. No keyservers — TOFU-over-TLS.
 - **Rotation (outline):** new key published alongside old; rpm signed with the new key; `fenris.repo` gpgkey lists both URLs (dnf accepts multiple); old key dropped after one release cycle. Procedure details stay in map fog.
 
@@ -53,9 +54,9 @@
 - **A Release is:** a version tag, its packages in the channel, a Gitea release entry with notes, and a clearsigned SHA256SUMS — all together. **Bare tags are forbidden** (tag without packages + release entry is not a Release).
 - **Cadence: on-demand.** Tag when user-visible changes or fixes accumulate; no calendar, no empty releases, no frequency SLA, no RC ceremony — fixes ship as a revision bump of the current version.
 - **Versioning: plain semver.** Major = breaking CLI/config/unit change; store schema changes ride the natural bump (the forward-only refusal handles old-reader/new-store).
-- **Promotion flow:** tag → `make release` (automated: `make package` → RPM signing via nfpm → SHA256SUMS generation → clearsign → prints registry PUTs + Gitea release steps). The ceremony is documented in `docs/install/signing-key-ceremony.md`.
+- **Promotion flow:** bump `pyproject.toml` and the matching dated `CHANGELOG.md` section, then push tag `v<version>`. The repository-scoped Gitea Actions workflow builds and validates the deb, rpm, checksums, and release entry. XBPS publication remains a manual dispatch option after host acceptance. `make release` is for local artifact preparation and does not replace the tag workflow as the supported publication path. The ceremony is documented in `docs/install/signing-key-ceremony.md`.
 - **Rollback:** installing an older package over a newer store is **unsupported** — the store's forward-only version refusal fails it by design. Documented rollback = restore the observation-store snapshot, then install the old Release. No automatic downgrade machinery exists or will be built.
-- **CI:** no runners are registered on the instance today ([Actions runner research](https://git.bongbetic.com/xavierk/Fenris/issues/37)), so the manual flow above is primary. A dormant `.gitea/workflows/release.yml` (`on: push: tags: ['v*']`, single job, host-mode runner) is committed alongside; if it fires, it replicates `make release`. Cheapest future upgrade: one `act_runner` static binary in host-label mode on the existing Gitea host.
+- **CI:** a repository-scoped self-hosted runner is registered and online (checked 2026-09-29). `.gitea/workflows/release.yml` is the tag-triggered release path; maintainers must confirm runner availability and required Gitea secrets before tagging. The workflow publishes deb/rpm packages and release assets; Void publication is withheld unless the signed XBPS host-release step is explicitly requested.
 
 ## 6. Package layout and ownership
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "fenris"
-version = "0.5.0"
+version = "0.6.0"
 description = "NVMe wear monitor with persistent TUI"
 requires-python = ">=3.10"
 license = {file = "LICENSE"}

src/fenris/activity_selection.py

@@ -0,0 +1,132 @@
+"""User choices for live and historical activity views."""
+from collections.abc import Sequence
+from dataclasses import dataclass
+
+HISTORY_RANGE_DEFAULT = 14
+
+
+@dataclass(frozen=True)
+class IntervalIdentity:
+    """Stable identity for one measured interval."""
+
+    start_ts: str
+    end_ts: str
+
+
+@dataclass(frozen=True)
+class HistoryDayIdentity:
+    """Stable identity for one recorded local-day summary."""
+
+    local_date: str
+    timezone: str
+    utc_start: str
+    utc_end: str
+
+
+class ActivitySelection:
+    """Keep activity navigation state separate from its Textual rendering."""
+
+    def __init__(self) -> None:
+        self.view = "live"
+        self.browse_date: str | None = None
+        self.selected_history_day: HistoryDayIdentity | None = None
+        self.selected_history_hour: str | None = None
+        self.history_range_days = HISTORY_RANGE_DEFAULT
+        self.measure = "written"
+        self.following_live = True
+        self.selected_live_interval: IntervalIdentity | None = None
+        self.live_interval_expired = False
+
+    def set_view(self, view: str) -> None:
+        """Select live, day, or history view."""
+        if view not in ("live", "day", "history"):
+            raise ValueError(f"unknown activity view: {view}")
+        self.view = view
+        if view == "live":
+            self.follow_live()
+
+    def select_history_date(
+        self,
+        date: str | None,
+        identity: HistoryDayIdentity | None = None,
+    ) -> None:
+        """Keep requested date and evidence identity across refresh."""
+        if date != self.browse_date or (
+            identity is not None and identity != self.selected_history_day
+        ):
+            self.selected_history_hour = None
+        if date != self.browse_date:
+            self.selected_history_day = None
+        self.browse_date = date
+        if identity is not None:
+            self.selected_history_day = identity
+        elif date is None:
+            self.selected_history_day = None
+
+    def select_history_hour(self, hour: str | None) -> None:
+        """Keep the chosen historical hour across refresh."""
+        self.selected_history_hour = hour
+
+    def toggle_measure(self) -> str:
+        """Switch read/write presentation without changing point selection."""
+        self.measure = "read" if self.measure == "written" else "written"
+        return self.measure
+
+    def follow_live(self) -> None:
+        """Resume following the newest interval and clear stale-pin notices."""
+        self.view = "live"
+        self.browse_date = None
+        self.selected_history_day = None
+        self.selected_history_hour = None
+        self.following_live = True
+        self.selected_live_interval = None
+        self.live_interval_expired = False
+
+    def update_live(self, intervals: Sequence[IntervalIdentity]) -> None:
+        """Reconcile the selected interval with the rolling live window."""
+        if self.following_live:
+            self.selected_live_interval = intervals[-1] if intervals else None
+            return
+        if self.selected_live_interval in intervals:
+            return
+
+        self.following_live = True
+        self.selected_live_interval = intervals[-1] if intervals else None
+        self.live_interval_expired = True
+
+    def move_live(
+        self, intervals: Sequence[IntervalIdentity], offset: int,
+    ) -> IntervalIdentity | None:
+        """Inspect an adjacent interval by stable identity."""
+        if not intervals:
+            return None
+        if self.selected_live_interval in intervals:
+            current = intervals.index(self.selected_live_interval)
+        else:
+            current = len(intervals) - 1
+        selected = intervals[max(0, min(len(intervals) - 1, current + offset))]
+        self._pin_live_interval(selected)
+        return selected
+
+    def inspect_live(
+        self,
+        intervals: Sequence[IntervalIdentity],
+        interval: IntervalIdentity,
+    ) -> None:
+        """Pin the interval chosen by mouse inspection."""
+        if interval in intervals:
+            self._pin_live_interval(interval)
+
+    def selected_live_index(self, intervals: Sequence[IntervalIdentity]) -> int:
+        """Return the selected point's current render index, or -1."""
+        if self.selected_live_interval is None:
+            return -1
+        try:
+            return intervals.index(self.selected_live_interval)
+        except ValueError:
+            return -1
+
+    def _pin_live_interval(self, interval: IntervalIdentity) -> None:
+        self.selected_live_interval = interval
+        self.following_live = False
+        self.live_interval_expired = False

src/fenris/collect.py

@@ -27,7 +27,6 @@ if VENV_DIR.exists():
         if site_packages:
             sys.path.insert(0, str(site_packages))
 
-from fenris.store import init_store, get_store_path
 from fenris.collector import run_collection
 
 
@@ -103,14 +102,6 @@ def main() -> None:
         config = load_config()
         device = config["device"]
 
-        # Interrogate the drive
-        smartctl_data = interrogate_drive(device)
-
-        # Find sysfs path
-        sysfs_path = find_nvme_sysfs()
-        if sysfs_path is None:
-            raise RuntimeError("No NVMe controller found in sysfs")
-
         # Inject a simple clock
         class SimpleClock:
             def utcnow(self):
@@ -118,11 +109,24 @@ def main() -> None:
 
         clock = SimpleClock()
 
+        def acquire():
+            """Interrogate the drive after pending-work preflight."""
+            smartctl_data = interrogate_drive(device)
+            sysfs_path = find_nvme_sysfs()
+            if sysfs_path is None:
+                raise RuntimeError("No NVMe controller found in sysfs")
+            return smartctl_data, sysfs_path
+
         # Run collection
-        result = run_collection(smartctl_data, sysfs_path, config, clock)
+        result = run_collection(config=config, clock=clock, acquire=acquire)
 
         if result["ok"]:
             print(f"Collection successful: {result['sample_count']} sample(s)")
+            if result.get("retention_error"):
+                print(
+                    f"Retention deferred: {result['retention_error']}",
+                    file=sys.stderr,
+                )
             sys.exit(0)
         else:
             print(f"Collection failed: {result['error']}", file=sys.stderr)

src/fenris/collector.py

@@ -5,19 +5,21 @@ This module implements the thinnest complete write path:
 - Acquire controller identity from sysfs
 - Normalize identity exactly once at write time
 - Validate every row against store invariants
+- Stage valid observations privately before derivation
 - Publish the sample and derived evidence in one collection-owned transaction
 
 No code path outside the collector interrogates the device.
 """
 import json
 import sqlite3
+from collections.abc import Callable
 from datetime import datetime, timezone
 from pathlib import Path
-from typing import Any, Dict, Optional, Tuple
+from typing import Any, Dict, Optional
 
-from .store import init_store, get_store_path
+from .derive import derive_hours_from_interval, find_previous_sample
 from .monitoring_periods import ensure_period_open
-from .derive import find_previous_sample, derive_hours_from_interval
+from .store import init_store, get_store_path
 
 
 class AcquisitionError(Exception):
@@ -30,6 +32,29 @@ class InvariantViolationError(Exception):
     pass
 
 
+class _PendingDerivationError(Exception):
+    """A valid staged observation failed while deriving dependent evidence."""
+
+    def __init__(self, cause: Exception):
+        super().__init__(str(cause))
+        self.cause = cause
+
+
+class _PendingRecoveryFailure(Exception):
+    """A publication attempt failed after earlier pending rows committed."""
+
+    def __init__(self, error: Exception, published_count: int):
+        derivation_error = isinstance(error, _PendingDerivationError)
+        cause = error.cause if derivation_error else error
+        super().__init__(str(cause))
+        self.cause = cause
+        self.published_count = published_count
+        self.derivation_error = derivation_error
+
+
+PENDING_PUBLICATION_LIMIT = 6720
+
+
 def acquire_from_smartctl(smartctl_data: Dict[str, Any]) -> Dict[str, Any]:
     """Acquire counters and thermal evidence from smartctl -a -j data.
     
@@ -195,7 +220,7 @@ def write_sample(
     sample: Dict[str, Any],
     identity: Dict[str, Any],
     conn: sqlite3.Connection,
-    clock,
+    observed_at: datetime,
 ) -> Dict[str, Any]:
     """Write one sample to the observation store.
     
@@ -219,8 +244,7 @@ def write_sample(
     # Open new segment if needed
     segment_opened = False
     if should_open:
-        now = clock.utcnow()
-        open_segment(conn, now, identity, identity_key, identity_degraded)
+        open_segment(conn, observed_at, identity, identity_key, identity_degraded)
         segment_opened = True
     
     # Get current segment_id for provenance
@@ -235,8 +259,8 @@ def write_sample(
             percentage_used, available_spare, media_errors, power_on_hours,
             power_cycles, unsafe_shutdowns, temperature_c,
             data_units_written, data_units_read, bytes_written, bytes_read,
-            critical_warning, segment_id
-        ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+            critical_warning, segment_id, local_tz
+        ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
         """,
         (
             sample["ts"],
@@ -259,6 +283,7 @@ def write_sample(
             sample["bytes_read"],
             sample["critical_warning"],
             segment_id,
+            sample.get("local_tz"),
         ),
     )
     
@@ -268,55 +293,34 @@ def write_sample(
         "identity_key": identity_key,
         "identity_degraded": identity_degraded,
         "segment_id": segment_id,
+        "sample_id": cursor.lastrowid,
     }
 
 
-def run_collection(
-    smartctl_data: Dict[str, Any],
-    sysfs_path: Path,
-    config: Dict[str, Any],
-    clock,
-) -> Dict[str, Any]:
-    """Run one collection run.
+def _observation_time(sample: Dict[str, Any]) -> datetime:
+    """Read an observation's original timestamp for ordered recovery."""
+    observed_at = datetime.fromisoformat(sample["ts"])
+    if observed_at.tzinfo is None:
+        return observed_at.replace(tzinfo=timezone.utc)
+    return observed_at.astimezone(timezone.utc)
 
-    This is the main entry point for the collector.
-    Returns the run outcome.
-    """
-    conn = None
-    try:
-        # Acquire counters and thermal evidence
-        counters = acquire_from_smartctl(smartctl_data)
-        
-        # Acquire controller identity
-        identity = acquire_from_sysfs(sysfs_path)
-        
-        # Build sample with injected clock
-        sample = {
-            "ts": clock.utcnow().isoformat(),
-            "device": config["device"],
-            **counters,
-            **identity,
-        }
-        
-        # Initialize store if needed
-        store_path = get_store_path(config)
-        conn = init_store(store_path)
-
-        # Run legacy import if needed (idempotent)
-        from .legacy import import_legacy_history
-        history_path = Path(config.get("data_dir", ".")) / "history.jsonl"
-        if history_path.exists():
-            import_legacy_history(conn, history_path, clock=clock)
-
-        # Collection owns one transaction for the sample and its evidence.
-        ensure_period_open(conn, clock.utcnow())
 
+def _publish_observation(
+    conn: sqlite3.Connection,
+    sample: Dict[str, Any],
+    identity: Dict[str, Any],
+    tz_name: str,
+) -> None:
+    """Publish one staged sample and all dependent evidence in caller transaction."""
+    observed_at = _observation_time(sample)
+    sample["local_tz"] = tz_name
     validate_sample_invariants(sample, conn)
-        seg_info = write_sample(sample, identity, conn, clock)
+    ensure_period_open(conn, observed_at)
+    seg_info = write_sample(sample, identity, conn, observed_at)
 
-        cursor = conn.execute("SELECT id FROM samples ORDER BY id DESC LIMIT 1")
-        current_id = cursor.fetchone()[0]
+    current_id = seg_info["sample_id"]
     prev = find_previous_sample(conn, seg_info.get("segment_id"), current_id)
+    try:
         if prev is not None:
             current = {
                 "id": current_id,
@@ -327,35 +331,252 @@ def run_collection(
                 "temperature_c": sample["temperature_c"],
                 "data_units_written": sample["data_units_written"],
                 "data_units_read": sample["data_units_read"],
+                "local_tz": tz_name,
             }
             derive_hours_from_interval(conn, prev, current)
+            from .local_day import record_local_activity_interval
+            record_local_activity_interval(
+                conn,
+                prev,
+                current,
+                start_sample_id=prev["id"],
+                end_sample_id=current_id,
+                segment_id=seg_info.get("segment_id"),
+            )
+        else:
+            previous_any_segment = find_previous_sample(conn, None, current_id)
+            if previous_any_segment is not None:
+                from .local_day import mark_local_activity_gap
+                mark_local_activity_gap(
+                    conn,
+                    sample["ts"],
+                    tz_name,
+                    previous_any_segment,
+                )
 
         from .day_aggregate import derive_all_days, persist_day_aggregate
-        for agg in derive_all_days(conn):
-            persist_day_aggregate(conn, agg)
+        for aggregate in derive_all_days(conn):
+            persist_day_aggregate(conn, aggregate)
 
-        from .tz_util import detect_system_tz
         from .local_day import derive_local_day_summary, persist_local_day
-        tz_name = detect_system_tz()
-        local_summary = derive_local_day_summary(conn, tz_name, clock.utcnow())
+        local_summary = derive_local_day_summary(conn, tz_name, observed_at)
         if local_summary is not None:
             persist_local_day(conn, local_summary)
+    except Exception as exc:  # noqa: BLE001 - preserve valid staged evidence for retry.
+        raise _PendingDerivationError(exc) from exc
 
+
+def _pending_count(conn: sqlite3.Connection) -> int:
+    return conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0]
+
+
+def _stage_observation(
+    conn: sqlite3.Connection,
+    sample: Dict[str, Any],
+    identity: Dict[str, Any],
+    tz_name: str,
+) -> int:
+    """Durably stage valid acquired evidence before attempting publication."""
+    payload = json.dumps(
+        {"sample": sample, "identity": identity, "tz_name": tz_name},
+        separators=(",", ":"),
+        sort_keys=True,
+    )
+    cursor = conn.execute(
+        "INSERT INTO pending_publications (sample_ts, payload) VALUES (?, ?)",
+        (sample["ts"], payload),
+    )
+    return cursor.lastrowid
+
+
+def _recover_pending(conn: sqlite3.Connection) -> int:
+    """Publish pending observations oldest first; stop at first failure."""
+    recovered = 0
+    while True:
+        conn.execute("BEGIN IMMEDIATE")
+        try:
+            row = conn.execute(
+                "SELECT id, payload FROM pending_publications ORDER BY id LIMIT 1"
+            ).fetchone()
+            if row is None:
+                conn.rollback()
+                return recovered
+
+            pending_id, payload = row
+            observation = json.loads(payload)
+            try:
+                _publish_observation(
+                    conn,
+                    observation["sample"],
+                    observation["identity"],
+                    observation["tz_name"],
+                )
+            except Exception as exc:
+                raise _PendingRecoveryFailure(exc, recovered) from exc
+            conn.execute("DELETE FROM pending_publications WHERE id = ?", (pending_id,))
             conn.commit()
+            recovered += 1
+        except Exception:
+            conn.rollback()
+            raise
 
+
+def _is_store_or_invariant_failure(exc: Exception) -> bool:
+    return (
+        isinstance(exc, sqlite3.Error)
+        and not isinstance(exc, sqlite3.IntegrityError)
+    ) or isinstance(exc, InvariantViolationError)
+
+
+def _pending_capacity_error(
+    reason: str,
+    *,
+    acquisition_skipped: bool = False,
+) -> RuntimeError:
+    outcome = "; no new observation acquired" if acquisition_skipped else ""
+    return RuntimeError(
+        "pending publication capacity full "
+        f"({PENDING_PUBLICATION_LIMIT} observations){outcome}; "
+        f"{reason}"
+    )
+
+
+def _recover_pending_for_collection(conn: sqlite3.Connection) -> int:
+    """Retry queued work and report exhausted capacity without masking store faults."""
+    try:
+        return _recover_pending(conn)
+    except Exception as exc:  # noqa: BLE001 - preserve pending evidence on recovery errors.
+        cause = exc.cause if isinstance(exc, _PendingRecoveryFailure) else exc
+        if _is_store_or_invariant_failure(cause):
+            raise cause
+        try:
+            capacity_full = _pending_count(conn) >= PENDING_PUBLICATION_LIMIT
+        except sqlite3.Error:
+            raise cause
+        if capacity_full:
+            raise _pending_capacity_error(f"recovery failed: {cause}") from cause
+        raise cause
+
+
+def _recover_pending_for_admission(conn: sqlite3.Connection) -> int:
+    """Recover in order, then admit one observation if bounded space remains."""
+    try:
+        recovered = _recover_pending(conn)
+    except _PendingRecoveryFailure as failure:
+        if (
+            not failure.derivation_error
+            or _is_store_or_invariant_failure(failure.cause)
+        ):
+            raise failure.cause
+        try:
+            pending_count = _pending_count(conn)
+        except sqlite3.Error:
+            raise failure.cause
+        if pending_count >= PENDING_PUBLICATION_LIMIT:
+            raise _pending_capacity_error(
+                f"recovery failed: {failure.cause}",
+                acquisition_skipped=True,
+            ) from failure.cause
+        return failure.published_count
+
+    if _pending_count(conn) >= PENDING_PUBLICATION_LIMIT:
+        raise _pending_capacity_error(
+            "recovery left the queue full",
+            acquisition_skipped=True,
+        )
+    return recovered
+
+
+def run_collection(
+    smartctl_data: Optional[Dict[str, Any]] = None,
+    sysfs_path: Optional[Path] = None,
+    config: Optional[Dict[str, Any]] = None,
+    clock=None,
+    *,
+    acquire: Optional[Callable[[], tuple[Dict[str, Any], Path]]] = None,
+) -> Dict[str, Any]:
+    """Recover old work, acquire one observation, then publish it atomically.
+
+    Production callers pass ``acquire`` so recovery and capacity checks run
+    before device interrogation. Direct sample arguments remain useful for
+    deterministic collector tests.
+    """
+    conn = None
+    try:
+        if config is None or clock is None:
+            raise ValueError("config and clock are required")
+
+        store_path = get_store_path(config)
+        conn = init_store(store_path)
+
+        from .legacy import import_legacy_history
+        history_path = Path(config.get("data_dir", ".")) / "history.jsonl"
+        if history_path.exists():
+            import_legacy_history(conn, history_path, clock=clock)
+
+        published_count = _recover_pending_for_admission(conn)
+        checked_pending_count = _pending_count(conn)
+
+        # Hold the writer reservation across the capacity check and acquisition.
+        # Recheck recovery if another collector appended work during preflight.
+        while True:
+            conn.execute("BEGIN IMMEDIATE")
+            waiting = _pending_count(conn)
+            if (
+                waiting >= PENDING_PUBLICATION_LIMIT
+                or waiting > checked_pending_count
+            ):
+                conn.rollback()
+                published_count += _recover_pending_for_admission(conn)
+                checked_pending_count = _pending_count(conn)
+                continue
+
+            from .tz_util import detect_system_tz
+            tz_name = detect_system_tz()
+            if acquire is not None:
+                smartctl_data, sysfs_path = acquire()
+            if smartctl_data is None or sysfs_path is None:
+                raise AcquisitionError("No acquired SMART data or sysfs identity")
+
+            counters = acquire_from_smartctl(smartctl_data)
+            identity = acquire_from_sysfs(sysfs_path)
+            sample = {
+                "ts": clock.utcnow().isoformat(),
+                "device": config["device"],
+                **counters,
+                **identity,
+            }
+            validate_sample_invariants(sample, conn)
+            _stage_observation(conn, sample, identity, tz_name)
+            conn.commit()
+            break
+
+        published_count += _recover_pending_for_collection(conn)
+        retention_error = None
+        try:
+            if _pending_count(conn) == 0:
+                from .pruning import prune_old_samples
+
+                prune_old_samples(conn, clock.utcnow())
+        except Exception as exc:  # noqa: BLE001 - publication already committed.
+            # Publication already committed. Keep collection successful and
+            # retry atomic retention on the next normal collection.
+            conn.rollback()
+            retention_error = str(exc)
         return {
             "ok": True,
-            "sample_count": 1,
+            "sample_count": published_count,
+            "retention_error": retention_error,
             "store_path": str(store_path),
         }
 
-    except Exception as e:
+    except Exception as exc:
         if conn is not None:
             conn.rollback()
         return {
             "ok": False,
-            "error": str(e),
-            "error_type": type(e).__name__,
+            "error": str(exc),
+            "error_type": type(exc).__name__,
         }
     finally:
         if conn is not None:

src/fenris/derive.py

@@ -29,7 +29,7 @@ def find_previous_sample(
     if segment_id is not None:
         cursor = conn.execute(
             "SELECT id, ts, bytes_written, bytes_read, power_on_hours, "
-            "       temperature_c, data_units_written, data_units_read "
+            "       temperature_c, data_units_written, data_units_read, local_tz "
             "FROM samples WHERE id < ? AND segment_id = ? "
             "ORDER BY id DESC LIMIT 1",
             (current_sample_id, segment_id),
@@ -37,7 +37,7 @@ def find_previous_sample(
     else:
         cursor = conn.execute(
             "SELECT id, ts, bytes_written, bytes_read, power_on_hours, "
-            "       temperature_c, data_units_written, data_units_read "
+            "       temperature_c, data_units_written, data_units_read, local_tz "
             "FROM samples WHERE id < ? "
             "ORDER BY id DESC LIMIT 1",
             (current_sample_id,),
@@ -49,7 +49,7 @@ def find_previous_sample(
         "id": row[0], "ts": row[1], "bytes_written": row[2],
         "bytes_read": row[3], "power_on_hours": row[4],
         "temperature_c": row[5], "data_units_written": row[6],
-        "data_units_read": row[7],
+        "data_units_read": row[7], "local_tz": row[8],
     }
 
 

src/fenris/monitoring_periods.py

@@ -9,7 +9,7 @@ Key contracts:
 - End causes: user_disabled, migrated, unknown_gap
 """
 import sqlite3
-from datetime import datetime
+from datetime import datetime, timezone
 
 
 def ensure_period_open(conn: sqlite3.Connection, run_time: datetime) -> None:
@@ -82,6 +82,46 @@ def is_inside_period(conn: sqlite3.Connection, ts: datetime) -> bool:
     return cursor.fetchone() is not None
 
 
+def interval_within_one_monitoring_period(
+    conn: sqlite3.Connection,
+    start: str | datetime,
+    end: str | datetime,
+) -> bool:
+    """Return whether one monitoring period contains the full interval.
+
+    Compare timestamps as UTC instants. Malformed timestamps fail closed.
+    """
+    try:
+        interval_start = as_utc_datetime(start)
+        interval_end = as_utc_datetime(end)
+    except (TypeError, ValueError, OverflowError):
+        return False
+    if interval_end <= interval_start:
+        return False
+
+    for started_at, ended_at in conn.execute(
+        "SELECT started_at, ended_at FROM monitoring_periods"
+    ):
+        try:
+            period_start = as_utc_datetime(started_at)
+            period_end = as_utc_datetime(ended_at) if ended_at is not None else None
+        except (TypeError, ValueError, OverflowError):
+            continue
+        if period_start <= interval_start and (
+            period_end is None or interval_end <= period_end
+        ):
+            return True
+    return False
+
+
+def as_utc_datetime(value: str | datetime) -> datetime:
+    """Parse a timestamp and normalize it to an aware UTC datetime."""
+    parsed = value if isinstance(value, datetime) else datetime.fromisoformat(value)
+    if parsed.tzinfo is None:
+        parsed = parsed.replace(tzinfo=timezone.utc)
+    return parsed.astimezone(timezone.utc)
+
+
 def wall_clock_in_periods(
     conn: sqlite3.Connection,
     start: datetime,

src/fenris/projection.py

@@ -28,6 +28,7 @@ from datetime import datetime, timedelta, timezone
 from enum import Enum
 from typing import Any, Dict, List, Optional, Tuple
 
+from .monitoring_periods import interval_within_one_monitoring_period
 
 # ---------------------------------------------------------------------------
 # Constants (spec §6)
@@ -187,7 +188,7 @@ def _wall_clock_in_range(conn, start, end):
 
 def _resolve_baseline(conn, current_segment):
     baseline = _get_baseline(conn)
-    facts = []
+    facts: list[str] = []
     if baseline is None:
         return BaselineTier.NONE, None, "no baseline", facts
 
@@ -494,18 +495,16 @@ def _has_complete_local_day(conn):
     within a monitoring period that has usable observation evidence.
     This is the prerequisite for showing an endurance outlook.
     """
-    row = conn.execute(
-        "SELECT 1 FROM local_days WHERE complete = 1 LIMIT 1"
-    ).fetchone()
-    return row is not None
-
-
-def _count_complete_local_days(conn):
-    """Count the number of complete local observation days."""
-    row = conn.execute(
-        "SELECT COUNT(*) FROM local_days WHERE complete = 1"
-    ).fetchone()
-    return row[0] if row else 0
+    local_days = conn.execute(
+        "SELECT utc_start, utc_end FROM local_days "
+        "WHERE complete = 1 "
+        "AND activity_precision IN ('measured', 'coarse') "
+        "AND activity_intervals > 0 ORDER BY id"
+    ).fetchall()
+    return any(
+        interval_within_one_monitoring_period(conn, utc_start, utc_end)
+        for utc_start, utc_end in local_days
+    )
 
 
 # ---------------------------------------------------------------------------

src/fenris/pruning.py

@@ -1,131 +1,321 @@
-"""Raw sample pruning per spec §3.4, ST-5.
+"""Expire detail only after UTC and local_days replacement evidence is durable.
 
-Raw samples are pruned opportunistically to 14 days.
-Hour observations and day aggregates are retained indefinitely.
-Boundary anchors required for successor evidence are retained.
-
-Local-day summaries (local_days table) are never touched by pruning.
-They are persisted at collection time from hour observations and survive
-raw-sample pruning because they depend on hour observations, not on raw
-samples.  This is the mechanism that keeps local-day history trustworthy
-after detail expires (issue #93).
+The 14-day cutoff never overrides local-day preservation, shared boundary
+evidence, publication recovery, or the newest sample's successor-anchor role.
 """
 import sqlite3
-from datetime import datetime, timedelta, timezone
+from datetime import date, datetime, timedelta
+
+from .monitoring_periods import (
+    as_utc_datetime,
+    interval_within_one_monitoring_period,
+)
 
-# Spec §3.4: Raw-sample retention
 RAW_SAMPLE_RETENTION_DAYS = 14
 
 
+def _parse_sample_time(value: str) -> datetime | None:
+    """Parse old timestamps conservatively; malformed legacy values stay."""
+    try:
+        parsed = datetime.fromisoformat(value)
+    except (TypeError, ValueError):
+        return None
+    return as_utc_datetime(parsed)
+
+
+def _sample_rows(conn: sqlite3.Connection) -> list[tuple]:
+    return conn.execute(
+        "SELECT id, ts, segment_id, bytes_written, bytes_read, local_tz "
+        "FROM samples ORDER BY id"
+    ).fetchall()
+
+
+def _required_hour_window(
+    start: datetime,
+    end: datetime,
+) -> tuple[datetime, datetime, int] | None:
+    """Return first hour, exclusive end and count for [start, end)."""
+    if end <= start:
+        return None
+    first_hour = start.replace(minute=0, second=0, microsecond=0)
+    last_hour = end.replace(minute=0, second=0, microsecond=0)
+    try:
+        exclusive_end = last_hour if end == last_hour else last_hour + timedelta(hours=1)
+    except OverflowError:
+        return None
+    count = int((exclusive_end - first_hour).total_seconds() // 3600)
+    return first_hour, exclusive_end, count
+
+
+def _has_utc_replacement(
+    conn: sqlite3.Connection,
+    previous: tuple,
+    current: tuple,
+    start: datetime,
+    end: datetime,
+) -> bool:
+    window = _required_hour_window(start, end)
+    if window is None:
+        return False
+    first_hour, exclusive_end, expected_hour_count = window
+
+    actual_hour_count = conn.execute(
+        "SELECT COUNT(*) FROM hour_observations WHERE hour >= ? AND hour < ?",
+        (
+            first_hour.strftime("%Y-%m-%dT%H:00:00+00:00"),
+            exclusive_end.strftime("%Y-%m-%dT%H:00:00+00:00"),
+        ),
+    ).fetchone()[0]
+    if actual_hour_count < expected_hour_count:
+        return False
+
+    start_written, end_written = previous[3], current[3]
+    start_read, end_read = previous[4], current[4]
+    if None in (start_written, end_written, start_read, end_read):
+        return False
+    if end_written < start_written or end_read < start_read:
+        return False
+    bytes_written = end_written - start_written
+    bytes_read = end_read - start_read
+    if expected_hour_count == 1:
+        totals = conn.execute(
+            "SELECT bytes_written_delta, bytes_read_delta FROM hour_observations "
+            "WHERE hour = ?",
+            (first_hour.strftime("%Y-%m-%dT%H:00:00+00:00"),),
+        ).fetchone()
+    else:
+        totals = conn.execute(
+            "SELECT unattributed_bytes_written, unattributed_bytes_read "
+            "FROM day_aggregates WHERE day = ?",
+            (start.date().isoformat(),),
+        ).fetchone()
+    if totals is None or totals[0] < bytes_written or totals[1] < bytes_read:
+        return False
+
+    last_day = (exclusive_end - timedelta(microseconds=1)).date()
+    expected_day_count = (last_day - first_hour.date()).days + 1
+    actual_day_count = conn.execute(
+        "SELECT COUNT(*) FROM day_aggregates WHERE day >= ? AND day <= ?",
+        (first_hour.date().isoformat(), last_day.isoformat()),
+    ).fetchone()[0]
+    return actual_day_count >= expected_day_count
+
+
+def _local_day_row_exists(
+    conn: sqlite3.Connection,
+    local_date: str,
+    tz_name: str,
+) -> bool:
+    return conn.execute(
+        "SELECT 1 FROM local_days WHERE local_date = ? AND tz_name = ?",
+        (local_date, tz_name),
+    ).fetchone() is not None
+
+
+def _local_day_bounds(
+    conn: sqlite3.Connection,
+    local_date: str,
+    tz_name: str,
+) -> tuple[datetime, datetime] | None:
+    row = conn.execute(
+        "SELECT utc_start, utc_end FROM local_days "
+        "WHERE local_date = ? AND tz_name = ?",
+        (local_date, tz_name),
+    ).fetchone()
+    if row is None:
+        return None
+    start = _parse_sample_time(row[0])
+    end = _parse_sample_time(row[1])
+    if start is None or end is None or end <= start:
+        return None
+    return start, end
+
+
+def _contains_instant(
+    conn: sqlite3.Connection,
+    local_date: str,
+    tz_name: str,
+    instant: datetime,
+) -> bool:
+    bounds = _local_day_bounds(conn, local_date, tz_name)
+    return bounds is not None and bounds[0] <= instant < bounds[1]
+
+
+def _monitoring_period_covers(
+    conn: sqlite3.Connection,
+    start: datetime,
+    end: datetime,
+) -> bool:
+    """Require continuous monitoring before a local-day total replaces detail."""
+    return interval_within_one_monitoring_period(conn, start, end)
+
+
+def _has_local_replacement(
+    conn: sqlite3.Connection,
+    previous: tuple,
+    current: tuple,
+    start: datetime,
+    end: datetime,
+) -> bool:
+    start_tz = previous[5]
+    end_tz = current[5]
+    if not end_tz:
+        return False
+
+    start_id, end_id = previous[0], current[0]
+    segment_id = current[2]
+    if (
+        start_tz
+        and start_tz == end_tz
+        and segment_id is not None
+        and _monitoring_period_covers(conn, start, end)
+    ):
+        known_days = conn.execute(
+            "SELECT local_days.utc_start, local_days.utc_end "
+            "FROM local_days JOIN local_day_segment_totals "
+            "  ON local_day_segment_totals.local_day_id = local_days.id "
+            "WHERE local_days.tz_name = ? "
+            "  AND local_days.activity_precision IN ('measured', 'coarse') "
+            "  AND local_days.last_sample_id >= ? "
+            "  AND local_days.activity_intervals > 0 "
+            "  AND local_day_segment_totals.segment_id = ? "
+            "  AND local_day_segment_totals.activity_intervals > 0",
+            (end_tz, end_id, segment_id),
+        ).fetchall()
+        for row in known_days:
+            bounds_start = _parse_sample_time(row[0])
+            bounds_end = _parse_sample_time(row[1])
+            if (
+                bounds_start is not None
+                and bounds_end is not None
+                and bounds_start <= start
+                and end < bounds_end
+            ):
+                return True
+
+    evidence = conn.execute(
+        "SELECT bytes_written, bytes_read, reason, start_local_date, end_local_date, "
+        "       start_tz_name, end_tz_name, started_at, ended_at "
+        "FROM local_day_unallocated_evidence "
+        "WHERE start_sample_id = ? AND end_sample_id = ?",
+        (start_id, end_id),
+    ).fetchone()
+    if evidence is None:
+        return False
+
+    start_written, end_written = previous[3], current[3]
+    start_read, end_read = previous[4], current[4]
+    if None in (start_written, end_written, start_read, end_read):
+        return False
+    if end_written < start_written or end_read < start_read:
+        return False
+    # The source IDs uniquely identify this interval. Confirm its durable
+    # deltas and retain every recorded local-day boundary it touches.
+    if (end_written - start_written, end_read - start_read) != evidence[:2]:
+        return False
+    if evidence[5] != start_tz or evidence[6] != end_tz:
+        return False
+    if evidence[2] not in {
+        "counter_discontinuity",
+        "legacy_timezone_unknown",
+        "timezone_change",
+        "monitoring_period",
+        "local_midnight",
+        "segment_unknown",
+    }:
+        return False
+    if _parse_sample_time(evidence[7]) != start or _parse_sample_time(evidence[8]) != end:
+        return False
+
+    affected_days: set[tuple[str, str]] = set()
+    if start_tz:
+        affected_days.add((evidence[3], start_tz))
+    affected_days.add((evidence[4], end_tz))
+    if start_tz and start_tz == end_tz and evidence[3] < evidence[4]:
+        day = date.fromisoformat(evidence[3]) + timedelta(days=1)
+        last_day = date.fromisoformat(evidence[4])
+        while day < last_day:
+            affected_days.add((day.isoformat(), end_tz))
+            day += timedelta(days=1)
+    if not all(
+        _local_day_row_exists(conn, local_date, zone)
+        for local_date, zone in affected_days
+    ):
+        return False
+    if start_tz and not _contains_instant(conn, evidence[3], start_tz, start):
+        return False
+    return _contains_instant(conn, evidence[4], end_tz, end)
+
+
+def _interval_has_replacement(
+    conn: sqlite3.Connection,
+    previous: tuple,
+    current: tuple,
+) -> bool:
+    start = _parse_sample_time(previous[1])
+    end = _parse_sample_time(current[1])
+    if start is None or end is None or end <= start:
+        return False
+    try:
+        return (
+            _has_utc_replacement(conn, previous, current, start, end)
+            and _has_local_replacement(conn, previous, current, start, end)
+        )
+    except (sqlite3.Error, ValueError, OverflowError, TypeError):
+        # Missing or unreadable derived evidence never authorizes deletion.
+        return False
+
+
+def _sample_needs_preservation(
+    conn: sqlite3.Connection,
+    rows: list[tuple],
+    index: int,
+    newest_id: int,
+) -> bool:
+    sample = rows[index]
+    segment_id = sample[2]
+    if segment_id is None or conn.execute(
+        "SELECT 1 FROM controller_segments WHERE id = ?",
+        (segment_id,),
+    ).fetchone() is None:
+        return True
+    if sample[0] == newest_id:
+        # Keep newest sample as the source anchor for the next collection.
+        return True
+
+    pairs = []
+    if index > 0 and rows[index - 1][2] == segment_id:
+        pairs.append((rows[index - 1], sample))
+    if index + 1 < len(rows) and rows[index + 1][2] == segment_id:
+        pairs.append((sample, rows[index + 1]))
+    return any(
+        not _interval_has_replacement(conn, previous, current)
+        for previous, current in pairs
+    )
+
+
 def needs_boundary_anchor(
     conn: sqlite3.Connection,
     sample_ts: str,
     now: datetime,
 ) -> bool:
-    """Check if a sample is needed as a boundary anchor for derivation.
-
-    A sample is a boundary anchor if:
-    1. It's older than retention_days (strictly before cutoff)
-    2. It has a next sample that forms an interval spanning the retention boundary
-    3. The interval hasn't been derived yet
-
-    The interval spans the boundary if:
-    - The sample is before the cutoff, AND
-    - The next sample is strictly after the cutoff (or within retention)
-    """
-    from .derive import _parse_ts
-    
-    sample_dt = _parse_ts(sample_ts)
-    retention_cutoff = now - timedelta(days=RAW_SAMPLE_RETENTION_DAYS)
-    
-    # If sample is within retention (strictly after cutoff), not an anchor
-    if sample_dt > retention_cutoff:
-        return False
-    
-    # Check if this sample has a next sample
-    cursor = conn.execute(
-        """SELECT ts, segment_id FROM samples WHERE ts > ? ORDER BY ts LIMIT 1""",
-        (sample_ts,),
-    )
-    next_row = cursor.fetchone()
-    
-    if next_row is None:
-        # No next sample - this is the last sample
-        # It's not needed for derivation (no interval to derive)
-        return False
-    
-    next_ts_str = next_row[0]
-    next_segment_id = next_row[1]
-    next_dt = _parse_ts(next_ts_str)
-    
-    # Check if the next sample is strictly after the cutoff (i.e., interval spans boundary)
-    if next_dt > retention_cutoff:
-        # The interval spans the retention boundary
-        # Check if the interval needs derivation
-        # Get current sample's segment_id
-        cursor = conn.execute(
-            "SELECT segment_id FROM samples WHERE ts = ?",
-            (sample_ts,),
-        )
-        current_segment_row = cursor.fetchone()
-        current_segment_id = current_segment_row[0] if current_segment_row else None
-        
-        # If different segments, no interval to derive
-        if current_segment_id != next_segment_id:
-            return False
-        
-        # Check if the interval [sample_ts, next_ts] needs derivation
-        # It needs derivation if any hour in the span lacks an observation
-        current_hour = sample_dt.replace(minute=0, second=0, microsecond=0)
-        end_hour = next_dt.replace(minute=0, second=0, microsecond=0)
-        
-        while current_hour <= end_hour:
-            cursor = conn.execute(
-                "SELECT id FROM hour_observations WHERE hour = ?",
-                (current_hour.isoformat(),),
-            )
-            if cursor.fetchone() is None:
-                # This hour lacks an observation - interval needs derivation
+    """Return whether an old sample still carries unreplaced evidence."""
+    sample_time = _parse_sample_time(sample_ts)
+    cutoff = as_utc_datetime(now) - timedelta(days=RAW_SAMPLE_RETENTION_DAYS)
+    if sample_time is None:
         return True
-            current_hour += timedelta(hours=1)
-        
-        # All hours in the span have observations - interval is derived
+    if sample_time >= cutoff:
         return False
-    else:
-        # The interval doesn't span the boundary (both samples are old)
-        # Check if the interval needs derivation
-        # Get current sample's segment_id
-        cursor = conn.execute(
-            "SELECT segment_id FROM samples WHERE ts = ?",
-            (sample_ts,),
+
+    rows = _sample_rows(conn)
+    matching = [index for index, row in enumerate(rows) if row[1] == sample_ts]
+    if not matching:
+        return False
+    newest_id = rows[-1][0] if rows else -1
+    return any(
+        _sample_needs_preservation(conn, rows, index, newest_id)
+        for index in matching
     )
-        current_segment_row = cursor.fetchone()
-        current_segment_id = current_segment_row[0] if current_segment_row else None
-        
-        # If different segments, no interval to derive
-        if current_segment_id != next_segment_id:
-            return False
-        
-        # Check if the interval [sample_ts, next_ts] needs derivation
-        current_hour = sample_dt.replace(minute=0, second=0, microsecond=0)
-        end_hour = next_dt.replace(minute=0, second=0, microsecond=0)
-        
-        while current_hour <= end_hour:
-            cursor = conn.execute(
-                "SELECT id FROM hour_observations WHERE hour = ?",
-                (current_hour.isoformat(),),
-            )
-            if cursor.fetchone() is None:
-                # This hour lacks an observation - interval needs derivation
-                # But only keep if the interval is significant (spans multiple hours)
-                # or if the next sample is the last sample before a gap
-                gap = (next_dt - sample_dt).total_seconds()
-                if gap > 24 * 3600:  # Significant gap (> 24 hours)
-                    return True
-            current_hour += timedelta(hours=1)
-        
-        # All hours in the span have observations or gap is not significant
-        return False
 
 
 def prune_old_samples(
@@ -133,36 +323,50 @@ def prune_old_samples(
     now: datetime,
     retention_days: int = RAW_SAMPLE_RETENTION_DAYS,
 ) -> int:
-    """Remove raw samples older than retention_days.
+    """Delete expired samples only after every dependent fact is durable.
 
-    Retains boundary anchors required for successor evidence.
-
-    Args:
-        conn: Connection to the observation store.
-        now: Current UTC time.
-        retention_days: Number of days to retain (default 14).
-
-    Returns:
-        Number of samples removed.
+    Pending publications are stored separately and never selected here. A
+    sample stays when it is the newest successor anchor, has an unpublishable
+    neighbour interval, lacks UTC or local-day replacement evidence, or has a
+    timestamp that cannot be safely interpreted.
     """
-    cutoff = (now - timedelta(days=retention_days)).isoformat()
-    
-    # Get all samples older than cutoff
-    cursor = conn.execute(
-        "SELECT id, ts FROM samples WHERE ts < ? ORDER BY ts",
-        (cutoff,),
-    )
-    old_samples = cursor.fetchall()
-    
-    removed = 0
-    for sample_id, sample_ts in old_samples:
-        # Check if this sample is a boundary anchor
-        if needs_boundary_anchor(conn, sample_ts, now):
-            continue  # Skip - it's a boundary anchor
-        
-        # Remove the sample
-        conn.execute("DELETE FROM samples WHERE id = ?", (sample_id,))
-        removed += 1
+    cutoff = as_utc_datetime(now) - timedelta(days=retention_days)
+    owns_transaction = not conn.in_transaction
+    savepoint = "fenris_sample_retention"
+    if owns_transaction:
+        conn.execute("BEGIN IMMEDIATE")
+    else:
+        conn.execute(f"SAVEPOINT {savepoint}")
 
+    try:
+        rows = _sample_rows(conn)
+        if not rows:
+            if owns_transaction:
                 conn.commit()
-    return removed
+            else:
+                conn.execute(f"RELEASE SAVEPOINT {savepoint}")
+            return 0
+
+        newest_id = rows[-1][0]
+        expired = []
+        for index, row in enumerate(rows):
+            sample_time = _parse_sample_time(row[1])
+            if sample_time is None or sample_time >= cutoff:
+                continue
+            if _sample_needs_preservation(conn, rows, index, newest_id):
+                continue
+            expired.append(row[0])
+
+        conn.executemany("DELETE FROM samples WHERE id = ?", ((sample_id,) for sample_id in expired))
+        if owns_transaction:
+            conn.commit()
+        else:
+            conn.execute(f"RELEASE SAVEPOINT {savepoint}")
+        return len(expired)
+    except Exception:
+        if owns_transaction:
+            conn.rollback()
+        else:
+            conn.execute(f"ROLLBACK TO SAVEPOINT {savepoint}")
+            conn.execute(f"RELEASE SAVEPOINT {savepoint}")
+        raise

src/fenris/repair.py

@@ -1,8 +1,9 @@
-"""Historical repair and retention (issue #74).
+"""Historical repair and retention (issue #74, #99).
 
 Re-derives hour observations and day aggregates from surviving raw samples,
-with idempotent and interruption-safe guarantees. Preserves import markers,
-boundary anchors, and valid historical summaries.
+rebuilds legacy local-day evidence from surviving samples or trustworthy UTC
+hours, and preserves import markers, boundary anchors, and valid summaries.
+Repair is idempotent and interruption-safe.
 
 Contracts:
 - Idempotent: running repair multiple times produces no duplicates
@@ -71,7 +72,6 @@ def _set_repair_in_progress(conn: sqlite3.Connection, in_progress: bool) -> None
         "INSERT OR REPLACE INTO store_metadata (key, value) VALUES (?, ?)",
         ("repair_in_progress", "true" if in_progress else "false"),
     )
-    conn.commit()
 
 
 def _update_repair_status(conn: sqlite3.Connection, result: RepairResult) -> None:
@@ -101,8 +101,6 @@ def _update_repair_status(conn: sqlite3.Connection, result: RepairResult) -> Non
         ("days_derived", str(days)),
     )
 
-    conn.commit()
-
 
 def get_repair_status(conn: sqlite3.Connection) -> RepairStatus:
     """Get current repair status for read-only views."""
@@ -364,20 +362,16 @@ def repair_derivation(
     elif hasattr(clock, 'utcnow'):
         clock = clock.utcnow()
     
-    # Check if repair is already in progress
-    if is_repair_in_progress(conn):
-        return RepairResult(
-            ok=False,
-            error="Repair already in progress",
-        )
-    
-    # Mark repair as in progress
-    _set_repair_in_progress(conn, True)
-    
     result = RepairResult(ok=True)
     
     try:
-        # Begin transaction
+        # A committed in-progress marker may be stale after process death.
+        # SQLite's write lock serializes active repairs; retry idempotently.
+        if conn.in_transaction:
+            conn.commit()
+        conn.execute("BEGIN IMMEDIATE")
+        _set_repair_in_progress(conn, True)
+        conn.commit()
         conn.execute("BEGIN IMMEDIATE")
         
         # 1. Find and derive intervals from sample pairs
@@ -408,6 +402,11 @@ def repair_derivation(
                 else:
                     result.days_updated += 1
 
+        # Rebuild only legacy or unavailable local-day activity. This shares
+        # the migration derivation and leaves trustworthy published totals intact.
+        from .local_day import repair_legacy_local_day_evidence
+        repair_legacy_local_day_evidence(conn)
+
         # 3. Count boundary anchors retained
         now = clock if isinstance(clock, datetime) else datetime.now(timezone.utc)
         cursor = conn.execute("SELECT ts FROM samples ORDER BY ts")
@@ -428,21 +427,22 @@ def repair_derivation(
         )
         result.legacy_summaries_preserved = cursor.fetchone()[0]
         
-        # Commit transaction
+        # Publish derived rows and completion status together. A process
+        # interruption rolls back the in-progress marker with the repair.
+        _update_repair_status(conn, result)
+        _set_repair_in_progress(conn, False)
         conn.commit()
         
-        # Update repair status
-        _update_repair_status(conn, result)
-        
     except Exception as e:
         conn.rollback()
+        try:
+            _set_repair_in_progress(conn, False)
+            conn.commit()
+        except sqlite3.Error:
+            conn.rollback()
         logger.error("Repair failed: %s", e)
         return RepairResult(
             ok=False,
             error=str(e),
         )
-    finally:
-        # Mark repair as complete
-        _set_repair_in_progress(conn, False)
-    
     return result

src/fenris/status_composition.py

@@ -115,6 +115,7 @@ class StatusComposition:
     # Sample counts for waiting explanations
     sample_count: int = 0
     day_count: int = 0
+    pending_publication_count: int = 0
 
     # Whether the status dot should blink (only Monitoring)
     should_blink: bool = False
@@ -286,6 +287,7 @@ def compose_status(
     freshness_age_s = None
     sample_count = 0
     day_count = 0
+    pending_publication_count = 0
     deliberately_paused = False
 
     if conn is None and store_fault is None and newer_schema is None:
@@ -315,6 +317,10 @@ def compose_status(
             sample_count = cursor.fetchone()[0]
             cursor = conn.execute("SELECT COUNT(*) FROM day_aggregates")
             day_count = cursor.fetchone()[0]
+            schema_version = conn.execute("PRAGMA user_version").fetchone()[0]
+            if schema_version >= 4:
+                cursor = conn.execute("SELECT COUNT(*) FROM pending_publications")
+                pending_publication_count = cursor.fetchone()[0]
         except sqlite3.Error as exc:
             store_fault = str(exc)
 
@@ -328,6 +334,7 @@ def compose_status(
         freshness = "unknown"
         freshness_age_s = None
         sample_count = day_count = 0
+        pending_publication_count = 0
         deliberately_paused = False
 
     # --- External stop detection ---
@@ -417,6 +424,7 @@ def compose_status(
         overlay_base=overlay_base,
         sample_count=sample_count,
         day_count=day_count,
+        pending_publication_count=pending_publication_count,
         continuity=continuity,
         paused_lines=paused_lines,
     )
@@ -459,6 +467,8 @@ def render_status_cli(comp: StatusComposition) -> str:
     facts.append("timer: %s" % (
         "unknown" if comp.timer_active is None else "active" if comp.timer_active else "inactive"
     ))
+    if comp.pending_publication_count:
+        facts.append(_pending_publication_text(comp.pending_publication_count))
 
     if facts:
         lines.append(" · ".join(facts))
@@ -524,6 +534,8 @@ def render_status_tui(comp: StatusComposition) -> str:
     facts.append("Timer: %s" % (
         "unknown" if comp.timer_active is None else "active" if comp.timer_active else "inactive"
     ))
+    if comp.pending_publication_count:
+        facts.append(_pending_publication_text(comp.pending_publication_count).capitalize())
 
     if facts:
         lines.append(" · ".join(facts))
@@ -539,3 +551,8 @@ def render_status_tui(comp: StatusComposition) -> str:
             lines.append(pl[:1].upper() + pl[1:])
 
     return "\n".join(lines)
+
+
+def _pending_publication_text(count: int) -> str:
+    noun = "observation" if count == 1 else "observations"
+    return f"pending publication: {count} {noun} retained for retry"

src/fenris/store.py

@@ -3,8 +3,7 @@
 This module handles:
 - Store initialization with WAL mode
 - Schema versioning with PRAGMA user_version
-- The six entities: samples, hour_observations, day_aggregates,
-  monitoring_periods, controller_segments, endurance_baseline
+- Observation records, derived activity, publication state, and metadata
 """
 import sqlite3
 from pathlib import Path
@@ -12,7 +11,7 @@ from typing import Optional
 
 
 # Schema version - increment on each migration
-SCHEMA_VERSION = 3
+SCHEMA_VERSION = 6
 
 
 # Packaged default placement (spec §8.3). The config may override it, but a
@@ -33,7 +32,7 @@ def get_store_path(config: dict) -> Path:
 def init_store(store_path: Path) -> sqlite3.Connection:
     """Initialize the observation store if not present.
     
-    Creates the database with WAL mode and all six entities.
+    Creates the observation schema, including private pending-publication storage.
     Returns a connection to the store.
     """
     conn = sqlite3.connect(str(store_path))
@@ -73,9 +72,11 @@ def init_store(store_path: Path) -> sqlite3.Connection:
         )
     elif current_version < SCHEMA_VERSION:
         # Older version - apply migrations
+        try:
             _apply_migrations(conn, current_version)
-        conn.execute(f"PRAGMA user_version={SCHEMA_VERSION}")
-        conn.commit()
+        except Exception:
+            conn.close()
+            raise
     
     return conn
 
@@ -107,7 +108,8 @@ def _create_schema(conn: sqlite3.Connection):
             bytes_written INTEGER,
             bytes_read INTEGER,
             critical_warning INTEGER,
-            segment_id INTEGER
+            segment_id INTEGER,
+            local_tz TEXT
         )
     """)
     
@@ -209,10 +211,18 @@ def _create_schema(conn: sqlite3.Connection):
             coverage REAL DEFAULT 0.0,
             sample_count INTEGER DEFAULT 0,
             complete BOOLEAN DEFAULT 0,
+            activity_seconds INTEGER NOT NULL DEFAULT 0,
+            activity_intervals INTEGER NOT NULL DEFAULT 0,
+            activity_incomplete BOOLEAN NOT NULL DEFAULT 0,
+            activity_precision TEXT NOT NULL DEFAULT 'measured',
+            last_sample_id INTEGER,
             UNIQUE(local_date, tz_name)
         )
     """)
 
+    _create_local_day_shared_evidence(conn)
+    _create_local_day_segment_totals(conn)
+
     # Metadata table for store state (e.g., legacy import marker)
     conn.execute("""
         CREATE TABLE IF NOT EXISTS store_metadata (
@@ -221,44 +231,125 @@ def _create_schema(conn: sqlite3.Connection):
         )
     """)
 
+    _create_pending_publications(conn)
+
+
+def _create_pending_publications(conn: sqlite3.Connection) -> None:
+    """Create private staging for valid observations awaiting derivation."""
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS pending_publications (
+            id INTEGER PRIMARY KEY AUTOINCREMENT,
+            sample_ts TEXT NOT NULL,
+            payload TEXT NOT NULL
+        )
+    """)
+
+
+def _create_local_day_shared_evidence(conn: sqlite3.Connection) -> None:
+    """Create once-only local activity evidence that cannot be day-allocated."""
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS local_day_unallocated_evidence (
+            id INTEGER PRIMARY KEY AUTOINCREMENT,
+            start_sample_id INTEGER NOT NULL,
+            end_sample_id INTEGER NOT NULL,
+            start_local_date TEXT NOT NULL,
+            end_local_date TEXT NOT NULL,
+            start_tz_name TEXT,
+            end_tz_name TEXT,
+            started_at TEXT NOT NULL,
+            ended_at TEXT NOT NULL,
+            bytes_written INTEGER NOT NULL DEFAULT 0,
+            bytes_read INTEGER NOT NULL DEFAULT 0,
+            reason TEXT NOT NULL,
+            segment_id INTEGER,
+            UNIQUE(start_sample_id, end_sample_id)
+        )
+    """)
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS local_day_evidence_start "
+        "ON local_day_unallocated_evidence(start_local_date, start_tz_name)"
+    )
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS local_day_evidence_end "
+        "ON local_day_unallocated_evidence(end_local_date, end_tz_name)"
+    )
+
+
+def _create_local_day_segment_totals(conn: sqlite3.Connection) -> None:
+    """Retain the controller-segment provenance behind known day totals."""
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS local_day_segment_totals (
+            id INTEGER PRIMARY KEY AUTOINCREMENT,
+            local_day_id INTEGER NOT NULL,
+            segment_id INTEGER NOT NULL,
+            bytes_written INTEGER NOT NULL DEFAULT 0,
+            bytes_read INTEGER NOT NULL DEFAULT 0,
+            activity_seconds INTEGER NOT NULL DEFAULT 0,
+            activity_intervals INTEGER NOT NULL DEFAULT 0,
+            UNIQUE(local_day_id, segment_id)
+        )
+    """)
+
 
 def _apply_migrations(conn: sqlite3.Connection, current_version: int):
     """Apply forward-only migrations from current_version to SCHEMA_VERSION.
 
-    Each migration step is a transactional block. Add new steps as sequential
-    elif branches when SCHEMA_VERSION increases.
-    
-    Spec: §3.6, §10.2
+    Commit each version transition independently. A failed step rolls back in
+    full while earlier successful steps remain versioned and retryable.
     """
-    # Migration 1→2: add segment_id provenance to samples,
-    # unattributed byte tracking to day_aggregates (issue #73)
-    if current_version < 2:
-        # Defensive: only ALTER if table exists (handles minimal v1 stores)
+    migrations = {
+        2: _migrate_1_to_2,
+        3: _migrate_2_to_3,
+        4: _migrate_3_to_4,
+        5: _migrate_4_to_5,
+        6: _migrate_5_to_6,
+    }
+    while current_version < SCHEMA_VERSION:
+        target_version = current_version + 1
+        migration = migrations.get(target_version)
+        if migration is None:
+            raise ValueError(f"No migration registered for schema {target_version}")
+        conn.execute("BEGIN IMMEDIATE")
+        try:
+            migration(conn)
+            conn.execute(f"PRAGMA user_version={target_version}")
+            conn.commit()
+        except Exception:
+            conn.rollback()
+            raise
+        current_version = target_version
+
+
+def _migrate_1_to_2(conn: sqlite3.Connection) -> None:
+    """Add segment provenance and unattributed UTC byte tracking."""
     tables = {row[0] for row in conn.execute(
         "SELECT name FROM sqlite_master WHERE type='table'"
     ).fetchall()}
     if "samples" in tables:
-            # Check if column already exists (idempotent)
-            cols = {row[1] for row in conn.execute("PRAGMA table_info(samples)").fetchall()}
+        cols = {row[1] for row in conn.execute(
+            "PRAGMA table_info(samples)"
+        ).fetchall()}
         if "segment_id" not in cols:
             conn.execute("ALTER TABLE samples ADD COLUMN segment_id INTEGER")
     if "day_aggregates" in tables:
-            cols = {row[1] for row in conn.execute("PRAGMA table_info(day_aggregates)").fetchall()}
-            if "unattributed_bytes_written" not in cols:
-                conn.execute("ALTER TABLE day_aggregates ADD COLUMN unattributed_bytes_written INTEGER DEFAULT 0")
-            if "unattributed_bytes_read" not in cols:
-                conn.execute("ALTER TABLE day_aggregates ADD COLUMN unattributed_bytes_read INTEGER DEFAULT 0")
-        current_version = 2
+        cols = {row[1] for row in conn.execute(
+            "PRAGMA table_info(day_aggregates)"
+        ).fetchall()}
+        for column in ("unattributed_bytes_written", "unattributed_bytes_read"):
+            if column not in cols:
+                conn.execute(
+                    f"ALTER TABLE day_aggregates ADD COLUMN {column} INTEGER DEFAULT 0"
+                )
 
-    # Migration 2→3: add local_days table for local-day activity totals
-    # (issue #90, ADR 0010).  Pure addition — no existing rows touched.
-    if current_version < 3:
+
+def _migrate_2_to_3(conn: sqlite3.Connection) -> None:
+    """Add local-day activity summaries."""
     tables = {row[0] for row in conn.execute(
         "SELECT name FROM sqlite_master WHERE type='table'"
     ).fetchall()}
     if "local_days" not in tables:
         conn.execute("""
-                CREATE TABLE IF NOT EXISTS local_days (
+            CREATE TABLE local_days (
                 id INTEGER PRIMARY KEY AUTOINCREMENT,
                 local_date TEXT NOT NULL,
                 tz_name TEXT NOT NULL,
@@ -273,7 +364,48 @@ def _apply_migrations(conn: sqlite3.Connection, current_version: int):
                 UNIQUE(local_date, tz_name)
             )
         """)
-        current_version = 3
+
+
+def _migrate_3_to_4(conn: sqlite3.Connection) -> None:
+    """Add private publication staging for acquired observations."""
+    _create_pending_publications(conn)
+
+
+def _migrate_4_to_5(conn: sqlite3.Connection) -> None:
+    """Add measured local-day evidence storage and mark old totals legacy."""
+    tables = {row[0] for row in conn.execute(
+        "SELECT name FROM sqlite_master WHERE type='table'"
+    ).fetchall()}
+    if "samples" in tables:
+        sample_cols = {row[1] for row in conn.execute(
+            "PRAGMA table_info(samples)"
+        ).fetchall()}
+        if "local_tz" not in sample_cols:
+            conn.execute("ALTER TABLE samples ADD COLUMN local_tz TEXT")
+    if "local_days" in tables:
+        local_cols = {row[1] for row in conn.execute(
+            "PRAGMA table_info(local_days)"
+        ).fetchall()}
+        for column, declaration in (
+            ("activity_seconds", "INTEGER NOT NULL DEFAULT 0"),
+            ("activity_intervals", "INTEGER NOT NULL DEFAULT 0"),
+            ("activity_incomplete", "BOOLEAN NOT NULL DEFAULT 0"),
+            ("activity_precision", "TEXT NOT NULL DEFAULT 'legacy'"),
+            ("last_sample_id", "INTEGER"),
+        ):
+            if column not in local_cols:
+                conn.execute(
+                    f"ALTER TABLE local_days ADD COLUMN {column} {declaration}"
+                )
+    _create_local_day_shared_evidence(conn)
+    _create_local_day_segment_totals(conn)
+
+
+def _migrate_5_to_6(conn: sqlite3.Connection) -> None:
+    """Rebuild local-day summaries from surviving trustworthy evidence."""
+    from .local_day import repair_legacy_local_day_evidence
+
+    repair_legacy_local_day_evidence(conn)
 
 
 def migrate_to_latest(store_path: Path) -> int:
@@ -310,9 +442,11 @@ def migrate_to_latest(store_path: Path) -> int:
         return SCHEMA_VERSION
     
     steps = SCHEMA_VERSION - current_version
+    try:
         _apply_migrations(conn, current_version)
-    conn.execute(f"PRAGMA user_version={SCHEMA_VERSION}")
-    conn.commit()
+    except Exception:
+        conn.close()
+        raise
     conn.close()
     return steps
 

tests/test_acceptance_sweep.py

@@ -121,8 +121,8 @@ def _insert_local_day(conn, local_date, tz_name="UTC", tz_offset="+00:00",
     conn.execute(
         "INSERT INTO local_days "
         "(local_date, tz_name, tz_offset, utc_start, utc_end, "
-        " bytes_written, bytes_read, coverage, sample_count, complete) "
-        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+        " bytes_written, bytes_read, coverage, sample_count, complete, activity_intervals) "
+        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 1)",
         (local_date, tz_name, tz_offset, utc_start, utc_end,
          bw, br, coverage, samples, complete),
     )

tests/test_activity_selection.py

@@ -0,0 +1,94 @@
+"""Activity-selection transitions independent from Textual widgets."""
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from fenris.activity_selection import ActivitySelection, IntervalIdentity
+
+
+def _interval(n):
+    return IntervalIdentity(f"start-{n}", f"end-{n}")
+
+
+def test_live_selection_starts_and_stays_on_newest_interval():
+    selection = ActivitySelection()
+    first = [_interval(1), _interval(2)]
+
+    selection.update_live(first)
+    assert selection.following_live
+    assert selection.selected_live_interval == first[-1]
+
+    second = first + [_interval(3)]
+    selection.update_live(second)
+    assert selection.selected_live_interval == second[-1]
+
+
+def test_keyboard_and_mouse_inspection_pin_interval_identity():
+    intervals = [_interval(1), _interval(2), _interval(3)]
+    keyboard = ActivitySelection()
+    mouse = ActivitySelection()
+
+    keyboard.update_live(intervals)
+    keyboard.move_live(intervals, -1)
+    mouse.update_live(intervals)
+    mouse.inspect_live(intervals, intervals[1])
+
+    assert keyboard.selected_live_interval == mouse.selected_live_interval == intervals[1]
+    assert not keyboard.following_live
+    assert not mouse.following_live
+
+    refreshed = [_interval(2), _interval(3), _interval(4)]
+    keyboard.update_live(refreshed)
+    mouse.update_live(refreshed)
+    assert keyboard.selected_live_interval == mouse.selected_live_interval == intervals[1]
+
+
+def test_expired_pin_explains_expiration_and_resumes_following():
+    selection = ActivitySelection()
+    intervals = [_interval(1), _interval(2), _interval(3)]
+    selection.update_live(intervals)
+    selection.inspect_live(intervals, intervals[0])
+
+    newest = _interval(4)
+    selection.update_live([_interval(2), _interval(3), newest])
+
+    assert selection.following_live
+    assert selection.selected_live_interval == newest
+    assert selection.live_interval_expired
+
+
+def test_today_resets_expired_pin_and_read_write_toggle_keeps_identity():
+    selection = ActivitySelection()
+    intervals = [_interval(1), _interval(2)]
+    selection.update_live(intervals)
+    selection.inspect_live(intervals, intervals[0])
+    selection.update_live([intervals[1]])
+    newest = _interval(3)
+    selection.update_live([intervals[1], newest])
+    assert selection.live_interval_expired
+
+    selection.toggle_measure()
+    assert selection.measure == "read"
+    assert selection.selected_live_interval == newest
+
+    selection.select_history_date("2026-09-27")
+    selection.set_view("history")
+    selection.set_view("live")
+    assert selection.browse_date is None
+    assert selection.following_live
+    assert not selection.live_interval_expired
+
+
+def test_live_identity_uses_both_interval_boundaries():
+    first = IntervalIdentity("2026-09-28T10:00:00Z", "2026-09-28T10:03:00Z")
+    changed_start = IntervalIdentity("2026-09-28T09:57:00Z", "2026-09-28T10:03:00Z")
+    selection = ActivitySelection()
+    selection.update_live([first])
+    selection.inspect_live([first], first)
+
+    selection.update_live([changed_start])
+
+    assert selection.following_live
+    assert selection.selected_live_interval == changed_start
+    assert selection.live_interval_expired

tests/test_collector_history_tracer.py

@@ -154,7 +154,7 @@ class TestSchemaMigration:
 
         # Migrate
         steps = migrate_to_latest(db)
-        assert steps == 2  # v1→v2→v3
+        assert steps == 5  # v1→v2→v3→v4→v5→v6
 
         # Verify data preserved
         conn = sqlite3.connect(str(db))
@@ -375,7 +375,9 @@ class TestCollectionAtomicity:
         }
         monkeypatch.setattr("fenris.status.query_service_state", lambda: service)
 
-        now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+        now = datetime.now(timezone.utc).replace(
+            minute=35, second=0, microsecond=0,
+        )
         first = {
             **smartctl_fixture,
             "nvme_smart_health_information_log": {

tests/test_collector_tracer.py

@@ -232,7 +232,7 @@ def test_store_initialization(config_fixture: Dict[str, Any]):
     # Initialize store
     conn = init_store(store_path)
     
-    # Verify all six entities exist
+    # Verify the observation, derived-history, and metadata tables exist.
     cursor = conn.execute("SELECT name FROM sqlite_master WHERE type='table'")
     tables = {row[0] for row in cursor.fetchall()}
     
@@ -249,6 +249,9 @@ def test_store_initialization(config_fixture: Dict[str, Any]):
     expected_tables.add("sqlite_sequence")
     expected_tables.add("store_metadata")
     expected_tables.add("local_days")
+    expected_tables.add("pending_publications")
+    expected_tables.add("local_day_unallocated_evidence")
+    expected_tables.add("local_day_segment_totals")
     assert expected_tables == tables
     
     conn.close()

tests/test_complete_observation_day_gate.py

@@ -98,8 +98,8 @@ def _insert_local_day(conn, local_date, tz_name="UTC", tz_offset="+00:00",
     conn.execute(
         "INSERT INTO local_days "
         "(local_date, tz_name, tz_offset, utc_start, utc_end, "
-        " bytes_written, bytes_read, coverage, sample_count, complete) "
-        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+        " bytes_written, bytes_read, coverage, sample_count, complete, activity_intervals) "
+        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 1)",
         (local_date, tz_name, tz_offset, utc_start, utc_end,
          bw, br, coverage, samples, complete),
     )
@@ -166,6 +166,80 @@ class TestGateNoCompleteDay:
         # The key assertion: gate message should NOT appear when gate IS met
         assert not any("full local observation day" in f for f in result.contributing_facts)
 
+    def test_complete_legacy_day_without_trusted_activity_does_not_open_gate(
+        self, store,
+    ):
+        """A complete flag cannot make unavailable local activity qualify."""
+        _insert_baseline(store)
+        _insert_segment(store)
+        _open_period(store)
+        for i in range(3):
+            day = (datetime(2026, 9, 27) + timedelta(days=i)).strftime("%Y-%m-%d")
+            _insert_day(store, day, bw=1024 * 1024 * 100)
+        _insert_sample(store, "2026-09-30T10:00:00+00:00", pu=5)
+        _insert_local_day(store, "2026-09-29", complete=True)
+        store.execute(
+            "UPDATE local_days SET activity_precision = 'measured', activity_intervals = 0 "
+            "WHERE local_date = '2026-09-29'"
+        )
+        result = compute_projection(store, _clock())
+        assert result.confidence_state == ConfidenceState.UNSUPPORTED
+        assert any("full local observation day" in fact
+                   for fact in result.contributing_facts)
+
+    def test_day_split_by_deliberate_pause_does_not_open_gate(self, store):
+        """A complete-looking summary cannot span separate monitoring periods."""
+        _insert_baseline(store)
+        _insert_segment(store)
+        for i in range(20):
+            day = (datetime(2026, 9, 10) + timedelta(days=i)).strftime("%Y-%m-%d")
+            _insert_day(store, day, bw=1024 * 1024 * 100)
+        _insert_sample(store, "2026-09-30T10:00:00+00:00", pu=5)
+
+        store.executemany(
+            "INSERT INTO monitoring_periods (started_at, ended_at, end_cause) "
+            "VALUES (?, ?, ?)",
+            [
+                ("2026-09-29T00:00:00+00:00", "2026-09-29T12:00:00+00:00", "user_disabled"),
+                ("2026-09-29T13:00:00+00:00", "2026-09-30T00:00:00+00:00", "user_disabled"),
+                ("2026-09-30T00:00:00+00:00", None, None),
+            ],
+        )
+        store.commit()
+        _insert_local_day(
+            store,
+            "2026-09-29",
+            utc_start="2026-09-29T00:00:00+00:00",
+            utc_end="2026-09-30T00:00:00+00:00",
+            complete=True,
+        )
+
+        result = compute_projection(store, _clock())
+
+        assert result.confidence_state == ConfidenceState.UNSUPPORTED
+        assert any("full local observation day" in fact
+                   for fact in result.contributing_facts)
+
+    def test_partial_but_trusted_local_activity_opens_gate(self, store):
+        """Known local intervals can coexist with an incomplete day total."""
+        _insert_baseline(store)
+        _insert_segment(store)
+        _open_period(store)
+        for i in range(3):
+            day = (datetime(2026, 9, 27) + timedelta(days=i)).strftime("%Y-%m-%d")
+            _insert_day(store, day, bw=1024 * 1024 * 100)
+        _insert_sample(store, "2026-09-30T10:00:00+00:00", pu=5)
+        _insert_local_day(store, "2026-09-29", complete=True)
+        store.execute(
+            "UPDATE local_days SET activity_incomplete = 1 "
+            "WHERE local_date = '2026-09-29'"
+        )
+
+        result = compute_projection(store, _clock())
+
+        assert not any("full local observation day" in fact
+                       for fact in result.contributing_facts)
+
 
 # ---------------------------------------------------------------------------
 # Gate-2: One complete local day → Limited confidence

tests/test_dashboard_design.py

@@ -102,6 +102,7 @@ async def test_tabs_date_entry_and_hourly_inspection_keep_context(dashboard):
         app.on_refresh_tick()
         await pilot.pause()
         assert str(app.query_one("#bar-readout").render()) == readout
+        assert "Reads" in str(app.query_one("#bar-legend").render())
         await pilot.press("g")
         await pilot.press(*list("2026-09-17"))
         await pilot.press("escape")

tests/test_init_system.py

@@ -685,11 +685,19 @@ class TestEdgeCases:
 
     def test_runit_collect_missing_script(self):
         """Runit collect exits when fenris-collect script not found."""
-        with patch("fenris.init_system.Path") as mock_path:
-            mock_path.return_value.exists.return_value = False
+        with patch("fenris.init_system.subprocess") as mock_sub, \
+             patch("fenris.init_system.Path") as mock_path:
+            missing = MagicMock()
+            missing.exists.return_value = False
+            # Fallback path is built via Path(...).parent... / "src" / ...
+            missing.parent = missing
+            missing.__truediv__.return_value = missing
+            mock_path.return_value = missing
+            mock_sub.run.return_value = MagicMock(returncode=1)
             with pytest.raises(SystemExit) as exc_info:
                 _runit_collect()
             assert exc_info.value.code == 1
+            mock_sub.run.assert_not_called()
 
     def test_systemd_query_state_timeout(self):
         """Systemd query state handles subprocess timeout."""

tests/test_issue_100.py

@@ -0,0 +1,341 @@
+"""Normal-collection retention acceptance tests for issue #100."""
+import sqlite3
+import sys
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from fenris.collector import run_collection
+from fenris.local_day import query_local_day_summary
+from fenris.pruning import prune_old_samples
+from fenris.status import read_status
+from fenris.store import init_store
+
+
+@pytest.fixture
+def sysfs_controller(tmp_path):
+    controller = tmp_path / "sys" / "class" / "nvme" / "nvme0"
+    controller.mkdir(parents=True)
+    (controller / "subsysnqn").write_text("nqn.example:drive-1\n")
+    (controller / "model").write_text("Fenris Test Drive\n")
+    (controller / "serial").write_text("drive-1\n")
+    (controller / "firmware_rev").write_text("1.0\n")
+    transport = controller / "transport"
+    transport.mkdir()
+    (transport / "address").write_text("0000:00:01.0\n")
+    (transport / "trstring").write_text("pcie\n")
+    return controller
+
+
+class FakeClock:
+    def __init__(self, now):
+        self.now = now
+
+    def utcnow(self):
+        return self.now
+
+
+def _smartctl(written, read):
+    return {
+        "nvme_smart_health_information_log": {
+            "critical_warning": 0,
+            "temperature": 35,
+            "available_spare": 100,
+            "percentage_used": 1,
+            "data_units_written": written,
+            "data_units_read": read,
+            "power_on_hours": 100,
+            "power_cycles": 1,
+            "unsafe_shutdowns": 0,
+            "media_errors": 0,
+        },
+        "user_capacity": {"bytes": 1_000_000_000_000},
+        "model_name": "Fenris Test Drive",
+        "serial_number": "drive-1",
+        "firmware_version": "1.0",
+    }
+
+
+def _collect(config, controller, now, written, read):
+    return run_collection(
+        smartctl_data=_smartctl(written, read),
+        sysfs_path=controller,
+        config=config,
+        clock=FakeClock(now),
+    )
+
+
+def test_normal_collection_expires_detail_and_keeps_local_day_evidence(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    monkeypatch.setenv("TZ", "UTC")
+    store_path = tmp_path / "observations.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    first_time = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    second_time = first_time + timedelta(days=15)
+
+    first = _collect(config, sysfs_controller, first_time, 1000, 2000)
+    assert first["ok"] is True, first
+    second = _collect(config, sysfs_controller, second_time, 1001, 2003)
+    assert second["ok"] is True, second
+
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT ts FROM samples ORDER BY id").fetchall() == [
+            (second_time.isoformat(),)
+        ]
+        assert conn.execute(
+            "SELECT bytes_written, bytes_read, reason "
+            "FROM local_day_unallocated_evidence"
+        ).fetchall() == [(512_000, 1_536_000, "local_midnight")]
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM controller_segments").fetchone()[0] == 1
+        assert conn.execute("SELECT COUNT(*) FROM day_aggregates").fetchone()[0] == 15
+
+    monkeypatch.setenv("TZ", "Asia/Kolkata")
+    with read_status(store_path, second_time, query_services=False) as (reader, state):
+        assert reader is not None
+        assert state.store_fault is None
+        old_day = query_local_day_summary(
+            reader, "2026-09-01", "UTC", second_time,
+        )
+        assert old_day["utc_start"] == "2026-09-01T00:00:00+00:00"
+        assert old_day["utc_end"] == "2026-09-02T00:00:00+00:00"
+        assert old_day["bytes_written"] is None
+        assert old_day["shared_bytes_written"] == 512_000
+        assert old_day["shared_bytes_read"] == 1_536_000
+        assert old_day["shared_evidence_count"] == 1
+
+
+def test_pruning_keeps_samples_when_replacement_evidence_is_missing(tmp_path):
+    store_path = tmp_path / "incomplete.db"
+    conn = init_store(store_path)
+    now = datetime(2026, 9, 30, tzinfo=timezone.utc)
+    old_time = now - timedelta(days=20)
+    next_time = now - timedelta(days=1)
+    conn.execute(
+        "INSERT INTO controller_segments (id, opened_at, identity_key) "
+        "VALUES (1, ?, 'test')",
+        (old_time.isoformat(),),
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+        "VALUES (?, '/dev/nvme0', 100, 100, 1, 'UTC')",
+        (old_time.isoformat(),),
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+        "VALUES (?, '/dev/nvme0', 200, 200, 1, 'UTC')",
+        (next_time.isoformat(),),
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, now) == 0
+    assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 2
+    conn.close()
+
+
+def test_later_same_day_activity_does_not_replace_gap_interval(tmp_path):
+    store_path = tmp_path / "unrelated-local-activity.db"
+    conn = init_store(store_path)
+    gap_start = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    gap_end = gap_start + timedelta(minutes=10)
+    later_end = gap_start + timedelta(minutes=20)
+    now = datetime(2026, 9, 30, tzinfo=timezone.utc)
+    conn.execute(
+        "INSERT INTO controller_segments (id, opened_at, identity_key) "
+        "VALUES (1, ?, 'test')",
+        (gap_start.isoformat(),),
+    )
+    for sample_time, written, read in (
+        (gap_start, 1000, 100),
+        (gap_end, 2000, 200),
+        (later_end, 2100, 210),
+    ):
+        conn.execute(
+            "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+            "VALUES (?, '/dev/nvme0', ?, ?, 1, 'UTC')",
+            (sample_time.isoformat(), written, read),
+        )
+
+    # UTC aggregates can replace the old counter pair, but the local-day
+    # summary contains only the later B->C interval, after monitoring resumed.
+    conn.execute(
+        "INSERT INTO hour_observations (hour, bytes_written_delta, bytes_read_delta) "
+        "VALUES ('2026-09-01T00:00:00+00:00', 1000, 100)"
+    )
+    conn.execute(
+        "INSERT INTO day_aggregates (day) VALUES ('2026-09-01')"
+    )
+    conn.execute(
+        "INSERT INTO monitoring_periods (started_at, ended_at) VALUES (?, ?)",
+        (gap_end.isoformat(), later_end.isoformat()),
+    )
+    last_sample_id = conn.execute("SELECT MAX(id) FROM samples").fetchone()[0]
+    conn.execute(
+        "INSERT INTO local_days "
+        "(local_date, tz_name, tz_offset, utc_start, utc_end, bytes_written, "
+        " bytes_read, activity_intervals, activity_precision, last_sample_id) "
+        "VALUES ('2026-09-01', 'UTC', '+00:00', ?, ?, 100, 10, 1, 'measured', ?)",
+        (
+            gap_start.isoformat(),
+            (gap_start + timedelta(days=1)).isoformat(),
+            last_sample_id,
+        ),
+    )
+    local_day_id = conn.execute("SELECT MAX(id) FROM local_days").fetchone()[0]
+    conn.execute(
+        "INSERT INTO local_day_segment_totals "
+        "(local_day_id, segment_id, bytes_written, bytes_read, activity_intervals) "
+        "VALUES (?, 1, 100, 10, 1)",
+        (local_day_id,),
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, now) == 0
+    assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 3
+    conn.close()
+
+
+def test_pruning_keeps_oldest_unpaired_sample_as_successor_anchor(tmp_path):
+    store_path = tmp_path / "anchor.db"
+    conn = init_store(store_path)
+    now = datetime(2026, 9, 30, tzinfo=timezone.utc)
+    old_time = now - timedelta(days=20)
+    conn.execute(
+        "INSERT INTO controller_segments (id, opened_at, identity_key) "
+        "VALUES (1, ?, 'test')",
+        (old_time.isoformat(),),
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+        "VALUES (?, '/dev/nvme0', 100, 100, 1, 'UTC')",
+        (old_time.isoformat(),),
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, now) == 0
+    assert conn.execute("SELECT ts FROM samples").fetchone()[0] == old_time.isoformat()
+    conn.close()
+
+
+def test_pruning_never_expires_pending_publications(tmp_path):
+    store_path = tmp_path / "pending.db"
+    conn = init_store(store_path)
+    old_time = datetime(2026, 9, 1, tzinfo=timezone.utc)
+    conn.execute(
+        "INSERT INTO pending_publications (sample_ts, payload) VALUES (?, '{}')",
+        (old_time.isoformat(),),
+    )
+    conn.commit()
+
+    assert prune_old_samples(
+        conn, datetime(2026, 9, 30, tzinfo=timezone.utc),
+    ) == 0
+    assert conn.execute(
+        "SELECT sample_ts FROM pending_publications"
+    ).fetchone()[0] == old_time.isoformat()
+    conn.close()
+
+
+def test_pruning_preserves_unparseable_legacy_timestamp(tmp_path):
+    store_path = tmp_path / "legacy.db"
+    conn = init_store(store_path)
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read) "
+        "VALUES ('legacy timestamp', '/dev/nvme0', 100, 100)"
+    )
+    conn.commit()
+
+    assert prune_old_samples(conn, datetime(2026, 9, 30, tzinfo=timezone.utc)) == 0
+    assert conn.execute("SELECT ts FROM samples").fetchone()[0] == "legacy timestamp"
+    conn.close()
+
+
+def test_collection_defers_interrupted_pruning_and_retries_next_run(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    monkeypatch.setenv("TZ", "UTC")
+    store_path = tmp_path / "interrupted.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    first_time = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    second_time = first_time + timedelta(days=15)
+    assert _collect(config, sysfs_controller, first_time, 1000, 2000)["ok"] is True
+
+    writer = sqlite3.connect(store_path)
+    writer.execute(
+        "CREATE TRIGGER stop_retention BEFORE DELETE ON samples "
+        "BEGIN SELECT RAISE(ABORT, 'simulated retention interruption'); END"
+    )
+    writer.commit()
+    writer.close()
+
+    second = _collect(config, sysfs_controller, second_time, 1001, 2001)
+    assert second["ok"] is True, second
+    assert "simulated retention interruption" in second["retention_error"]
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 2
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM local_day_unallocated_evidence").fetchone()[0] == 1
+
+    writer = sqlite3.connect(store_path)
+    writer.execute("DROP TRIGGER stop_retention")
+    writer.commit()
+    writer.close()
+
+    third_time = second_time + timedelta(days=15)
+    third = _collect(config, sysfs_controller, third_time, 1002, 2002)
+    assert third["ok"] is True, third
+    assert third["retention_error"] is None
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT ts FROM samples ORDER BY id").fetchall() == [
+            (third_time.isoformat(),)
+        ]
+
+
+def test_collection_recovers_old_pending_work_before_pruning(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    monkeypatch.setenv("TZ", "UTC")
+    store_path = tmp_path / "pending-recovery.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    first_time = datetime(2026, 9, 1, 0, 0, tzinfo=timezone.utc)
+    second_time = first_time + timedelta(days=15)
+    assert _collect(config, sysfs_controller, first_time, 1000, 2000)["ok"] is True
+
+    writer = sqlite3.connect(store_path)
+    writer.execute(
+        "CREATE TRIGGER interrupt_publication "
+        "BEFORE INSERT ON local_day_unallocated_evidence "
+        "BEGIN SELECT RAISE(ABORT, 'simulated publication interruption'); END"
+    )
+    writer.commit()
+    writer.close()
+
+    failed = _collect(config, sysfs_controller, second_time, 1001, 2001)
+    assert failed["ok"] is False
+    assert "simulated publication interruption" in failed["error"]
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 1
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 1
+
+    writer = sqlite3.connect(store_path)
+    writer.execute("DROP TRIGGER interrupt_publication")
+    writer.commit()
+    writer.close()
+
+    third_time = second_time + timedelta(minutes=5)
+    recovered = _collect(config, sysfs_controller, third_time, 1002, 2002)
+    assert recovered["ok"] is True, recovered
+    assert recovered["retention_error"] is None
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+        assert conn.execute("SELECT ts FROM samples ORDER BY id").fetchall() == [
+            (second_time.isoformat(),),
+            (third_time.isoformat(),),
+        ]
+        assert conn.execute(
+            "SELECT COUNT(*) FROM local_day_unallocated_evidence"
+        ).fetchone()[0] == 1

tests/test_issue_101.py

@@ -0,0 +1,257 @@
+"""Bound pending-publication admission before device acquisition."""
+import json
+import sqlite3
+import sys
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from fenris import collector
+
+
+class FakeClock:
+    def __init__(self, now):
+        self.now = now
+
+    def utcnow(self):
+        return self.now
+
+
+@pytest.fixture
+def sysfs_controller(tmp_path):
+    controller_path = tmp_path / "sys" / "class" / "nvme" / "nvme0"
+    controller_path.mkdir(parents=True)
+    (controller_path / "subsysnqn").write_text("nqn.test:drive\n")
+    (controller_path / "model").write_text("Test NVMe\n")
+    (controller_path / "serial").write_text("test-serial\n")
+    (controller_path / "firmware_rev").write_text("1.0\n")
+    transport = controller_path / "transport"
+    transport.mkdir()
+    (transport / "trstring").write_text("pcie\n")
+    return controller_path
+
+
+def _smartctl(units_written):
+    return {
+        "nvme_smart_health_information_log": {
+            "critical_warning": 0,
+            "temperature": 35,
+            "available_spare": 100,
+            "percentage_used": 5,
+            "data_units_written": units_written,
+            "data_units_read": 500,
+            "power_on_hours": 100,
+        },
+        "user_capacity": {"bytes": 1_024_000_000_000},
+        "model_name": "Test NVMe",
+        "serial_number": "test-serial",
+        "firmware_version": "1.0",
+    }
+
+
+def _collect(config, controller, now, units_written):
+    return collector.run_collection(
+        config=config,
+        clock=FakeClock(now),
+        acquire=lambda: (_smartctl(units_written), controller),
+    )
+
+
+def test_partial_recovery_frees_admission_and_keeps_pending_order(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    """A freed slot admits one observation behind older pending evidence."""
+    assert collector.PENDING_PUBLICATION_LIMIT == 6_720
+    monkeypatch.setenv("TZ", "UTC")
+    monkeypatch.setattr(collector, "PENDING_PUBLICATION_LIMIT", 2)
+
+    store_path = tmp_path / "observations.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    old_time = datetime(2026, 8, 1, tzinfo=timezone.utc)
+    controller = sysfs_controller
+
+    assert _collect(config, controller, old_time, 1000)["ok"] is True
+
+    original_derive = collector.derive_hours_from_interval
+
+    def unavailable_derivation(*_args):
+        raise RuntimeError("derivation unavailable")
+
+    monkeypatch.setattr(collector, "derive_hours_from_interval", unavailable_derivation)
+    first_pending = _collect(
+        config, controller, old_time + timedelta(minutes=5), 1010,
+    )
+    assert first_pending["ok"] is False
+
+    second_pending = _collect(
+        config, controller, old_time + timedelta(minutes=10), 1020,
+    )
+    assert second_pending["ok"] is False
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute(
+            "SELECT COUNT(*) FROM pending_publications"
+        ).fetchone()[0] == 2
+
+    called = False
+
+    def must_not_acquire():
+        nonlocal called
+        called = True
+        return _smartctl(1030), controller
+
+    full = collector.run_collection(
+        config=config,
+        clock=FakeClock(old_time + timedelta(days=40)),
+        acquire=must_not_acquire,
+    )
+    assert full["ok"] is False
+    assert "capacity full (2 observations)" in full["error"]
+    assert called is False
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute(
+            "SELECT COUNT(*), MIN(sample_ts) FROM pending_publications"
+        ).fetchone() == (
+            2,
+            (old_time + timedelta(minutes=5)).isoformat(),
+        )
+        assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 1
+
+    def fail_second_pending(conn, previous, current):
+        if current["data_units_written"] == 1020:
+            raise RuntimeError("second observation still blocked")
+        return original_derive(conn, previous, current)
+
+    monkeypatch.setattr(collector, "derive_hours_from_interval", fail_second_pending)
+    acquired = False
+
+    def acquire_after_partial_recovery():
+        nonlocal acquired
+        acquired = True
+        return _smartctl(1030), controller
+
+    partial = collector.run_collection(
+        config=config,
+        clock=FakeClock(old_time + timedelta(days=40, minutes=5)),
+        acquire=acquire_after_partial_recovery,
+    )
+    assert partial["ok"] is False
+    assert acquired is True, partial
+    assert "no new observation acquired" not in partial["error"]
+
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute(
+            "SELECT data_units_written FROM samples ORDER BY id"
+        ).fetchall() == [(1000,), (1010,)]
+        queued_units = [
+            json.loads(row[0])["sample"]["data_units_written"]
+            for row in conn.execute(
+                "SELECT payload FROM pending_publications ORDER BY id"
+            )
+        ]
+        assert queued_units == [1020, 1030]
+        assert conn.execute("SELECT COUNT(*) FROM monitoring_periods").fetchone()[0] == 1
+
+    recovered_in_order = []
+
+    def record_recovery_order(conn, previous, current):
+        recovered_in_order.append(current["data_units_written"])
+        return original_derive(conn, previous, current)
+
+    monkeypatch.setattr(collector, "derive_hours_from_interval", record_recovery_order)
+    recovered = collector.run_collection(
+        config=config,
+        clock=FakeClock(old_time + timedelta(days=40, minutes=10)),
+        acquire=lambda: (_smartctl(1040), controller),
+    )
+    assert recovered["ok"] is True, recovered
+    assert recovered_in_order == [1020, 1030, 1040]
+
+    with sqlite3.connect(store_path) as conn:
+        assert conn.execute(
+            "SELECT data_units_written FROM samples ORDER BY id DESC LIMIT 1"
+        ).fetchone() == (1040,)
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+
+
+def test_actual_pending_limit_fails_before_acquisition_and_keeps_old_evidence(
+    tmp_path, sysfs_controller, monkeypatch,
+):
+    """The production limit retains every old row and refuses device acquisition."""
+    monkeypatch.setenv("TZ", "UTC")
+    store_path = tmp_path / "full-observations.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    controller = sysfs_controller
+    old_time = datetime(2026, 7, 1, tzinfo=timezone.utc)
+    assert _collect(config, controller, old_time, 1000)["ok"] is True
+
+    def unavailable_derivation(*_args):
+        raise RuntimeError("derivation unavailable")
+
+    monkeypatch.setattr(collector, "derive_hours_from_interval", unavailable_derivation)
+    failed = _collect(
+        config, controller, old_time + timedelta(minutes=5), 1010,
+    )
+    assert failed["ok"] is False
+
+    with sqlite3.connect(store_path) as conn:
+        payload_row = conn.execute(
+            "SELECT payload FROM pending_publications ORDER BY id LIMIT 1"
+        ).fetchone()
+        assert payload_row is not None
+        base = json.loads(payload_row[0])
+        base_observed_at = datetime.fromisoformat(base["sample"]["ts"])
+        base_units = base["sample"]["data_units_written"]
+        rows = []
+        for offset in range(1, collector.PENDING_PUBLICATION_LIMIT):
+            observation = json.loads(payload_row[0])
+            units_written = base_units + offset
+            observed_at = base_observed_at + timedelta(minutes=5 * offset)
+            observation["sample"]["ts"] = observed_at.isoformat()
+            observation["sample"]["data_units_written"] = units_written
+            observation["sample"]["bytes_written"] = units_written * 512_000
+            rows.append((observed_at.isoformat(), json.dumps(observation)))
+        conn.executemany(
+            "INSERT INTO pending_publications (sample_ts, payload) VALUES (?, ?)",
+            rows,
+        )
+        conn.commit()
+
+    called = False
+
+    def must_not_acquire():
+        nonlocal called
+        called = True
+        return _smartctl(9000), controller
+
+    result = collector.run_collection(
+        config=config,
+        clock=FakeClock(old_time + timedelta(days=40)),
+        acquire=must_not_acquire,
+    )
+    assert result["ok"] is False
+    assert f"capacity full ({collector.PENDING_PUBLICATION_LIMIT} observations)" in result[
+        "error"
+    ]
+    assert called is False
+
+    with sqlite3.connect(store_path) as conn:
+        pending_rows = conn.execute(
+            "SELECT sample_ts, payload FROM pending_publications ORDER BY id"
+        ).fetchall()
+        assert len(pending_rows) == collector.PENDING_PUBLICATION_LIMIT
+        assert (pending_rows[0][0], pending_rows[-1][0]) == (
+            (old_time + timedelta(minutes=5)).isoformat(),
+            (
+                old_time
+                + timedelta(minutes=5 * collector.PENDING_PUBLICATION_LIMIT)
+            ).isoformat(),
+        )
+        assert [
+            json.loads(payload)["sample"]["data_units_written"]
+            for _, payload in pending_rows
+        ] == list(range(1010, 1010 + collector.PENDING_PUBLICATION_LIMIT))
+        assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 1
+        assert conn.execute("SELECT COUNT(*) FROM monitoring_periods").fetchone()[0] == 1

tests/test_issue_72_edge_cases.py

@@ -90,8 +90,8 @@ def _insert_local_day(conn, local_date, tz_name="UTC", tz_offset="+00:00",
     conn.execute(
         "INSERT INTO local_days "
         "(local_date, tz_name, tz_offset, utc_start, utc_end, "
-        " bytes_written, bytes_read, coverage, sample_count, complete) "
-        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+        " bytes_written, bytes_read, coverage, sample_count, complete, activity_intervals) "
+        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 1)",
         (local_date, tz_name, tz_offset, utc_start, utc_end,
          bw, br, coverage, samples, complete),
     )
@@ -211,6 +211,8 @@ class TestStalenessFact:
         _insert_baseline(conn)
 
         base = datetime(2026, 9, 15, 12, 0, 0, tzinfo=timezone.utc)
+        ensure_period_open(conn, base)
+        conn.commit()
         for i in range(14):
             day = (base + timedelta(days=i)).strftime("%Y-%m-%d")
             _insert_day(conn, day, bw=10*1024*1024*1024)
@@ -234,6 +236,8 @@ class TestStalenessFact:
         _insert_baseline(conn)
 
         base = datetime(2026, 9, 18, 12, 0, 0, tzinfo=timezone.utc)
+        ensure_period_open(conn, base)
+        conn.commit()
         for i in range(14):
             day = (base + timedelta(days=i)).strftime("%Y-%m-%d")
             _insert_day(conn, day, bw=10*1024*1024*1024)

tests/test_issue_77.py

@@ -98,8 +98,8 @@ def _insert_local_day(conn, local_date, tz_name="UTC", tz_offset="+00:00",
     conn.execute(
         "INSERT INTO local_days "
         "(local_date, tz_name, tz_offset, utc_start, utc_end, "
-        " bytes_written, bytes_read, coverage, sample_count, complete) "
-        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+        " bytes_written, bytes_read, coverage, sample_count, complete, activity_intervals) "
+        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 1)",
         (local_date, tz_name, tz_offset, utc_start, utc_end,
          bw, br, coverage, samples, complete),
     )

tests/test_issue_92.py

@@ -24,13 +24,12 @@ from fenris.store import init_store
 from fenris.monitoring_periods import ensure_period_open
 from fenris.local_day import (
     LocalDaySummary,
+    local_day_boundaries,
     persist_local_day,
-    query_local_day_summary,
 )
+from fenris.tz_util import detect_system_tz
 from fenris.tui import (
     FenrisTuiApp,
-    HistoryGraph,
-    _RANGE_OPTIONS,
     _RANGE_DEFAULT,
 )
 
@@ -70,13 +69,18 @@ def _open_period(conn, start="2026-09-01T00:00:00+00:00"):
 
 
 def _insert_local_day(conn, local_date, tz_name="UTC", tz_offset="+00:00",
-                       bw=300, br=130, complete=True):
+                       bw=300, br=130, complete=True,
+                       utc_start=None, utc_end=None):
+    if utc_start is None:
+        utc_start = f"{local_date}T00:00:00+00:00"
+    if utc_end is None:
+        utc_end = f"{(datetime.fromisoformat(local_date) + timedelta(days=1)).strftime('%Y-%m-%d')}T00:00:00+00:00"
     summary = LocalDaySummary(
         local_date=local_date,
         tz_name=tz_name,
         tz_offset=tz_offset,
-        utc_start=f"{local_date}T00:00:00+00:00",
-        utc_end=f"{(datetime.fromisoformat(local_date) + timedelta(days=1)).strftime('%Y-%m-%d')}T00:00:00+00:00",
+        utc_start=utc_start,
+        utc_end=utc_end,
         bytes_written=bw,
         bytes_read=br,
         coverage=0.95,
@@ -124,6 +128,10 @@ def _make_app(tmp_path, clock=None):
     return app, patcher
 
 
+def _visible_selected_date(app):
+    return str(app.query_one("#local-day").render())[:10]
+
+
 # ---------------------------------------------------------------------------
 # AC92-1: Keyboard bindings exist
 # ---------------------------------------------------------------------------
@@ -178,14 +186,13 @@ class TestBracketNavigation:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
             # Move to an older day first so ] can advance
             for _ in range(3):
                 await pilot.press("left_square_bracket")
-            day_before = graph._day_data[graph.selected_index].get("day")
+            day_before = _visible_selected_date(app)
 
             await pilot.press("right_square_bracket")
-            day_after = graph._day_data[graph.selected_index].get("day")
+            day_after = _visible_selected_date(app)
             # ] should move to a newer day (closer to today)
             assert day_after > day_before
 
@@ -198,11 +205,10 @@ class TestBracketNavigation:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
-            day_before = graph._day_data[graph.selected_index].get("day")
+            day_before = _visible_selected_date(app)
 
             await pilot.press("left_square_bracket")
-            day_after = graph._day_data[graph.selected_index].get("day")
+            day_after = _visible_selected_date(app)
             # [ should move to an older day (further from today)
             assert day_after < day_before
 
@@ -224,9 +230,8 @@ class TestBracketNavigation:
                 for _ in range(29):
                     await pilot.press("left_square_bracket")
 
-                # Graph should now show a range that includes the oldest day
-                assert graph.selected_index >= 0
-                selected_day = graph._day_data[graph.selected_index].get("day")
+                # Visible readout identifies oldest selected date.
+                selected_day = _visible_selected_date(app)
                 oldest_day = (app_clock - timedelta(days=29)).strftime("%Y-%m-%d")
                 assert selected_day == oldest_day
         finally:
@@ -241,11 +246,10 @@ class TestBracketNavigation:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
-            initial_idx = graph.selected_index
+            initial_date = _visible_selected_date(app)
 
             await pilot.press("right_square_bracket")
-            assert graph.selected_index == initial_idx
+            assert _visible_selected_date(app) == initial_date
 
 
 # ---------------------------------------------------------------------------
@@ -262,36 +266,54 @@ class TestArrowInspection:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
             await pilot.press("v", "v")
-            graph.focus()
+            app.query_one("#usage-history").focus()
             await pilot.pause()
-            day_before = graph._day_data[graph.selected_index].get("day")
+            day_before = _visible_selected_date(app)
 
             await pilot.press("left")
-            day_after = graph._day_data[graph.selected_index].get("day")
+            day_after = _visible_selected_date(app)
             assert day_after != day_before
 
     @pytest.mark.asyncio
     async def test_arrow_keys_inspect_hours_in_drill(self, tmp_path):
         """Arrow keys navigate hours when in hourly drill-down."""
         conn, clock = _setup_store(tmp_path, n_days=14)
+        local_date = datetime.now().astimezone().date().isoformat()
+        timezone_name = detect_system_tz()
+        utc_start, utc_end, tz_offset = local_day_boundaries(local_date, timezone_name)
+        _insert_local_day(
+            conn, local_date, tz_name=timezone_name, tz_offset=tz_offset,
+            utc_start=utc_start.isoformat(), utc_end=utc_end.isoformat(),
+        )
+        hour = utc_start.replace(minute=0, second=0, microsecond=0)
+        if hour < utc_start:
+            hour += timedelta(hours=1)
+        while hour + timedelta(hours=1) <= utc_end:
+            conn.execute(
+                "INSERT INTO hour_observations "
+                "(hour, bytes_written_delta, bytes_read_delta, coverage, sample_count) "
+                "VALUES (?, 1000000, 2000000, 1, 1)",
+                (hour.isoformat(),),
+            )
+            hour += timedelta(hours=1)
+        conn.commit()
         conn.close()
 
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
             await pilot.press("v", "v")
             await pilot.click("#usage-history")
 
             # Enter drill-down
             await pilot.press("enter")
-            assert graph.view_mode == "hourly"
+            assert app.query_one("#activity-tabs").active == "view-day"
 
             # Arrow keys navigate hours
+            readout_before = str(app.query_one("#bar-readout").render())
             await pilot.press("left")
-            assert graph._hourly_selected >= 0
+            assert str(app.query_one("#bar-readout").render()) != readout_before
 
 
 # ---------------------------------------------------------------------------
@@ -308,18 +330,17 @@ class TestTodayBinding:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
 
             # Browse to an older day using [
             for _ in range(5):
                 await pilot.press("left_square_bracket")
-            browsed_day = graph._day_data[graph.selected_index].get("day")
-            today = clock.strftime("%Y-%m-%d")
+            browsed_day = _visible_selected_date(app)
+            today = datetime.now().astimezone().date().isoformat()
             assert browsed_day != today
 
             # Press t to return to today
             await pilot.press("t")
-            assert graph.selected_index == len(graph._day_data) - 1
+            assert app.query_one("#activity-tabs").active == "view-live"
 
     @pytest.mark.asyncio
     async def test_t_exits_hourly_drill(self, tmp_path):
@@ -330,17 +351,16 @@ class TestTodayBinding:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
             await pilot.press("v", "v")
             await pilot.click("#usage-history")
 
             # Enter drill
             await pilot.press("enter")
-            assert graph.view_mode == "hourly"
+            assert app.query_one("#activity-tabs").active == "view-day"
 
             # t returns to daily
             await pilot.press("t")
-            assert graph.view_mode == "daily"
+            assert app.query_one("#activity-tabs").active == "view-live"
 
 
 # ---------------------------------------------------------------------------
@@ -360,7 +380,7 @@ class TestDateEntry:
             await pilot.press("g")
             await pilot.pause()
             # Date picker should be pushed as a screen
-            assert len(app.screen_stack) > 1
+            assert app.screen.query_one("#date-input") is not None
 
     @pytest.mark.asyncio
     async def test_valid_date_navigates(self, tmp_path):
@@ -375,13 +395,12 @@ class TestDateEntry:
                 await pilot.pause()
                 target_date = (app_clock - timedelta(days=5)).strftime("%Y-%m-%d")
 
-                # Simulate the date picker callback directly
-                app._go_to_date_callback(target_date)
+                await pilot.press("g")
+                await pilot.press(*target_date)
+                await pilot.press("enter")
                 await pilot.pause()
 
-                graph = app.query_one("#usage-history")
-                selected_day = graph._day_data[graph.selected_index].get("day")
-                assert selected_day == target_date
+                assert _visible_selected_date(app) == target_date
         finally:
             patcher.stop()
 
@@ -400,7 +419,8 @@ class TestDateEntry:
             await pilot.press("enter")
             await pilot.pause()
             # Should remain on date picker (error shown, not dismissed)
-            assert len(app.screen_stack) > 1
+            error = str(app.screen.query_one("#date-error").render())
+            assert "Invalid date format" in error
 
     @pytest.mark.asyncio
     async def test_cancel_returns_to_prior(self, tmp_path):
@@ -411,8 +431,7 @@ class TestDateEntry:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
-            prior_idx = graph.selected_index
+            prior_readout = str(app.query_one("#bar-readout").render())
 
             await pilot.press("g")
             await pilot.pause()
@@ -420,7 +439,7 @@ class TestDateEntry:
             await pilot.pause()
 
             # Should be back to the graph with the same selection
-            assert graph.selected_index == prior_idx
+            assert str(app.query_one("#bar-readout").render()) == prior_readout
 
 
 # ---------------------------------------------------------------------------
@@ -442,8 +461,6 @@ class TestLocalDayEvidence:
         try:
             async with app.run_test(size=(100, 40)) as pilot:
                 await pilot.pause()
-                graph = app.query_one("#usage-history")
-
                 # Navigate to the target day
                 for _ in range(4):
                     await pilot.press("left_square_bracket")
@@ -489,12 +506,10 @@ class TestBrowsingStability:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=0.1)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
-
             # Browse to a specific day
             for _ in range(3):
-                await pilot.press("right_square_bracket")
-            browsed_day = graph._day_data[graph.selected_index].get("day")
+                await pilot.press("left_square_bracket")
+            browsed_day = _visible_selected_date(app)
 
             # Wait for at least one refresh tick
             import asyncio
@@ -502,8 +517,7 @@ class TestBrowsingStability:
             await pilot.pause()
 
             # Selection should be preserved
-            assert graph.selected_index >= 0
-            assert graph._day_data[graph.selected_index].get("day") == browsed_day
+            assert _visible_selected_date(app) == browsed_day
 
     @pytest.mark.asyncio
     async def test_navigate_away_does_not_jump_on_refresh(self, tmp_path):
@@ -514,14 +528,12 @@ class TestBrowsingStability:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=0.1)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
-
             # Browse to older day using [
             for _ in range(5):
                 await pilot.press("left_square_bracket")
 
-            browsed_day = graph._day_data[graph.selected_index].get("day")
-            today = clock.strftime("%Y-%m-%d")
+            browsed_day = _visible_selected_date(app)
+            today = datetime.now().astimezone().date().isoformat()
             assert browsed_day != today
 
             # Wait for refresh
@@ -530,8 +542,8 @@ class TestBrowsingStability:
             await pilot.pause()
 
             # Should NOT jump back to today
-            selected_day = graph._day_data[graph.selected_index].get("day")
-            assert selected_day != today
+            assert _visible_selected_date(app) == browsed_day
+            assert _visible_selected_date(app) != today
 
 
 # ---------------------------------------------------------------------------
@@ -548,17 +560,16 @@ class TestDrillDownData:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(100, 40)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
             await pilot.press("v", "v")
             await pilot.click("#usage-history")
 
             # Enter drill-down
             await pilot.press("enter")
-            assert graph.view_mode == "hourly"
+            assert app.query_one("#activity-tabs").active == "view-day"
 
-            # After drill callback, data should be populated
-            # (empty since we have no hour_observations)
-            assert graph.view_mode == "hourly"
+            assert "hourly evidence unavailable" in str(
+                app.query_one("#bar-readout").render()
+            )
 
 
 # ---------------------------------------------------------------------------
@@ -575,11 +586,10 @@ class TestConstrainedWidth:
         app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
         async with app.run_test(size=(80, 24)) as pilot:
             await pilot.pause()
-            graph = app.query_one("#usage-history")
-            day_before = graph._day_data[graph.selected_index].get("day")
+            day_before = _visible_selected_date(app)
 
             await pilot.press("left_square_bracket")
-            day_after = graph._day_data[graph.selected_index].get("day")
+            day_after = _visible_selected_date(app)
             assert day_after < day_before
 
     @pytest.mark.asyncio
@@ -593,4 +603,235 @@ class TestConstrainedWidth:
             await pilot.pause()
             await pilot.press("g")
             await pilot.pause()
-            assert len(app.screen_stack) > 1
+            assert app.screen.query_one("#date-input") is not None
+
+
+class TestClickableNavigation:
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("size", [(80, 24), (70, 20)])
+    async def test_click_and_keyboard_select_same_local_date(self, tmp_path, size):
+        """Clickable and keyboard navigation share date-entry transitions."""
+        app_clock = _clock()
+        conn, _ = _setup_store(tmp_path, n_days=14, today=app_clock)
+        conn.close()
+        app, patcher = _make_app(tmp_path, clock=app_clock)
+        expected = (app_clock - timedelta(days=1)).strftime("%Y-%m-%d")
+        direct = (app_clock - timedelta(days=5)).strftime("%Y-%m-%d")
+        try:
+            async with app.run_test(size=size) as pilot:
+                await pilot.press("left_square_bracket")
+                keyboard_date = _visible_selected_date(app)
+                assert keyboard_date == expected
+
+                await pilot.press("t")
+                await pilot.click("#activity-tools", offset=(2, 0))
+                assert _visible_selected_date(app) == keyboard_date
+
+                await pilot.press("g")
+                await pilot.press(*direct)
+                await pilot.press("enter")
+                await pilot.pause()
+                assert _visible_selected_date(app) == direct
+        finally:
+            patcher.stop()
+
+
+class TestHistoricalSelectionIdentity:
+    @pytest.mark.asyncio
+    async def test_unavailable_entered_date_stays_visible_after_refresh(self, tmp_path):
+        """An unavailable requested date remains selected and explained."""
+        conn, clock = _setup_store(tmp_path, n_days=30)
+        conn.close()
+        app, patcher = _make_app(tmp_path, clock=clock)
+        target = (clock - timedelta(days=120)).strftime("%Y-%m-%d")
+        try:
+            async with app.run_test(size=(100, 40)) as pilot:
+                await pilot.press("g")
+                await pilot.press(*target)
+                await pilot.press("enter")
+                await pilot.pause()
+
+                local_day = str(app.query_one("#local-day").render())
+                assert target in local_day
+                assert "unavailable" in local_day
+
+                app.on_refresh_tick()
+                await pilot.pause()
+                local_day = str(app.query_one("#local-day").render())
+                assert target in local_day
+                assert "unavailable" in local_day
+        finally:
+            patcher.stop()
+
+    @pytest.mark.asyncio
+    async def test_retained_summary_outside_recent_window_remains_available(self, tmp_path):
+        """Direct date entry can inspect durable local summaries beyond 90 days."""
+        clock = _clock()
+        conn, _ = _setup_store(tmp_path, n_days=30, today=clock)
+        target = (clock - timedelta(days=120)).strftime("%Y-%m-%d")
+        utc_start, utc_end, tz_offset = local_day_boundaries(
+            target, "Asia/Kolkata",
+        )
+        _insert_local_day(
+            conn,
+            target,
+            tz_name="Asia/Kolkata",
+            tz_offset=tz_offset,
+            bw=2_000_000_000,
+            br=1_000_000_000,
+            utc_start=utc_start.isoformat(),
+            utc_end=utc_end.isoformat(),
+        )
+        conn.close()
+        app, patcher = _make_app(tmp_path, clock=clock)
+        try:
+            async with app.run_test(size=(100, 40)) as pilot:
+                await pilot.press("g")
+                await pilot.press(*target)
+                await pilot.press("enter")
+                await pilot.pause()
+                await pilot.press("v")
+                await pilot.pause()
+
+                readout = str(app.query_one("#bar-readout").render())
+                assert target in readout
+                assert "Asia/Kolkata +05:30" in readout
+                assert "W 2.000 GB known · R 1.000 GB known" in readout
+        finally:
+            patcher.stop()
+
+    @pytest.mark.asyncio
+    async def test_constrained_history_keeps_known_totals_and_evidence_state(self, tmp_path):
+        """Small textual view retains selected day, totals, and shared evidence."""
+        clock = _clock()
+        conn, _ = _setup_store(tmp_path, n_days=14, today=clock)
+        target = (clock - timedelta(days=2)).strftime("%Y-%m-%d")
+        utc_start, utc_end, tz_offset = local_day_boundaries(
+            target, "Asia/Kolkata",
+        )
+        _insert_local_day(
+            conn,
+            target,
+            tz_name="Asia/Kolkata",
+            tz_offset=tz_offset,
+            bw=1_500_000_000,
+            br=500_000_000,
+            utc_start=utc_start.isoformat(),
+            utc_end=utc_end.isoformat(),
+        )
+        conn.execute(
+            "INSERT INTO local_day_unallocated_evidence "
+            "(start_sample_id, end_sample_id, start_local_date, end_local_date, "
+            " start_tz_name, end_tz_name, started_at, ended_at, bytes_written, "
+            " bytes_read, reason) VALUES (1, 2, ?, ?, ?, ?, ?, ?, ?, ?, 'local_midnight')",
+            (target, "2026-09-29", "Asia/Kolkata", "Asia/Kolkata",
+             utc_start.isoformat(), utc_start.isoformat(), 200_000_000, 300_000_000),
+        )
+        conn.commit()
+        conn.close()
+        app, patcher = _make_app(tmp_path, clock=clock)
+        try:
+            async with app.run_test(size=(70, 20)) as pilot:
+                await pilot.press("g")
+                await pilot.press(*target)
+                await pilot.press("enter")
+                await pilot.press("v")
+                await pilot.pause()
+
+                readout = str(app.query_one("#bar-readout").render())
+                assert target in readout
+                assert "Asia/Kolkata +05:30" in readout
+                assert "incomplete" in readout
+                assert "W 1.500 GB known · R 0.500 GB known" in readout
+                assert "shared at midnight W 0.200 GB · R 0.300 GB" in readout
+        finally:
+            patcher.stop()
+
+    @pytest.mark.asyncio
+    async def test_date_picker_keystrokes_do_not_toggle_measurement(self, tmp_path):
+        """Typing in date entry does not trigger dashboard shortcuts."""
+        conn, _ = _setup_store(tmp_path, n_days=14)
+        conn.close()
+        app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
+        async with app.run_test(size=(100, 40)) as pilot:
+            await pilot.pause()
+            await pilot.press("g")
+            await pilot.press("w")
+            await pilot.press("escape")
+            await pilot.pause()
+
+            header = str(app.query_one("#live-header").render())
+            assert "Writes" in header
+
+    @pytest.mark.asyncio
+    async def test_historical_hour_uses_recorded_timezone_and_utc_identity(self, tmp_path):
+        """Repeated local clock labels retain distinct UTC-hour identities."""
+        clock = _clock(2026, 11, 2, 12)
+        conn, _ = _setup_store(tmp_path, n_days=30, today=clock)
+        target = "2026-11-01"
+        _insert_local_day(
+            conn,
+            target,
+            tz_name="America/New_York",
+            tz_offset="-04:00",
+            utc_start="2026-11-01T04:00:00+00:00",
+            utc_end="2026-11-02T05:00:00+00:00",
+        )
+        conn.execute(
+            "INSERT INTO local_day_unallocated_evidence "
+            "(start_sample_id, end_sample_id, start_local_date, end_local_date, "
+            " start_tz_name, end_tz_name, started_at, ended_at, bytes_written, "
+            " bytes_read, reason) VALUES (1, 2, ?, ?, ?, ?, ?, ?, ?, ?, 'local_midnight')",
+            (target, "2026-11-02", "America/New_York", "America/New_York",
+             "2026-11-02T03:59:00+00:00", "2026-11-02T04:01:00+00:00",
+             2_000_000_000, 3_000_000_000),
+        )
+        for index in range(25):
+            hour = datetime(2026, 11, 1, 4, tzinfo=timezone.utc) + timedelta(hours=index)
+            conn.execute(
+                "INSERT INTO hour_observations "
+                "(hour, bytes_written_delta, bytes_read_delta, coverage, sample_count) "
+                "VALUES (?, ?, ?, 1, 1)",
+                (hour.isoformat(), (index + 1) * 1_000_000_000,
+                 (index + 1) * 2_000_000_000),
+            )
+        conn.execute(
+            "UPDATE hour_observations SET coverage = 0.5, unknown_seconds = 1800 "
+            "WHERE hour = ?",
+            ("2026-11-01T04:00:00+00:00",),
+        )
+        conn.commit()
+        conn.close()
+        app, patcher = _make_app(tmp_path, clock=clock)
+        try:
+            async with app.run_test(size=(100, 40)) as pilot:
+                await pilot.click("#view-history")
+                await pilot.press("g")
+                await pilot.press(*target)
+                await pilot.press("enter")
+                await pilot.pause()
+
+                local_day = str(app.query_one("#local-day").render())
+                assert "America/New_York -04:00" in local_day
+                assert "incomplete" in local_day
+                assert "shared at midnight W 2.000 GB · R 3.000 GB" in local_day
+
+                # Move from final local hour to the second 01:00 occurrence.
+                await pilot.press(*(["left"] * 22))
+                second_occurrence = str(app.query_one("#bar-readout").render())
+                assert "01:00 -0500" in second_occurrence
+                assert "America/New_York -0500" in second_occurrence
+                assert "06:00 UTC" in second_occurrence
+
+                await pilot.press("left")
+                first_occurrence = str(app.query_one("#bar-readout").render())
+                assert "01:00 -0400" in first_occurrence
+                assert "America/New_York -0400" in first_occurrence
+                assert "05:00 UTC" in first_occurrence
+
+                await pilot.press("left")
+                incomplete_hour = str(app.query_one("#bar-readout").render())
+                assert "00:00 -0400" in incomplete_hour
+                assert "50% coverage · incomplete" in incomplete_hour
+        finally:
+            patcher.stop()

tests/test_issue_93.py

@@ -151,8 +151,8 @@ class TestDetailExpiresSummariesRemain:
         # Run pruning
         pruned = prune_old_samples(conn, now, retention_days=14)
 
-        # Old samples should be pruned, recent ones retained
-        assert pruned >= 1
+        # The handcrafted totals lack replacement provenance, so retain samples.
+        assert pruned == 0
 
         # Local-day summaries must still be queryable
         old_result = query_local_day_summary(conn, "2026-09-16")
@@ -277,7 +277,7 @@ class TestTimezonePreservation:
         now = datetime(2026, 10, 1, 12, 0, 0, tzinfo=timezone.utc)
 
         old = LocalDaySummary(
-            local_date="2026-09-16", tz_name="US/Eastern", tz_offset="-04:00",
+            local_date="2026-09-16", tz_name="America/New_York", tz_offset="-04:00",
             utc_start="2026-09-16T04:00:00+00:00",
             utc_end="2026-09-17T04:00:00+00:00",
             bytes_written=5000, bytes_read=2000,
@@ -288,7 +288,7 @@ class TestTimezonePreservation:
         entries = query_local_day_history(conn, "2026-09-16", "2026-09-16", now)
         assert len(entries) == 1
         entry = entries[0]
-        assert entry.tz_name == "US/Eastern"
+        assert entry.tz_name == "America/New_York"
         assert entry.tz_offset == "-04:00"
         assert entry.utc_start == "2026-09-16T04:00:00+00:00"
         assert entry.detail_available is False
@@ -389,10 +389,10 @@ class TestIdempotencySafety:
 # ---------------------------------------------------------------------------
 
 class TestVolumeConservation:
-    """AC6: Summary volume plus unallocated evidence conserves each delta once."""
+    """AC6: UTC-hour totals never become fabricated local-day totals."""
 
-    def test_midnight_spanning_not_double_counted(self, tmp_path):
-        """Midnight-spanning interval appears once, not in both days."""
+    def test_hourly_bytes_are_unavailable_at_local_precision(self, tmp_path):
+        """UTC-day hour bytes cannot be split across local-day boundaries."""
         conn = init_store(tmp_path / "obs.db")
 
         # Insert hour observations that span midnight
@@ -409,13 +409,10 @@ class TestVolumeConservation:
         summary_17 = derive_local_day_summary(conn, "UTC", clock_17)
         assert summary_17 is not None
 
-        # Each day should only have its own hour's bytes
-        # Day 16 has the 23:00 hour (100 bw)
-        # Day 17 has the 00:00 hour (200 bw)
-        assert summary_16.bytes_written == 100
-        assert summary_17.bytes_written == 200
-        # Total is conserved: 100 + 200 = 300
-        assert summary_16.bytes_written + summary_17.bytes_written == 300
+        assert summary_16.bytes_written == 0
+        assert summary_17.bytes_written == 0
+        assert summary_16.activity_precision == "unavailable"
+        assert summary_17.activity_precision == "unavailable"
         conn.close()
 
 
@@ -437,10 +434,10 @@ class TestExistingEntryPoints:
             _insert_sample(conn, ts, bw=days_ago * 100)
 
         pruned = prune_old_samples(conn, now, retention_days=14)
-        assert pruned == 15
+        assert pruned == 0
 
         cursor = conn.execute("SELECT COUNT(*) FROM samples")
-        assert cursor.fetchone()[0] == 14
+        assert cursor.fetchone()[0] == 29
         conn.close()
 
     def test_repair_with_real_temp_store(self, tmp_path):

tests/test_issue_97.py

@@ -0,0 +1,294 @@
+"""Pending publication behavior from issue #97."""
+import sqlite3
+import sys
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from fenris import collector
+from fenris.projection import compute_projection
+from fenris.status import _query_drive_facts, get_status, read_status
+from fenris.status_composition import render_status_tui
+from fenris.store import init_store
+
+
+class FakeClock:
+    def __init__(self, now):
+        self.now = now
+
+    def utcnow(self):
+        return self.now
+
+
+@pytest.fixture
+def sysfs_fixture_tree(tmp_path):
+    ctrl_dir = tmp_path / "sys" / "class" / "nvme" / "nvme0"
+    ctrl_dir.mkdir(parents=True)
+    (ctrl_dir / "subsysnqn").write_text("nqn.test:drive\n")
+    (ctrl_dir / "model").write_text("Test NVMe\n")
+    (ctrl_dir / "serial").write_text("test-serial\n")
+    (ctrl_dir / "firmware_rev").write_text("1.0\n")
+    transport_dir = ctrl_dir / "transport"
+    transport_dir.mkdir()
+    (transport_dir / "trstring").write_text("pcie\n")
+    return tmp_path
+
+
+def _smartctl(unit_written, unit_read=500):
+    return {
+        "nvme_smart_health_information_log": {
+            "critical_warning": 0,
+            "temperature": 35,
+            "available_spare": 100,
+            "percentage_used": 5,
+            "data_units_written": unit_written,
+            "data_units_read": unit_read,
+            "power_on_hours": 100,
+        },
+        "user_capacity": {"bytes": 1_024_000_000_000},
+        "model_name": "Test NVMe",
+        "serial_number": "test-serial",
+        "firmware_version": "1.0",
+    }
+
+
+def _acquirer(sysfs_path, unit_written, unit_read=500):
+    return lambda: (_smartctl(unit_written, unit_read), sysfs_path)
+
+
+def _public_snapshot(store_path, clock_now):
+    with read_status(store_path, clock_now) as (conn, comp):
+        assert conn is not None
+        return {
+            "state": comp.state,
+            "freshness": comp.freshness,
+            "sample_count": comp.sample_count,
+            "day_count": comp.day_count,
+            "boot_enabled": comp.boot_enabled,
+            "timer_active": comp.timer_active,
+            "last_collect_ok": comp.last_collect_ok,
+            "latest_sample": tuple(conn.execute(
+                "SELECT ts, subnqn, mn, sn, fr, capacity_bytes, temperature_c, "
+                "available_spare, data_units_written, bytes_written, critical_warning, "
+                "media_errors, unsafe_shutdowns "
+                "FROM samples ORDER BY id DESC LIMIT 1"
+            ).fetchone()),
+            "drive_facts": tuple(_query_drive_facts(conn)),
+            "projection": repr(compute_projection(conn, clock_now)),
+            "segments": conn.execute("SELECT COUNT(*) FROM controller_segments").fetchone()[0],
+            "periods": conn.execute("SELECT COUNT(*) FROM monitoring_periods").fetchone()[0],
+            "hours": tuple(conn.execute(
+                "SELECT hour, bytes_written_delta, bytes_read_delta, sample_count "
+                "FROM hour_observations ORDER BY hour"
+            ).fetchall()),
+            "days": tuple(conn.execute(
+                "SELECT day, bytes_written_delta, bytes_read_delta "
+                "FROM day_aggregates ORDER BY day"
+            ).fetchall()),
+            "local_days": tuple(conn.execute(
+                "SELECT local_date, bytes_written, bytes_read, sample_count "
+                "FROM local_days ORDER BY local_date, tz_name"
+            ).fetchall()),
+            "pending": comp.pending_publication_count,
+        }
+
+
+def test_failed_publication_survives_restart_and_readers_keep_last_consistent_view(
+    tmp_path, sysfs_fixture_tree, monkeypatch
+):
+    """A failed derivation stays private, visible as pending, and replays in order."""
+    monkeypatch.setenv("TZ", "UTC")
+    service = {"boot_enabled": True, "timer_active": True, "last_collect_ok": True}
+    monkeypatch.setattr("fenris.status.query_service_state", lambda: service)
+    now = datetime.now(timezone.utc).replace(
+        minute=35, second=0, microsecond=0,
+    )
+    store_path = tmp_path / "observations.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    sysfs_path = sysfs_fixture_tree / "sys" / "class" / "nvme" / "nvme0"
+
+    for units, at in ((1000, now - timedelta(minutes=10)), (1010, now - timedelta(minutes=5))):
+        result = collector.run_collection(
+            config=config,
+            clock=FakeClock(at),
+            acquire=_acquirer(sysfs_path, units),
+        )
+        assert result["ok"] is True, result
+
+    previous = _public_snapshot(store_path, now)
+    baseline_cli = get_status(store_path, now, query_services=False, query_journal=False)
+    original_derive = collector.derive_hours_from_interval
+    reader_during_write = {}
+
+    def fail_derivation(conn, previous_sample, next_sample):
+        # Read while writer has inserted the unpublished sample but not committed.
+        with read_status(store_path, now, query_services=False) as (reader, comp):
+            assert reader is not None
+            reader_during_write["samples"] = reader.execute(
+                "SELECT COUNT(*) FROM samples"
+            ).fetchone()[0]
+            reader_during_write["pending"] = comp.pending_publication_count
+            reader_during_write["latest"] = tuple(reader.execute(
+                "SELECT data_units_written FROM samples ORDER BY id DESC LIMIT 1"
+            ).fetchone())
+        raise RuntimeError("injected non-invariant derivation failure")
+
+    monkeypatch.setattr(collector, "derive_hours_from_interval", fail_derivation)
+    failed = collector.run_collection(
+        config=config,
+        clock=FakeClock(now),
+        acquire=_acquirer(sysfs_path, 1025),
+    )
+    assert failed["ok"] is False
+    assert "injected non-invariant derivation failure" in failed["error"]
+    assert reader_during_write == {"samples": 2, "pending": 1, "latest": (1010,)}
+
+    after_failure = _public_snapshot(store_path, now)
+    assert {**after_failure, "pending": previous["pending"]} == previous
+    assert after_failure["pending"] == 1
+
+    cli_text = get_status(store_path, now, query_services=False, query_journal=False)
+    pending_fact = " · pending publication: 1 observation retained for retry"
+    assert pending_fact in cli_text
+    assert cli_text.replace(pending_fact, "", 1) == baseline_cli
+
+    with read_status(store_path, now) as (_, comp):
+        assert comp.state == previous["state"]
+        assert comp.boot_enabled is True
+        assert comp.timer_active is True
+        assert "pending publication: 1 observation retained for retry" in render_status_tui(comp).lower()
+
+    called = False
+
+    def should_wait_for_recovery():
+        nonlocal called
+        called = True
+        return _smartctl(1030), sysfs_path
+
+    monkeypatch.setattr(collector, "PENDING_PUBLICATION_LIMIT", 1)
+    blocked = collector.run_collection(
+        config=config,
+        clock=FakeClock(now + timedelta(minutes=5)),
+        acquire=should_wait_for_recovery,
+    )
+    assert blocked["ok"] is False
+    assert called is False
+    assert _public_snapshot(store_path, now)["pending"] == 1
+
+    monkeypatch.setattr(collector, "derive_hours_from_interval", original_derive)
+    monkeypatch.setenv("TZ", "Asia/Kolkata")
+    recovered = collector.run_collection(
+        config=config,
+        clock=FakeClock(now + timedelta(minutes=5)),
+        acquire=_acquirer(sysfs_path, 1040),
+    )
+    assert recovered["ok"] is True, recovered
+
+    conn = sqlite3.connect(store_path)
+    try:
+        assert conn.execute(
+            "SELECT data_units_written FROM samples ORDER BY id"
+        ).fetchall() == [(1000,), (1010,), (1025,), (1040,)]
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+        assert conn.execute(
+            "SELECT SUM(bytes_written_delta) FROM day_aggregates"
+        ).fetchone()[0] == (1040 - 1000) * 512_000
+        assert conn.execute(
+            "SELECT sample_count FROM local_days WHERE tz_name = 'UTC'"
+        ).fetchone()[0] == 4
+        assert conn.execute(
+            "SELECT COUNT(*) FROM local_days WHERE tz_name = 'Asia/Kolkata'"
+        ).fetchone()[0] == 1
+    finally:
+        conn.close()
+
+
+def test_invalid_sample_writes_no_sample_segment_period_or_pending_row(tmp_path, sysfs_fixture_tree):
+    """Sample invariants run before staging or public monitoring changes."""
+    now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+    store_path = tmp_path / "observations.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    sysfs_path = sysfs_fixture_tree / "sys" / "class" / "nvme" / "nvme0"
+    result = collector.run_collection(
+        config=config,
+        clock=FakeClock(now),
+        acquire=_acquirer(sysfs_path, -1),
+    )
+
+    assert result["ok"] is False
+    assert result["error_type"] == "InvariantViolationError"
+    conn = sqlite3.connect(store_path)
+    try:
+        assert conn.execute("SELECT COUNT(*) FROM samples").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM controller_segments").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM monitoring_periods").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+    finally:
+        conn.close()
+
+
+def test_pending_capacity_retries_before_skipping_acquisition(tmp_path, sysfs_fixture_tree, monkeypatch):
+    now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+    store_path = tmp_path / "observations.db"
+    config = {"device": "/dev/nvme0", "store_path": str(store_path)}
+    sysfs_path = sysfs_fixture_tree / "sys" / "class" / "nvme" / "nvme0"
+    first = collector.run_collection(
+        config=config,
+        clock=FakeClock(now - timedelta(minutes=5)),
+        acquire=_acquirer(sysfs_path, 1000),
+    )
+    assert first["ok"] is True, first
+
+    monkeypatch.setattr(collector, "PENDING_PUBLICATION_LIMIT", 1)
+
+    def fail_derivation(*_):
+        raise RuntimeError("derivation still unavailable")
+
+    monkeypatch.setattr(collector, "derive_hours_from_interval", fail_derivation)
+    failed = collector.run_collection(
+        config=config,
+        clock=FakeClock(now),
+        acquire=_acquirer(sysfs_path, 1010),
+    )
+    assert failed["ok"] is False
+    assert "capacity full (1 observations)" in failed["error"]
+
+    called = False
+
+    def must_not_acquire():
+        nonlocal called
+        called = True
+        return _smartctl(1020), sysfs_path
+
+    retry = collector.run_collection(
+        config=config,
+        clock=FakeClock(now + timedelta(minutes=5)),
+        acquire=must_not_acquire,
+    )
+    assert retry["ok"] is False
+    assert "capacity full (1 observations)" in retry["error"]
+    assert called is False
+
+
+def test_v3_readers_ignore_pending_table_until_store_migrates(tmp_path):
+    store_path = tmp_path / "observations.db"
+    conn = init_store(store_path)
+    conn.execute("DROP TABLE pending_publications")
+    conn.execute("PRAGMA user_version=3")
+    conn.commit()
+    conn.close()
+
+    with read_status(store_path, datetime.now(timezone.utc), query_services=False) as (reader, comp):
+        assert reader is not None
+        assert comp.store_fault is None
+        assert comp.pending_publication_count == 0
+
+    migrated = init_store(store_path)
+    try:
+        assert migrated.execute("PRAGMA user_version").fetchone()[0] == 6
+        assert migrated.execute("SELECT COUNT(*) FROM pending_publications").fetchone()[0] == 0
+    finally:
+        migrated.close()

tests/test_issue_99.py

@@ -0,0 +1,520 @@
+"""Historical local-day migration and repair acceptance tests for issue #99."""
+import sqlite3
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from fenris.local_day import (
+    query_local_day_summary,
+    record_local_activity_interval,
+)
+from fenris.status import read_status
+from fenris.store import init_store
+
+SOURCE_SCHEMA_VERSION = 5
+LOCAL_DATE = "2026-09-02"
+TIMEZONE = "Asia/Kolkata"
+UTC_START = "2026-09-01T18:30:00+00:00"
+UTC_END = "2026-09-02T18:30:00+00:00"
+
+
+def _legacy_store(path: Path) -> sqlite3.Connection:
+    """Build a v5 store with one old coarse local-day summary."""
+    conn = init_store(path)
+    conn.execute(
+        "INSERT INTO local_days "
+        "(local_date, tz_name, tz_offset, utc_start, utc_end, bytes_written, "
+        " bytes_read, coverage, sample_count, complete, activity_incomplete, "
+        " activity_precision) "
+        "VALUES (?, ?, '+05:30', ?, ?, 999999, 888888, 1.0, 24, 1, 0, 'legacy')",
+        (LOCAL_DATE, TIMEZONE, UTC_START, UTC_END),
+    )
+    conn.execute(
+        "INSERT INTO monitoring_periods (started_at, ended_at) VALUES (?, ?)",
+        (UTC_START, "2026-09-03T00:00:00+00:00"),
+    )
+    conn.execute(
+        "INSERT INTO controller_segments "
+        "(opened_at, identity_key, identity_degraded) VALUES (?, ?, 0)",
+        (UTC_START, "controller-a"),
+    )
+    conn.execute(
+        "INSERT INTO hour_observations "
+        "(hour, bytes_written_delta, bytes_read_delta, sample_count, coverage) "
+        "VALUES ('2026-09-02T00:00:00+00:00', 123, 45, 2, 1.0)"
+    )
+    conn.execute(
+        "INSERT INTO day_aggregates "
+        "(day, bytes_written_delta, bytes_read_delta, coverage) "
+        "VALUES ('2026-09-02', 678, 90, 1.0)"
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, power_on_hours, "
+        " segment_id, local_tz) VALUES (?, '/dev/nvme0', 1000, 2000, 100, 1, ?)",
+        ("2026-09-01T19:00:00+00:00", TIMEZONE),
+    )
+    conn.execute(
+        "INSERT INTO samples (ts, device, bytes_written, bytes_read, power_on_hours, "
+        " segment_id, local_tz) VALUES (?, '/dev/nvme0', 11000, 5000, 115, 1, ?)",
+        ("2026-09-02T10:00:00+00:00", TIMEZONE),
+    )
+    conn.execute(f"PRAGMA user_version={SOURCE_SCHEMA_VERSION}")
+    conn.commit()
+    return conn
+
+
+def test_upgrade_rebuilds_legacy_day_from_samples_and_keeps_recorded_bounds(
+    tmp_path, monkeypatch,
+):
+    db = tmp_path / "legacy.db"
+    conn = _legacy_store(db)
+    conn.close()
+    monkeypatch.setenv("TZ", "UTC")
+
+    conn = init_store(db)
+    conn.close()
+
+    with read_status(
+        db, datetime(2026, 9, 4, tzinfo=timezone.utc), query_services=False,
+    ) as (reader, composition):
+        assert reader is not None
+        assert composition.store_fault is None
+        summary = query_local_day_summary(
+            reader, LOCAL_DATE, TIMEZONE,
+            datetime(2026, 9, 4, tzinfo=timezone.utc),
+        )
+        assert summary is not None
+        assert summary["local_date"] == LOCAL_DATE
+        assert summary["tz_name"] == TIMEZONE
+        assert summary["tz_offset"] == "+05:30"
+        assert summary["utc_start"] == UTC_START
+        assert summary["utc_end"] == UTC_END
+        assert summary["bytes_written"] == 10_000
+        assert summary["bytes_read"] == 3_000
+        assert summary["activity_precision"] == "measured"
+        assert summary["segments"] == [{
+            "segment_id": 1,
+            "bytes_written": 10_000,
+            "bytes_read": 3_000,
+            "activity_seconds": 54_000,
+            "activity_intervals": 1,
+        }]
+        assert summary["activity_state"] == "incomplete"
+        assert reader.execute(
+            "SELECT bytes_written_delta FROM hour_observations"
+        ).fetchone()[0] == 123
+        assert reader.execute(
+            "SELECT bytes_written_delta FROM day_aggregates"
+        ).fetchone()[0] == 678
+
+    conn = init_store(db)
+    conn.close()
+    with read_status(
+        db, datetime(2026, 9, 4, tzinfo=timezone.utc), query_services=False,
+    ) as (reader, _composition):
+        assert reader is not None
+        summary = query_local_day_summary(
+            reader, LOCAL_DATE, TIMEZONE,
+            datetime(2026, 9, 4, tzinfo=timezone.utc),
+        )
+        assert summary["bytes_written"] == 10_000
+        assert summary["bytes_read"] == 3_000
+
+
+def test_repair_rebuilds_legacy_local_day_once_across_restarts(tmp_path):
+    from fenris.repair import repair_derivation
+
+    db = tmp_path / "repair.db"
+    conn = _legacy_store(db)
+
+    first = repair_derivation(conn)
+    assert first.ok
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] == 10_000
+    assert summary["bytes_read"] == 3_000
+
+    conn.close()
+    conn = sqlite3.connect(db)
+    second = repair_derivation(conn)
+    assert second.ok
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] == 10_000
+    assert summary["bytes_read"] == 3_000
+    assert conn.execute(
+        "SELECT bytes_written, bytes_read, activity_intervals "
+        "FROM local_day_segment_totals"
+    ).fetchall() == [(10_000, 3_000, 1)]
+    conn.close()
+
+
+def test_interrupted_repair_keeps_local_day_rebuild_retryable(tmp_path):
+    from fenris.repair import repair_derivation
+
+    db = tmp_path / "repair-interrupted.db"
+    conn = _legacy_store(db)
+    conn.execute(
+        "CREATE TRIGGER interrupt_repair_local_rebuild "
+        "BEFORE INSERT ON local_day_segment_totals "
+        "BEGIN SELECT RAISE(ABORT, 'simulated repair interruption'); END"
+    )
+    conn.commit()
+
+    failed = repair_derivation(conn)
+    assert not failed.ok
+    assert query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )["bytes_written"] is None
+    assert conn.execute(
+        "SELECT activity_precision, bytes_written FROM local_days"
+    ).fetchone() == ("legacy", 999_999)
+    conn.execute("DROP TRIGGER interrupt_repair_local_rebuild")
+    conn.commit()
+    conn.close()
+
+    conn = sqlite3.connect(db)
+    retried = repair_derivation(conn)
+    assert retried.ok
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] == 10_000
+    assert summary["bytes_read"] == 3_000
+    conn.close()
+
+
+def test_stale_repair_marker_does_not_block_crash_retry(tmp_path):
+    from fenris.repair import repair_derivation
+
+    conn = _legacy_store(tmp_path / "stale-repair.db")
+    conn.execute(
+        "INSERT OR REPLACE INTO store_metadata (key, value) "
+        "VALUES ('repair_in_progress', 'true')"
+    )
+    conn.commit()
+
+    result = repair_derivation(conn)
+
+    assert result.ok
+    assert conn.execute(
+        "SELECT value FROM store_metadata WHERE key = 'repair_in_progress'"
+    ).fetchone() == ("false",)
+    assert query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )["bytes_written"] == 10_000
+    conn.close()
+
+
+def test_upgrade_uses_whole_utc_hours_but_skips_midnight_straddling_hours(
+    tmp_path,
+):
+    db = tmp_path / "coarse.db"
+    conn = _legacy_store(db)
+    conn.execute("DELETE FROM samples")
+    conn.execute(
+        "INSERT INTO hour_observations "
+        "(hour, bytes_written_delta, bytes_read_delta, sample_count, "
+        " active_seconds, coverage) "
+        "VALUES ('2026-09-01T18:00:00+00:00', 99999, 9999, 2, 3600, 1.0)"
+    )
+    conn.execute(
+        "INSERT INTO hour_observations "
+        "(hour, bytes_written_delta, bytes_read_delta, sample_count, "
+        " active_seconds, coverage) "
+        "VALUES ('2026-09-02T18:00:00+00:00', 88888, 8888, 2, 3600, 1.0)"
+    )
+    conn.commit()
+    conn.close()
+
+    conn = init_store(db)
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] == 123
+    assert summary["bytes_read"] == 45
+    assert summary["activity_precision"] == "coarse"
+    assert summary["activity_state"] == "incomplete"
+    assert summary["utc_start"] == UTC_START
+    assert summary["utc_end"] == UTC_END
+
+    conn.execute(
+        "INSERT INTO samples "
+        "(id, ts, device, bytes_written, bytes_read, segment_id, local_tz) "
+        "VALUES (10, '2026-09-02T12:00:00+00:00', '/dev/nvme0', 1000, 2000, 1, ?), "
+        "       (11, '2026-09-02T12:05:00+00:00', '/dev/nvme0', 2000, 2500, 1, ?)",
+        (TIMEZONE, TIMEZONE),
+    )
+    conn.commit()
+    assert record_local_activity_interval(
+        conn,
+        {
+            "id": 10, "ts": "2026-09-02T12:00:00+00:00",
+            "bytes_written": 1_000, "bytes_read": 2_000,
+            "local_tz": TIMEZONE,
+        },
+        {
+            "id": 11, "ts": "2026-09-02T12:05:00+00:00",
+            "bytes_written": 2_000, "bytes_read": 2_500,
+            "local_tz": TIMEZONE,
+        },
+        start_sample_id=10, end_sample_id=11, segment_id=1,
+    ) == "known"
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] == 1_000
+    assert summary["bytes_read"] == 500
+    assert summary["activity_precision"] == "measured"
+    assert summary["segments"] == [{
+        "segment_id": 1,
+        "bytes_written": 1_000,
+        "bytes_read": 500,
+        "activity_seconds": 300,
+        "activity_intervals": 1,
+    }]
+    conn.close()
+
+
+def test_upgrade_does_not_double_count_coarse_hours_in_overlapping_days(tmp_path):
+    db = tmp_path / "overlapping-coarse-days.db"
+    conn = _legacy_store(db)
+    conn.execute("DELETE FROM samples")
+    conn.execute(
+        "INSERT INTO local_days "
+        "(local_date, tz_name, tz_offset, utc_start, utc_end, bytes_written, "
+        " bytes_read, coverage, sample_count, complete, activity_precision) "
+        "VALUES ('2026-09-02', 'UTC', '+00:00', "
+        "'2026-09-02T00:00:00+00:00', '2026-09-03T00:00:00+00:00', "
+        "999999, 888888, 1.0, 24, 1, 'legacy')"
+    )
+    conn.commit()
+    conn.close()
+
+    conn = init_store(db)
+
+    for tz_name in (TIMEZONE, "UTC"):
+        summary = query_local_day_summary(
+            conn, LOCAL_DATE, tz_name,
+            datetime(2026, 9, 4, tzinfo=timezone.utc),
+        )
+        assert summary["bytes_written"] is None
+        assert summary["bytes_read"] is None
+        assert summary["activity_precision"] == "unavailable"
+        assert summary["activity_state"] == "unavailable"
+    assert conn.execute(
+        "SELECT bytes_written_delta, bytes_read_delta FROM hour_observations "
+        "WHERE hour = '2026-09-02T00:00:00+00:00'"
+    ).fetchone() == (123, 45)
+    conn.close()
+
+
+def test_upgrade_retains_local_midnight_interval_once_as_shared_evidence(tmp_path):
+    db = tmp_path / "shared.db"
+    conn = _legacy_store(db)
+    conn.execute(
+        "INSERT INTO local_days "
+        "(local_date, tz_name, tz_offset, utc_start, utc_end, bytes_written, "
+        " bytes_read, coverage, sample_count, complete, activity_precision) "
+        "VALUES ('2026-09-01', ?, '+05:30', "
+        "'2026-08-31T18:30:00+00:00', '2026-09-01T18:30:00+00:00', "
+        "500, 600, 1.0, 24, 1, 'legacy')",
+        (TIMEZONE,),
+    )
+    conn.execute(
+        "UPDATE monitoring_periods SET started_at = '2026-08-31T18:00:00+00:00'"
+    )
+    conn.execute(
+        "UPDATE controller_segments SET opened_at = '2026-08-31T18:00:00+00:00'"
+    )
+    conn.execute(
+        "UPDATE samples SET ts = '2026-09-01T18:25:00+00:00', "
+        "bytes_written = 1000, bytes_read = 2000 WHERE id = 1"
+    )
+    conn.execute(
+        "UPDATE samples SET ts = '2026-09-01T18:35:00+00:00', "
+        "bytes_written = 1100, bytes_read = 2050 WHERE id = 2"
+    )
+    conn.commit()
+    conn.close()
+
+    conn = init_store(db)
+    shared = conn.execute(
+        "SELECT COUNT(*), SUM(bytes_written), SUM(bytes_read) "
+        "FROM local_day_unallocated_evidence"
+    ).fetchone()
+    assert shared == (1, 100, 50)
+    for local_date in ("2026-09-01", "2026-09-02"):
+        summary = query_local_day_summary(
+            conn, local_date, TIMEZONE,
+            datetime(2026, 9, 4, tzinfo=timezone.utc),
+        )
+        assert summary["bytes_written"] is None
+        assert summary["shared_bytes_written"] == 100
+        assert summary["shared_bytes_read"] == 50
+        assert summary["activity_state"] == "incomplete"
+    from fenris.repair import repair_derivation
+    assert repair_derivation(conn).ok
+    assert conn.execute(
+        "SELECT COUNT(*), SUM(bytes_written), SUM(bytes_read) "
+        "FROM local_day_unallocated_evidence"
+    ).fetchone() == (1, 100, 50)
+    conn.close()
+
+
+def test_interrupted_upgrade_keeps_legacy_day_retryable(tmp_path):
+    db = tmp_path / "interrupted.db"
+    conn = _legacy_store(db)
+    conn.execute(
+        "CREATE TRIGGER interrupt_local_rebuild "
+        "BEFORE INSERT ON local_day_segment_totals "
+        "BEGIN SELECT RAISE(ABORT, 'simulated interruption'); END"
+    )
+    conn.commit()
+    conn.close()
+
+    with pytest.raises(sqlite3.IntegrityError, match="simulated interruption"):
+        init_store(db)
+
+    conn = sqlite3.connect(db)
+    assert conn.execute("PRAGMA user_version").fetchone() == (SOURCE_SCHEMA_VERSION,)
+    assert conn.execute(
+        "SELECT activity_precision, bytes_written FROM local_days"
+    ).fetchone() == ("legacy", 999_999)
+    conn.execute("DROP TRIGGER interrupt_local_rebuild")
+    conn.commit()
+    conn.close()
+
+    conn = init_store(db)
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] == 10_000
+    assert summary["bytes_read"] == 3_000
+    conn.close()
+
+
+def test_prior_migration_step_commits_before_local_day_repair(tmp_path):
+    db = tmp_path / "migration-steps.db"
+    conn = _legacy_store(db)
+    conn.execute("PRAGMA user_version=4")
+    conn.execute(
+        "CREATE TRIGGER interrupt_local_rebuild BEFORE UPDATE OF bytes_written "
+        "ON local_days BEGIN SELECT RAISE(ABORT, 'simulated interruption'); END"
+    )
+    conn.commit()
+    conn.close()
+
+    with pytest.raises(sqlite3.IntegrityError, match="simulated interruption"):
+        init_store(db)
+
+    conn = sqlite3.connect(db)
+    assert conn.execute("PRAGMA user_version").fetchone() == (5,)
+    assert conn.execute(
+        "SELECT 1 FROM sqlite_master WHERE type='table' "
+        "AND name='local_day_segment_totals'"
+    ).fetchone() is not None
+    assert conn.execute(
+        "SELECT bytes_written FROM local_days WHERE local_date=?",
+        (LOCAL_DATE,),
+    ).fetchone() == (999_999,)
+    conn.close()
+
+
+def test_upgrade_does_not_allocate_unzoned_interval_to_overlapping_timezones(
+    tmp_path,
+):
+    db = tmp_path / "overlapping-timezones.db"
+    conn = _legacy_store(db)
+    conn.execute(
+        "INSERT INTO local_days "
+        "(local_date, tz_name, tz_offset, utc_start, utc_end, bytes_written, "
+        " bytes_read, coverage, sample_count, complete, activity_precision) "
+        "VALUES ('2026-09-02', 'UTC', '+00:00', "
+        "'2026-09-02T00:00:00+00:00', '2026-09-03T00:00:00+00:00', "
+        "999999, 888888, 1.0, 24, 1, 'legacy')"
+    )
+    conn.execute(
+        "UPDATE samples SET ts = '2026-09-02T01:00:00+00:00', local_tz = NULL "
+        "WHERE id = 1"
+    )
+    conn.execute(
+        "UPDATE samples SET ts = '2026-09-02T10:00:00+00:00', local_tz = NULL "
+        "WHERE id = 2"
+    )
+    conn.commit()
+    conn.close()
+
+    conn = init_store(db)
+    for tz_name in (TIMEZONE, "UTC"):
+        summary = query_local_day_summary(
+            conn, LOCAL_DATE, tz_name,
+            datetime(2026, 9, 4, tzinfo=timezone.utc),
+        )
+        assert summary["bytes_written"] is None
+        assert summary["bytes_read"] is None
+        assert summary["activity_precision"] == "unavailable"
+        assert summary["activity_state"] == "unavailable"
+    conn.close()
+
+
+def test_upgrade_rejects_interval_before_recorded_controller_segment(
+    tmp_path,
+):
+    db = tmp_path / "segment-boundary.db"
+    conn = _legacy_store(db)
+    conn.execute(
+        "UPDATE controller_segments "
+        "SET opened_at = '2026-09-01T20:00:00+00:00' WHERE id = 1"
+    )
+    conn.commit()
+    conn.close()
+
+    conn = init_store(db)
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] is None
+    assert summary["bytes_read"] is None
+    assert summary["activity_precision"] == "unavailable"
+    conn.close()
+
+
+def test_sample_at_local_midnight_does_not_hide_coarse_reconstruction(tmp_path):
+    db = tmp_path / "midnight-sample.db"
+    conn = _legacy_store(db)
+    conn.execute(
+        "UPDATE samples SET ts = '2026-09-01T17:00:00+00:00' WHERE id = 1"
+    )
+    conn.execute(
+        "UPDATE samples SET ts = ?, local_tz = ? WHERE id = 2",
+        (UTC_END, TIMEZONE),
+    )
+    conn.commit()
+    conn.close()
+
+    conn = init_store(db)
+    summary = query_local_day_summary(
+        conn, LOCAL_DATE, TIMEZONE,
+        datetime(2026, 9, 4, tzinfo=timezone.utc),
+    )
+    assert summary["bytes_written"] == 123
+    assert summary["bytes_read"] == 45
+    assert summary["activity_precision"] == "coarse"
+    assert summary["activity_state"] == "incomplete"
+    conn.close()

tests/test_live_graph.py

@@ -6,26 +6,24 @@ Covers:
 - Three-minute cadence constants
 - TUI integration with live graph
 """
+import sys
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
 
 import pytest
-import sys
 
 sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
 
-from fenris.store import init_store
+from fenris.activity_selection import IntervalIdentity
 from fenris.monitoring_periods import ensure_period_open
+from fenris.status import ACCURACY_SEC, CADENCE_DEFAULT_S, FRESH_THRESHOLD_S
+from fenris.store import init_store
 from fenris.tui import (
+    LIVE_WINDOW_H,
     FenrisTuiApp,
     LiveActivityGraph,
     _query_live_graph_data,
-    LIVE_WINDOW_H,
-    _query_daily_graph_data,
-    _RANGE_OPTIONS,
 )
-from fenris.status import CADENCE_DEFAULT_S, FRESH_THRESHOLD_S, ACCURACY_SEC
-
 
 # ---------------------------------------------------------------------------
 # Helpers
@@ -63,6 +61,27 @@ def _open_period(conn, start="2026-09-01T00:00:00+00:00"):
     ensure_period_open(conn, datetime.fromisoformat(start))
 
 
+def _live_intervals(count=4):
+    first = datetime(2026, 9, 28, 12, 0, tzinfo=timezone.utc)
+    intervals = []
+    for index in range(count):
+        start = first + timedelta(minutes=3 * index)
+        end = start + timedelta(minutes=3)
+        intervals.append({
+            "start_ts": start.isoformat(),
+            "end_ts": end.isoformat(),
+            "start_label": start.strftime("%H:%M"),
+            "end_label": end.strftime("%H:%M"),
+            "bytes_written": (index + 1) * 1000,
+            "bytes_read": (index + 1) * 500,
+            "elapsed_s": 180,
+            "is_gap": False,
+            "is_zero": False,
+            "is_segment_boundary": False,
+        })
+    return intervals
+
+
 # ---------------------------------------------------------------------------
 # Cadence constants (issue #91 AC1)
 # ---------------------------------------------------------------------------
@@ -322,3 +341,85 @@ class TestTUILiveIntegration:
             await pilot.pause()
             main_grid = app.query_one("#main-grid")
             assert main_grid.has_class("constrained")
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("size", [(100, 40), (70, 20)])
+    async def test_follow_pin_expiry_and_today_use_visible_readout(self, tmp_path, size):
+        """Live selection follows, pins, expires, and resets at both layout sizes."""
+        app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
+        intervals = _live_intervals()
+        async with app.run_test(size=size) as pilot:
+            await pilot.pause()
+            graph = app.query_one("#live-activity")
+            graph.set_data(intervals[:3])
+
+            readout = str(app.query_one("#live-readout").render())
+            assert "12:09 UTC" in readout
+            assert graph.selection.following_live
+            assert graph.selection.selected_live_interval == IntervalIdentity(
+                intervals[2]["start_ts"], intervals[2]["end_ts"]
+            )
+
+            # Refresh while following selects the newly measured newest point.
+            graph.set_data(intervals)
+            assert "12:12 UTC" in str(app.query_one("#live-readout").render())
+
+            # Live graph receives focus at launch; first arrow pins immediately.
+            await pilot.press("left")
+            pinned = graph.selection.selected_live_interval
+            readout = str(app.query_one("#live-readout").render())
+            assert "12:09 UTC" in readout
+            assert not graph.selection.following_live
+
+            graph.set_data(intervals)
+            readout = str(app.query_one("#live-readout").render())
+            assert "12:09 UTC" in readout
+            assert graph.selection.selected_live_interval == pinned
+
+            # Read/write changes presentation only; pinned point stays stable.
+            await pilot.press("w")
+            graph.set_data(intervals)
+            assert graph.measure == "read"
+            assert graph.selection.selected_live_interval == pinned
+            assert "12:09 UTC" in str(app.query_one("#live-readout").render())
+
+            # Remove pinned interval from rolling window. Explain expiration.
+            graph.set_data(intervals[3:])
+            readout = str(app.query_one("#live-readout").render())
+            assert "Inspected interval expired; following live" in readout
+            assert "12:12 UTC" in readout
+            assert graph.selection.following_live
+
+            # Today action clears expiration and resumes newest-point following.
+            await pilot.press("t")
+            assert graph.selection.following_live
+            assert not graph.selection.live_interval_expired
+            graph.set_data(intervals)
+            readout = str(app.query_one("#live-readout").render())
+            assert "12:12 UTC" in readout
+            assert "Inspected interval expired" not in readout
+            assert graph.selection.following_live
+
+    @pytest.mark.asyncio
+    async def test_mouse_inspection_pins_same_interval_identity(self, tmp_path):
+        """Mouse inspection pins the interval at its visible graph location."""
+        app = FenrisTuiApp(store_path=tmp_path / "test.db", refresh_interval_s=999)
+        intervals = _live_intervals()
+        async with app.run_test(size=(100, 40)) as pilot:
+            await pilot.pause()
+            graph = app.query_one("#live-activity")
+            graph.set_data(intervals[:3])
+            await pilot.pause()
+            render = app.query_one("#live-render")
+            await pilot.click(
+                "#live-render",
+                offset=(8, max(0, render.content_size.height // 2)),
+            )
+            await pilot.pause()
+
+            expected = IntervalIdentity(
+                intervals[0]["start_ts"], intervals[0]["end_ts"]
+            )
+            assert graph.selection.selected_live_interval == expected
+            assert not graph.selection.following_live
+            assert "12:03 UTC" in str(app.query_one("#live-readout").render())

tests/test_local_day.py

@@ -25,6 +25,7 @@ from fenris.store import init_store, SCHEMA_VERSION
 from fenris.local_day import (
     derive_local_day_summary,
     persist_local_day,
+    record_local_activity_interval,
     query_local_day_summary,
     query_current_local_day,
     LocalDaySummary,
@@ -135,13 +136,56 @@ class TestSchemaMigration:
         assert count == 0
         conn.close()
 
+    def test_v4_totals_keep_values_but_are_not_treated_as_local_precision(
+        self, tmp_path,
+    ):
+        db = tmp_path / "v4.db"
+        conn = sqlite3.connect(db)
+        conn.execute(
+            "CREATE TABLE samples (id INTEGER PRIMARY KEY, ts TEXT, device TEXT)"
+        )
+        conn.execute(
+            "CREATE TABLE pending_publications "
+            "(id INTEGER PRIMARY KEY, sample_ts TEXT, payload TEXT)"
+        )
+        conn.execute(
+            "CREATE TABLE local_days ("
+            "id INTEGER PRIMARY KEY, local_date TEXT NOT NULL, tz_name TEXT NOT NULL, "
+            "tz_offset TEXT NOT NULL, utc_start TEXT NOT NULL, utc_end TEXT NOT NULL, "
+            "bytes_written INTEGER DEFAULT 0, bytes_read INTEGER DEFAULT 0, "
+            "coverage REAL DEFAULT 0, sample_count INTEGER DEFAULT 0, "
+            "complete BOOLEAN DEFAULT 0, UNIQUE(local_date, tz_name))"
+        )
+        conn.execute(
+            "INSERT INTO local_days VALUES "
+            "(1, '2026-09-01', 'UTC', '+00:00', "
+            "'2026-09-01T00:00:00+00:00', '2026-09-02T00:00:00+00:00', "
+            "5120000, 2048000, 0.5, 10, 0)"
+        )
+        conn.execute("PRAGMA user_version=4")
+        conn.commit()
+        conn.close()
+
+        conn = init_store(db)
+        stored = conn.execute(
+            "SELECT bytes_written, activity_precision FROM local_days"
+        ).fetchone()
+        assert stored == (5_120_000, "unavailable")
+        visible = query_local_day_summary(conn, "2026-09-01", "UTC")
+        assert visible["bytes_written"] is None
+        assert visible["activity_state"] == "unavailable"
+        assert "local_tz" in {
+            row[1] for row in conn.execute("PRAGMA table_info(samples)")
+        }
+        conn.close()
+
 
 # ---------------------------------------------------------------------------
-# Local-day derivation from UTC hours
+# Local-day coverage derivation from UTC hours
 # ---------------------------------------------------------------------------
 
 class TestDeriveLocalDay:
-    """Derive local-day summaries from UTC hour observations."""
+    """Derive coverage from UTC hours without allocating local-day bytes."""
 
     def test_utc_plus_zero(self, tmp_path):
         """UTC timezone: local day boundaries = UTC day boundaries."""
@@ -158,8 +202,9 @@ class TestDeriveLocalDay:
         assert summary.local_date == "2026-09-01"
         assert summary.tz_name == "UTC"
         assert summary.tz_offset == "+00:00"
-        assert summary.bytes_written == 300
-        assert summary.bytes_read == 130
+        assert summary.bytes_written == 0
+        assert summary.bytes_read == 0
+        assert summary.activity_precision == "unavailable"
         assert summary.complete is False  # not all 24 hours covered
         conn.close()
 
@@ -179,12 +224,13 @@ class TestDeriveLocalDay:
         assert summary.local_date == "2026-09-01"
         assert summary.tz_name == "Asia/Kolkata"
         assert summary.tz_offset == "+05:30"
-        assert summary.bytes_written == 300
-        assert summary.bytes_read == 130
+        assert summary.bytes_written == 0
+        assert summary.bytes_read == 0
+        assert summary.activity_precision == "unavailable"
         conn.close()
 
-    def test_midnight_spanning_hour_included(self, tmp_path):
-        """UTC hour straddling local midnight is included in the local day."""
+    def test_hourly_bytes_cannot_be_allocated_to_local_midnight(self, tmp_path):
+        """UTC-hour totals do not establish exact local-day byte totals."""
         conn = init_store(tmp_path / "obs.db")
 
         # For UTC+5:30, local day 2026-09-01 starts at 2026-08-31T18:30 UTC
@@ -196,9 +242,9 @@ class TestDeriveLocalDay:
         summary = derive_local_day_summary(conn, "Asia/Kolkata", clock)
 
         assert summary is not None
-        # The midnight-spanning hour (18:00) is included
-        assert summary.bytes_written == 300
-        assert summary.bytes_read == 130
+        assert summary.bytes_written == 0
+        assert summary.bytes_read == 0
+        assert summary.activity_precision == "unavailable"
         conn.close()
 
     def test_no_hours_returns_none(self, tmp_path):
@@ -387,6 +433,411 @@ class TestCollectorIntegration:
         assert row[1] == 30 * 512000  # reads
         conn.close()
 
+    def test_interval_crossing_utc_hour_stays_wholly_in_local_day(
+        self, tmp_path, sysfs_tree, monkeypatch,
+    ):
+        """A measured interval inside one local date keeps its full delta."""
+        from fenris.collector import run_collection
+
+        monkeypatch.setenv("TZ", "Asia/Kolkata")
+        store = str(tmp_path / "obs.db")
+        cfg = {"device": "/dev/nvme0", "store_path": store}
+        sysfs_nvme = sysfs_tree / "sys" / "class" / "nvme" / "nvme0"
+        first_at = datetime(2026, 9, 1, 12, 59, tzinfo=timezone.utc)
+        second_at = datetime(2026, 9, 1, 13, 1, tzinfo=timezone.utc)
+
+        first = run_collection(
+            _make_smartctl(10_000_000, 8_000_000),
+            sysfs_nvme,
+            cfg,
+            _Clock(first_at),
+        )
+        second = run_collection(
+            _make_smartctl(10_000_010, 8_000_004),
+            sysfs_nvme,
+            cfg,
+            _Clock(second_at),
+        )
+
+        assert first["ok"] and second["ok"]
+        conn = sqlite3.connect(store)
+        summary = query_local_day_summary(conn, "2026-09-01")
+
+        assert summary is not None
+        assert summary["bytes_written"] == 5_120_000
+        assert summary["bytes_read"] == 2_048_000
+        assert summary["segments"] == [{
+            "segment_id": 1,
+            "bytes_written": 5_120_000,
+            "bytes_read": 2_048_000,
+            "activity_seconds": 120,
+            "activity_intervals": 1,
+        }]
+        segment_totals = conn.execute(
+            "SELECT segment_id, bytes_written, bytes_read, activity_intervals "
+            "FROM local_day_segment_totals"
+        ).fetchone()
+        assert segment_totals == (1, 5_120_000, 2_048_000, 1)
+
+        samples = conn.execute(
+            "SELECT id, ts, bytes_written, bytes_read, local_tz "
+            "FROM samples ORDER BY id"
+        ).fetchall()
+        first_sample, second_sample = samples
+        assert record_local_activity_interval(
+            conn,
+            dict(zip(("id", "ts", "bytes_written", "bytes_read", "local_tz"), first_sample)),
+            dict(zip(("id", "ts", "bytes_written", "bytes_read", "local_tz"), second_sample)),
+            start_sample_id=first_sample[0],
+            end_sample_id=second_sample[0],
+            segment_id=1,
+        ) == "known"
+        repeated = query_local_day_summary(conn, "2026-09-01", "Asia/Kolkata")
+        assert repeated["bytes_written"] == 5_120_000
+        assert repeated["bytes_read"] == 2_048_000
+        assert conn.execute(
+            "SELECT SUM(bytes_written), SUM(bytes_read), SUM(activity_intervals) "
+            "FROM local_day_segment_totals"
+        ).fetchone() == (5_120_000, 2_048_000, 1)
+        conn.close()
+
+    @pytest.mark.parametrize(
+        (
+            "first_at", "second_at", "local_date", "expected_hours",
+            "expected_start", "expected_end", "expected_offset",
+        ),
+        [
+            (
+                datetime(2026, 3, 8, 6, 30, tzinfo=timezone.utc),
+                datetime(2026, 3, 8, 7, 30, tzinfo=timezone.utc),
+                "2026-03-08", 23,
+                "2026-03-08T05:00:00+00:00",
+                "2026-03-09T04:00:00+00:00", "-05:00",
+            ),
+            (
+                datetime(2026, 11, 1, 5, 30, tzinfo=timezone.utc),
+                datetime(2026, 11, 1, 6, 30, tzinfo=timezone.utc),
+                "2026-11-01", 25,
+                "2026-11-01T04:00:00+00:00",
+                "2026-11-02T05:00:00+00:00", "-04:00",
+            ),
+        ],
+        ids=("spring-forward", "fall-back-repeated-hour"),
+    )
+    def test_collector_retains_exact_interval_on_dst_local_day(
+        self,
+        tmp_path,
+        sysfs_tree,
+        monkeypatch,
+        first_at,
+        second_at,
+        local_date,
+        expected_hours,
+        expected_start,
+        expected_end,
+        expected_offset,
+    ):
+        """Collection preserves exact activity on short and long local days."""
+        from fenris.collector import run_collection
+
+        zone = "America/New_York"
+        monkeypatch.setenv("TZ", zone)
+        store = str(tmp_path / "obs.db")
+        cfg = {"device": "/dev/nvme0", "store_path": store}
+        sysfs_nvme = sysfs_tree / "sys" / "class" / "nvme" / "nvme0"
+        assert run_collection(
+            _make_smartctl(10_000_000, 8_000_000),
+            sysfs_nvme,
+            cfg,
+            _Clock(first_at),
+        )["ok"]
+        assert run_collection(
+            _make_smartctl(10_000_001, 8_000_002),
+            sysfs_nvme,
+            cfg,
+            _Clock(second_at),
+        )["ok"]
+
+        conn = sqlite3.connect(store)
+        summary = query_local_day_summary(
+            conn, local_date, zone, second_at + timedelta(minutes=1)
+        )
+        assert summary is not None
+        assert summary["bytes_written"] == 512_000
+        assert summary["bytes_read"] == 1_024_000
+        assert summary["activity_seconds"] == 3_600
+        assert summary["utc_start"] == expected_start
+        assert summary["utc_end"] == expected_end
+        assert summary["tz_offset"] == expected_offset
+        assert (
+            datetime.fromisoformat(summary["utc_end"])
+            - datetime.fromisoformat(summary["utc_start"])
+        ).total_seconds() == expected_hours * 3_600
+        conn.close()
+
+    @pytest.mark.asyncio
+    async def test_local_midnight_interval_is_shared_once(
+        self, tmp_path, sysfs_tree, monkeypatch,
+    ):
+        """A real counter interval crossing local midnight stays unallocated."""
+        from fenris.collector import run_collection
+
+        monkeypatch.setenv("TZ", "Asia/Kolkata")
+        store = str(tmp_path / "obs.db")
+        cfg = {"device": "/dev/nvme0", "store_path": store}
+        sysfs_nvme = sysfs_tree / "sys" / "class" / "nvme" / "nvme0"
+        first_at = datetime(2026, 9, 1, 18, 25, tzinfo=timezone.utc)
+        second_at = datetime(2026, 9, 1, 18, 35, tzinfo=timezone.utc)
+
+        first = run_collection(
+            _make_smartctl(10_000_000, 8_000_000),
+            sysfs_nvme,
+            cfg,
+            _Clock(first_at),
+        )
+        second = run_collection(
+            _make_smartctl(10_000_100, 8_000_060),
+            sysfs_nvme,
+            cfg,
+            _Clock(second_at),
+        )
+        assert first["ok"] and second["ok"]
+
+        conn = sqlite3.connect(store)
+        evidence = conn.execute(
+            "SELECT COUNT(*), SUM(bytes_written), SUM(bytes_read) "
+            "FROM local_day_unallocated_evidence"
+        ).fetchone()
+        assert evidence == (1, 51_200_000, 30_720_000)
+        assert conn.execute(
+            "SELECT SUM(bytes_written), SUM(bytes_read) FROM local_days"
+        ).fetchone() == (0, 0)
+
+        for local_date in ("2026-09-01", "2026-09-02"):
+            summary = query_local_day_summary(
+                conn, local_date, "Asia/Kolkata", second_at
+            )
+            assert summary is not None
+            assert summary["bytes_written"] is None
+            assert summary["shared_bytes_written"] == 51_200_000
+            assert summary["shared_bytes_read"] == 30_720_000
+            assert summary["activity_state"] in ("incomplete", "so_far")
+        conn.close()
+
+        from fenris.status import read_status
+        from fenris.tui import FenrisTuiApp
+
+        app = FenrisTuiApp(store_path=Path(store), refresh_interval_s=999)
+        async with app.run_test(size=(100, 32)):
+            app._browse_date = "2026-09-02"
+            with read_status(
+                Path(store), second_at, query_services=False
+            ) as (reader, _):
+                assert reader is not None
+                app._render_local_day(reader)
+            visible = str(app.query_one("#local-day").render())
+            assert "2026-09-02" in visible
+            assert "Asia/Kolkata +05:30" in visible
+            assert "W unavailable" in visible
+            assert "shared at midnight W 0.051 GB" in visible
+
+    def test_synthetic_shared_bytes_conserve_one_hundred_bytes(self, tmp_path):
+        """Shared evidence conserves byte values before display rounding."""
+        conn = init_store(tmp_path / "obs.db")
+        first_at = datetime(2026, 9, 1, 18, 25, tzinfo=timezone.utc)
+        second_at = datetime(2026, 9, 3, 18, 35, tzinfo=timezone.utc)
+        ensure_period_open(conn, first_at)
+        conn.execute(
+            "INSERT INTO samples (ts, device, bytes_written, bytes_read, local_tz) "
+            "VALUES (?, ?, ?, ?, ?)",
+            (first_at.isoformat(), "/dev/nvme0", 1_000, 2_000, "Asia/Kolkata"),
+        )
+        first_id = conn.execute("SELECT last_insert_rowid()").fetchone()[0]
+        conn.execute(
+            "INSERT INTO samples (ts, device, bytes_written, bytes_read, local_tz) "
+            "VALUES (?, ?, ?, ?, ?)",
+            (second_at.isoformat(), "/dev/nvme0", 1_100, 2_050, "Asia/Kolkata"),
+        )
+        second_id = conn.execute("SELECT last_insert_rowid()").fetchone()[0]
+
+        result = record_local_activity_interval(
+            conn,
+            {
+                "id": first_id,
+                "ts": first_at.isoformat(),
+                "bytes_written": 1_000,
+                "bytes_read": 2_000,
+                "local_tz": "Asia/Kolkata",
+            },
+            {
+                "id": second_id,
+                "ts": second_at.isoformat(),
+                "bytes_written": 1_100,
+                "bytes_read": 2_050,
+                "local_tz": "Asia/Kolkata",
+            },
+            start_sample_id=first_id,
+            end_sample_id=second_id,
+            segment_id=1,
+        )
+
+        assert result == "local_midnight"
+        stored = conn.execute(
+            "SELECT COUNT(*), SUM(bytes_written), SUM(bytes_read) "
+            "FROM local_day_unallocated_evidence"
+        ).fetchone()
+        assert stored == (1, 100, 50)
+        record_local_activity_interval(
+            conn,
+            {
+                "id": first_id,
+                "ts": first_at.isoformat(),
+                "bytes_written": 1_000,
+                "bytes_read": 2_000,
+                "local_tz": "Asia/Kolkata",
+            },
+            {
+                "id": second_id,
+                "ts": second_at.isoformat(),
+                "bytes_written": 1_100,
+                "bytes_read": 2_050,
+                "local_tz": "Asia/Kolkata",
+            },
+            start_sample_id=first_id,
+            end_sample_id=second_id,
+            segment_id=1,
+        )
+        assert conn.execute(
+            "SELECT COUNT(*), SUM(bytes_written), SUM(bytes_read) "
+            "FROM local_day_unallocated_evidence"
+        ).fetchone() == (1, 100, 50)
+        assert conn.execute(
+            "SELECT SUM(bytes_written), SUM(bytes_read) FROM local_days"
+        ).fetchone() == (0, 0)
+        middle_day = query_local_day_summary(
+            conn, "2026-09-03", "Asia/Kolkata", second_at + timedelta(days=1)
+        )
+        assert middle_day["shared_bytes_written"] == 100
+        assert middle_day["shared_bytes_read"] == 50
+        assert middle_day["shared_evidence_count"] == 1
+        conn.close()
+
+    def test_timezone_change_keeps_interval_unallocated(
+        self, tmp_path, sysfs_tree, monkeypatch,
+    ):
+        from fenris.collector import run_collection
+
+        store = str(tmp_path / "obs.db")
+        cfg = {"device": "/dev/nvme0", "store_path": store}
+        sysfs_nvme = sysfs_tree / "sys" / "class" / "nvme" / "nvme0"
+        first_at = datetime(2026, 9, 1, 22, 0, tzinfo=timezone.utc)
+        second_at = datetime(2026, 9, 1, 22, 5, tzinfo=timezone.utc)
+
+        monkeypatch.setenv("TZ", "UTC")
+        assert run_collection(
+            _make_smartctl(10_000_000, 8_000_000),
+            sysfs_nvme,
+            cfg,
+            _Clock(first_at),
+        )["ok"]
+        monkeypatch.setenv("TZ", "Asia/Kolkata")
+        assert run_collection(
+            _make_smartctl(10_000_010, 8_000_004),
+            sysfs_nvme,
+            cfg,
+            _Clock(second_at),
+        )["ok"]
+
+        conn = sqlite3.connect(store)
+        event = conn.execute(
+            "SELECT COUNT(*), SUM(bytes_written), MIN(reason) "
+            "FROM local_day_unallocated_evidence"
+        ).fetchone()
+        assert event == (1, 5_120_000, "timezone_change")
+        old_zone = query_local_day_summary(conn, "2026-09-01", "UTC")
+        new_zone = query_local_day_summary(conn, "2026-09-02", "Asia/Kolkata")
+        assert old_zone["bytes_written"] is None
+        assert new_zone["bytes_written"] is None
+        assert old_zone["unallocated_bytes_written"] == 5_120_000
+        assert new_zone["unallocated_bytes_written"] == 5_120_000
+        conn.close()
+
+    def test_pause_boundary_keeps_interval_unallocated(
+        self, tmp_path, sysfs_tree, monkeypatch,
+    ):
+        from fenris.collector import run_collection
+
+        monkeypatch.setenv("TZ", "UTC")
+        store = str(tmp_path / "obs.db")
+        cfg = {"device": "/dev/nvme0", "store_path": store}
+        sysfs_nvme = sysfs_tree / "sys" / "class" / "nvme" / "nvme0"
+        first_at = datetime(2026, 9, 1, 12, 0, tzinfo=timezone.utc)
+        second_at = datetime(2026, 9, 1, 12, 10, tzinfo=timezone.utc)
+        assert run_collection(
+            _make_smartctl(10_000_000, 8_000_000),
+            sysfs_nvme,
+            cfg,
+            _Clock(first_at),
+        )["ok"]
+
+        conn = sqlite3.connect(store)
+        close_period(conn, first_at + timedelta(minutes=5), "user_disabled")
+        conn.close()
+
+        assert run_collection(
+            _make_smartctl(10_000_010, 8_000_004),
+            sysfs_nvme,
+            cfg,
+            _Clock(second_at),
+        )["ok"]
+
+        conn = sqlite3.connect(store)
+        reason = conn.execute(
+            "SELECT reason FROM local_day_unallocated_evidence"
+        ).fetchone()
+        summary = query_local_day_summary(conn, "2026-09-01", "UTC")
+        assert reason == ("monitoring_period",)
+        assert summary["bytes_written"] is None
+        assert summary["unallocated_bytes_written"] == 5_120_000
+        assert summary["shared_bytes_written"] == 0
+        conn.close()
+
+    def test_controller_segment_boundary_marks_local_activity_incomplete(
+        self, tmp_path, sysfs_tree, monkeypatch,
+    ):
+        from fenris.collector import run_collection
+
+        monkeypatch.setenv("TZ", "UTC")
+        store = str(tmp_path / "obs.db")
+        cfg = {"device": "/dev/nvme0", "store_path": store}
+        sysfs_nvme = sysfs_tree / "sys" / "class" / "nvme" / "nvme0"
+        first_at = datetime(2026, 9, 1, 12, 0, tzinfo=timezone.utc)
+        second_at = datetime(2026, 9, 1, 12, 5, tzinfo=timezone.utc)
+        assert run_collection(
+            _make_smartctl(10_000_000, 8_000_000),
+            sysfs_nvme,
+            cfg,
+            _Clock(first_at),
+        )["ok"]
+        assert run_collection(
+            _make_smartctl(10, 8_000_004),
+            sysfs_nvme,
+            cfg,
+            _Clock(second_at),
+        )["ok"]
+
+        conn = sqlite3.connect(store)
+        summary = query_local_day_summary(
+            conn, "2026-09-01", "UTC", second_at
+        )
+        assert summary is not None
+        assert summary["bytes_written"] is None
+        assert summary["activity_state"] == "incomplete"
+        assert conn.execute(
+            "SELECT COUNT(*) FROM local_day_unallocated_evidence"
+        ).fetchone()[0] == 0
+        conn.close()
+
 
 # ---------------------------------------------------------------------------
 # DST handling (23/25-hour days)
@@ -402,17 +853,72 @@ class TestDSTHandling:
         # Simulate a 23-hour day in a timezone with DST
         # For simplicity, just verify the summary records the correct UTC range
         clock = datetime(2026, 3, 8, 12, 0, 0, tzinfo=timezone.utc)
-        # US/Eastern springs forward on 2026-03-08
+        # New York springs forward on 2026-03-08
         # Local day 2026-03-08 is 23 hours: UTC [07:00, 06:00+1d)
         _insert_hour(conn, "2026-03-08T08:00:00+00:00", bw=100, active=3600)
         _insert_hour(conn, "2026-03-08T12:00:00+00:00", bw=200, active=3600)
 
-        summary = derive_local_day_summary(conn, "US/Eastern", clock)
+        summary = derive_local_day_summary(conn, "America/New_York", clock)
         assert summary is not None
         assert summary.local_date == "2026-03-08"
         # UTC range should be approximately 23 hours
         utc_start = datetime.fromisoformat(summary.utc_start)
         utc_end = datetime.fromisoformat(summary.utc_end)
         duration = (utc_end - utc_start).total_seconds()
-        assert 22 * 3600 <= duration <= 24 * 3600  # ~23h ± tolerance
+        assert duration == 23 * 3600
+        conn.close()
+
+    def test_long_day_25_hours(self):
+        from fenris.local_day import local_day_boundaries
+
+        start, end, offset = local_day_boundaries(
+            "2026-11-01", "America/New_York"
+        )
+        assert (end - start).total_seconds() == 25 * 3600
+        assert offset == "-04:00"
+
+    def test_half_hour_offset_records_exact_local_boundaries(self):
+        from fenris.local_day import local_day_boundaries
+
+        start, end, offset = local_day_boundaries(
+            "2026-09-01", "Asia/Kolkata"
+        )
+        assert start.isoformat() == "2026-08-31T18:30:00+00:00"
+        assert end.isoformat() == "2026-09-01T18:30:00+00:00"
+        assert offset == "+05:30"
+
+    def test_repeated_local_clock_labels_remain_one_recorded_day(
+        self, tmp_path, sysfs_tree, monkeypatch,
+    ):
+        from fenris.collector import run_collection
+
+        monkeypatch.setenv("TZ", "America/New_York")
+        store = str(tmp_path / "obs.db")
+        cfg = {"device": "/dev/nvme0", "store_path": store}
+        sysfs_nvme = sysfs_tree / "sys" / "class" / "nvme" / "nvme0"
+        # Both endpoints display as 01:30 locally, on opposite UTC offsets.
+        first_at = datetime(2026, 11, 1, 5, 30, tzinfo=timezone.utc)
+        second_at = datetime(2026, 11, 1, 6, 30, tzinfo=timezone.utc)
+        assert run_collection(
+            _make_smartctl(10_000_000, 8_000_000),
+            sysfs_nvme,
+            cfg,
+            _Clock(first_at),
+        )["ok"]
+        assert run_collection(
+            _make_smartctl(10_000_001, 8_000_002),
+            sysfs_nvme,
+            cfg,
+            _Clock(second_at),
+        )["ok"]
+
+        conn = sqlite3.connect(store)
+        summary = query_local_day_summary(
+            conn, "2026-11-01", "America/New_York", second_at
+        )
+        assert summary is not None
+        assert summary["bytes_written"] == 512_000
+        assert summary["bytes_read"] == 1_024_000
+        assert summary["utc_start"] == "2026-11-01T04:00:00+00:00"
+        assert summary["utc_end"] == "2026-11-02T05:00:00+00:00"
         conn.close()

tests/test_projection.py

@@ -100,8 +100,8 @@ def _insert_local_day(conn, local_date, tz_name="UTC", tz_offset="+00:00",
     conn.execute(
         "INSERT INTO local_days "
         "(local_date, tz_name, tz_offset, utc_start, utc_end, "
-        " bytes_written, bytes_read, coverage, sample_count, complete) "
-        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+        " bytes_written, bytes_read, coverage, sample_count, complete, activity_intervals) "
+        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 1)",
         (local_date, tz_name, tz_offset, utc_start, utc_end,
          bw, br, coverage, samples, complete),
     )

tests/test_pruning.py

@@ -27,9 +27,14 @@ def store_conn(tmp_path: Path):
 
 
 def _insert_sample(conn, ts_iso, device="/dev/nvme0"):
+    conn.execute(
+        "INSERT OR IGNORE INTO controller_segments "
+        "(id, opened_at, identity_key) VALUES (1, '2026-01-01T00:00:00+00:00', 'test')"
+    )
     conn.execute(
         "INSERT INTO samples (ts, device, data_units_written, data_units_read, "
-        " bytes_written, bytes_read, percentage_used) VALUES (?, ?, 0, 0, 0, 0, 0)",
+        " bytes_written, bytes_read, percentage_used, segment_id, local_tz) "
+        "VALUES (?, ?, 0, 0, 0, 0, 0, 1, 'UTC')",
         (ts_iso, device),
     )
     conn.commit()
@@ -50,33 +55,30 @@ class TestPruneOldSamples:
         cursor = store_conn.execute("SELECT COUNT(*) FROM samples")
         assert cursor.fetchone()[0] == 1
 
-    def test_removes_old_samples(self, store_conn):
+    def test_keeps_old_samples_without_replacement_evidence(self, store_conn):
         now = datetime(2026, 9, 15, 12, 0, 0, tzinfo=timezone.utc)
-        # Insert samples at 10, 14, and 15 days ago
-        # All three are before the cutoff (2026-09-01T12:00:00)
-        # The 15-day-old sample is not a boundary anchor because
-        # the next sample (14 days ago) is also before the cutoff
+        # Old samples stay until UTC and local-day evidence replace them.
         for days_ago in [10, 14, 15]:
             ts = (now - timedelta(days=days_ago)).isoformat()
             _insert_sample(store_conn, ts)
 
         pruned = prune_old_samples(store_conn, now, retention_days=14)
-        assert pruned == 1  # Only the 15-day-old sample removed
+        assert pruned == 0
 
         cursor = store_conn.execute("SELECT COUNT(*) FROM samples")
-        assert cursor.fetchone()[0] == 2
+        assert cursor.fetchone()[0] == 3
 
-    def test_removes_many_old_samples(self, store_conn):
+    def test_keeps_many_old_samples_without_replacement_evidence(self, store_conn):
         now = datetime(2026, 9, 15, 12, 0, 0, tzinfo=timezone.utc)
         for days_ago in range(1, 30):
             ts = (now - timedelta(days=days_ago)).isoformat()
             _insert_sample(store_conn, ts)
 
         pruned = prune_old_samples(store_conn, now, retention_days=14)
-        assert pruned == 15  # Days 15-29 removed
+        assert pruned == 0
 
         cursor = store_conn.execute("SELECT COUNT(*) FROM samples")
-        assert cursor.fetchone()[0] == 14  # Days 1-14 kept
+        assert cursor.fetchone()[0] == 29
 
     def test_empty_store_no_error(self, store_conn):
         now = datetime(2026, 9, 15, 12, 0, 0, tzinfo=timezone.utc)
@@ -130,8 +132,8 @@ class TestPruneOldSamples:
         )
         assert cursor.fetchone()[0] == 1
 
-    def test_old_sample_with_derived_interval_removed(self, store_conn):
-        """Old samples with fully derived intervals are removed."""
+    def test_hour_rows_alone_do_not_replace_local_day_evidence(self, store_conn):
+        """An hour row alone cannot authorize source-sample deletion."""
         now = datetime(2026, 9, 30, 12, 0, 0, tzinfo=timezone.utc)
         
         # Old sample with derived interval
@@ -148,8 +150,8 @@ class TestPruneOldSamples:
         # Run pruning
         pruned = prune_old_samples(store_conn, now, retention_days=14)
         
-        # Old sample should be removed (interval is derived)
+        # Keep the source samples until local-day replacement evidence exists.
         cursor = store_conn.execute(
             "SELECT COUNT(*) FROM samples WHERE ts = '2026-09-10T10:00:00+00:00'"
         )
-        assert cursor.fetchone()[0] == 0
+        assert cursor.fetchone()[0] == 1

tests/test_repair.py

@@ -235,8 +235,8 @@ class TestBoundaryAnchorRetention:
         # the interval spans the retention boundary
         assert needs_boundary_anchor(store_conn, "2026-09-15T23:55:00+00:00", now)
 
-    def test_not_boundary_anchor_if_fully_derived(self, store_conn):
-        """Sample that's fully derived is not a boundary anchor."""
+    def test_hour_only_derivation_does_not_replace_local_day_evidence(self, store_conn):
+        """UTC hour evidence alone does not make a sample safe to prune."""
         now = datetime(2026, 9, 30, 12, 0, 0, tzinfo=timezone.utc)
         
         # Insert sample and fully derive its interval
@@ -247,8 +247,7 @@ class TestBoundaryAnchorRetention:
         _insert_hour(store_conn, "2026-09-14T10:00:00+00:00", 
                      bytes_written_delta=1000000)
         
-        # Not a boundary anchor
-        assert not needs_boundary_anchor(store_conn, "2026-09-14T10:00:00+00:00", now)
+        assert needs_boundary_anchor(store_conn, "2026-09-14T10:00:00+00:00", now)
 
     def test_pruning_retains_boundary_anchors(self, store_conn):
         """Pruning keeps samples needed as boundary anchors."""
@@ -274,8 +273,8 @@ class TestBoundaryAnchorRetention:
         )
         assert cursor.fetchone()[0] == 1
 
-    def test_pruning_removes_old_sample_with_derived_interval(self, store_conn):
-        """Pruning removes old samples when interval is fully derived."""
+    def test_pruning_keeps_samples_without_local_day_replacement(self, store_conn):
+        """Pruning keeps samples when only the UTC-hour row exists."""
         now = datetime(2026, 9, 30, 12, 0, 0, tzinfo=timezone.utc)
         
         # Old sample with derived interval
@@ -289,11 +288,11 @@ class TestBoundaryAnchorRetention:
         # Run pruning
         pruned = prune_old_samples(store_conn, now, retention_days=14)
         
-        # Old sample should be removed (interval is derived)
+        # The source remains until local-day evidence is also durable.
         cursor = store_conn.execute(
             "SELECT COUNT(*) FROM samples WHERE ts = '2026-09-10T10:00:00+00:00'"
         )
-        assert cursor.fetchone()[0] == 0
+        assert cursor.fetchone()[0] == 1
 
 
 # ---------------------------------------------------------------------------

tests/test_signing.py

@@ -259,8 +259,8 @@ class TestKeyCeremonyDoc:
 
     def test_documents_private_key_storage(self):
         content = self._doc_content()
-        assert "password manager" in content.lower(), \
-            "Must document that private key lives in password manager"
+        assert "gitea repository actions secret `gpg_private_key`" in content.lower(), \
+            "Must document that Gitea Actions stores the private signing key"
 
 
 # ---------------------------------------------------------------------------

tests/test_tui.py

@@ -60,6 +60,13 @@ def _clock(year=2026, month=9, day=30, hour=12):
     return datetime(year, month, day, hour, 0, 0, tzinfo=timezone.utc)
 
 
+class _FrozenTuiDateTime(datetime):
+    @classmethod
+    def now(cls, tz=None):
+        instant = datetime(2026, 9, 28, 12, 0, 0, tzinfo=timezone.utc)
+        return instant.replace(tzinfo=None) if tz is None else instant.astimezone(tz)
+
+
 def _insert_baseline(conn, tbw_tb=1.0, verified=True,
                       model="Samsung SSD 970 EVO Plus 1TB"):
     conn.execute(
@@ -124,8 +131,8 @@ def _insert_local_day(conn, local_date, tz_name="UTC", tz_offset="+00:00",
     conn.execute(
         "INSERT INTO local_days "
         "(local_date, tz_name, tz_offset, utc_start, utc_end, "
-        " bytes_written, bytes_read, coverage, sample_count, complete) "
-        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+        " bytes_written, bytes_read, coverage, sample_count, complete, activity_intervals) "
+        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 1)",
         (local_date, tz_name, tz_offset, utc_start, utc_end,
          bw, br, coverage, samples, complete),
     )
@@ -927,6 +934,7 @@ class TestBarGraphTUI:
         for i in range(14):
             d = (datetime(2026, 9, 15) + timedelta(days=i)).strftime("%Y-%m-%d")
             _insert_day(conn, d, bw=1024*1024*100)
+            _insert_local_day(conn, d, bw=1024*1024*100)
         _insert_sample(conn, "2026-09-30T10:00:00+00:00")
         conn.close()
 
@@ -941,11 +949,9 @@ class TestBarGraphTUI:
             assert "Unalloc" in legend
             assert "Zero" in legend
             range_label = str(app.query_one("#bar-range").render())
-            assert "14 days" in range_label
-            assert "UTC" in range_label
+            assert "14 local days" in range_label
             plotted = str(app.query_one("#bar-render").render())
             assert "Writes (" in legend
-            assert "UTC" in range_label
             assert "─" in plotted
 
     @pytest.mark.asyncio
@@ -957,6 +963,7 @@ class TestBarGraphTUI:
         for i in range(14):
             d = (datetime(2026, 9, 15) + timedelta(days=i)).strftime("%Y-%m-%d")
             _insert_day(conn, d, bw=1024*1024*100)
+            _insert_local_day(conn, d, bw=1024*1024*100)
         _insert_sample(conn, "2026-09-30T10:00:00+00:00")
         conn.close()
 
@@ -1022,7 +1029,7 @@ class TestBarGraphTUI:
             assert len(graph._day_data) == 90
 
     @pytest.mark.asyncio
-    async def test_readout_updates_on_selection(self, tmp_path):
+    async def test_readout_updates_on_selection(self, tmp_path, monkeypatch):
         """Readout shows selected day info."""
         conn = init_store(tmp_path / "test.db")
         _insert_segment(conn)
@@ -1030,9 +1037,11 @@ class TestBarGraphTUI:
         for i in range(14):
             d = (datetime(2026, 9, 15) + timedelta(days=i)).strftime("%Y-%m-%d")
             _insert_day(conn, d, bw=1024*1024*100)
+            _insert_local_day(conn, d, bw=1024*1024*100)
         _insert_sample(conn, "2026-09-30T10:00:00+00:00")
         conn.close()
 
+        monkeypatch.setattr("fenris.tui.datetime", _FrozenTuiDateTime)
         app = FenrisTuiApp(store_path=tmp_path / "test.db")
         async with app.run_test(size=(80, 24)) as pilot:
             await pilot.press("v", "v")
@@ -1043,18 +1052,19 @@ class TestBarGraphTUI:
 
             # Newest day is selected initially.
             readout = str(app.query_one("#bar-readout").render())
-            assert "UTC" in readout
-            assert "GB" in readout
+            assert "totals so far" in readout
+            assert "W 0.105 GB known" in readout
 
             # Moving left updates the selected-day readout.
             await pilot.press("left")
             await pilot.pause()
             readout = str(app.query_one("#bar-readout").render())
             assert "2026-09" in readout
-            assert "GB" in readout
+            assert "incomplete" in readout
+            assert "W 0.105 GB known" in readout
 
     @pytest.mark.asyncio
-    async def test_hourly_drill_down_and_back(self, tmp_path):
+    async def test_hourly_drill_down_and_back(self, tmp_path, monkeypatch):
         """Enter drills into hourly view, Esc returns to daily."""
         conn = init_store(tmp_path / "test.db")
         _insert_segment(conn)
@@ -1062,6 +1072,7 @@ class TestBarGraphTUI:
         for i in range(14):
             d = (datetime(2026, 9, 15) + timedelta(days=i)).strftime("%Y-%m-%d")
             _insert_day(conn, d, bw=1024*1024*100)
+            _insert_local_day(conn, d, bw=1024*1024*100)
             # Insert hours for each day
             for h in range(24):
                 hour = "%sT%02d:00:00+00:00" % (d, h)
@@ -1070,6 +1081,7 @@ class TestBarGraphTUI:
         _insert_sample(conn, "2026-09-30T10:00:00+00:00")
         conn.close()
 
+        monkeypatch.setattr("fenris.tui.datetime", _FrozenTuiDateTime)
         app = FenrisTuiApp(store_path=tmp_path / "test.db")
         async with app.run_test(size=(80, 24)) as pilot:
             await pilot.press("v", "v")
@@ -1078,30 +1090,27 @@ class TestBarGraphTUI:
             await pilot.pause()
             await pilot.pause()
 
-            # The newest day is selected automatically.
-            assert graph.selected_index == len(graph._day_data) - 1
-            assert graph.view_mode == "daily"
+            assert app.query_one("#activity-tabs").active == "view-history"
+            daily_readout = str(app.query_one("#bar-readout").render())
 
             # Enter drill-down
             await pilot.press("enter")
             await pilot.pause()
-            assert graph.view_mode == "hourly"
-            assert graph.drill_day is not None
-            assert len(graph._hour_data) == 24
+            assert app.query_one("#activity-tabs").active == "view-day"
+            assert "UTC" in str(app.query_one("#bar-readout").render())
 
-            # Five-minute data refreshes retain the selected hour.
-            graph._hourly_selected = 12
-            selected_hour = graph._hour_data[12]["hour"]
+            # Inspection identity survives refresh through visible readout.
+            await pilot.press("left")
+            selected_hour = str(app.query_one("#bar-readout").render())
             app._refresh()
             await pilot.pause()
-            assert graph.view_mode == "hourly"
-            assert graph._hour_data[graph._hourly_selected]["hour"] == selected_hour
+            assert str(app.query_one("#bar-readout").render()) == selected_hour
 
             # Esc returns to daily
             await pilot.press("escape")
             await pilot.pause()
-            assert graph.view_mode == "daily"
-            assert graph.drill_day is None
+            assert app.query_one("#activity-tabs").active == "view-history"
+            assert str(app.query_one("#bar-readout").render()) == daily_readout
 
     @pytest.mark.asyncio
     async def test_empty_store_graph(self, tmp_path):