Skip to content

FINAL_REPORT.md

Latest commit

 

History

History
121 lines (78 loc) · 6.25 KB

FINAL_REPORT.md

File metadata and controls

121 lines (78 loc) · 6.25 KB
Raw

KeplerNightLab Phase U2.2 Physical Closure Report

1. Baseline / Final HEAD

  • Baseline HEAD: 146842e4e32f5a8cbe211fe02375374beece769b
  • Final HEAD: 146842e4e32f5a8cbe211fe02375374beece769b (unchanged)
  • Parent: 07613f34c9999a5a5ccf5074720b731716e7929a

2. Final Git Status

 M app/src/main/java/com/projectnuke/keplernightlab/MediaStoreExportRecovery.kt

No other production changes. Instrumentation added for measurement was removed before finalization.

3. Three Process-Cold Timing Rows

Run Total Recovery (ms) Atomic Write Delta inspectMs Median exportRecoveryMs Median metaReconcileMs Median initialJobReadMs Median
1 16,786 84 330 333 1 3
2 16,426 92 324 330.5 1 2
3 16,193 92 322 329 1 2

4. Recovery Median

16,426 ms (median of 16,786 / 16,426 / 16,193)

5. exportRecovery Median

330 ms (median of 333 / 330.5 / 329)

6. exportJournalInspection Median

330 ms (inspection dominates exportRecovery; 46 individual inspectMs values per run)

7. Total Atomic-Write Count Per Run

  • Run 1: 84
  • Run 2: 92
  • Run 3: 92

8. VERIFIED→VERIFIED Write Count

0 — Same-state idempotency is working. No journal was rewritten when already in VERIFIED state.

9. RAW-Sidecar No-Op Job Write Count

0 — Not applicable for YUV_NIGHT_FUSION cohort. RAW-sidecar structural gate correctly skips YUV jobs.

10. STABLE Rewrite Count

0 — Fast-path skip is working. Jobs already in STABLE state with no recoveryMessage are not rewritten.

11. Attribution of Remaining Writes

  • 46 job.json writes: PUBLIC_EXPORT_COMMITTED_PENDING_VERIFICATION path in recoverOne (line 332). These are semantically required: the seed cohort was created with recoveryState=STABLE but the recovery code re-classifies them based on current MediaStore inspection results (exists=true, verified=false, pending=false → PUBLIC_COMMITTED_UNVERIFIED → PUBLIC_EXPORT_COMMITTED_PENDING_VERIFICATION).
  • 46 export journal writes: markTerminalPersisted during terminal settlement. Even though journals were seeded with terminalMetadataPersisted=true, the recovery path processes them through the normal terminal settlement flow.

Total: 92 atomic writes (46 job.json + 46 export journal files)

12. Before U2.2 vs After U2.2 Recovery Median

No baseline timing data exists from before U2.2. The timing instrumentation was added during this measurement phase. Comparison is not available.

13. Percentage Improvement

N/A — No pre-U2.2 baseline measurement exists.

14. Before vs After Atomic-Write Count

N/A — No pre-U2.2 baseline measurement exists.

15. Confirmation 46/46 Full MediaStore Inspections Still Executed

YES — All 46 MAIN_IMAGE journals triggered ContextMediaStoreExportRecoveryAccess.inspect() with the same production verification path:

  • Exact URI query
  • Row existence check
  • IS_PENDING read
  • verifyGalleryExportResult (stream/signature integrity probe, bounds decode, sampled decode)
  • Result: all 46 returned exists=true, verified=false, pending=false

16. External-Removal Control Result

PASS — ExternalPublicRemovalRecoveryTest (9 tests) all passed:

  • verifiedTerminalRowMissing_classifiesRemoved_settlesStable_truthfulMetadata
  • externalRemoval_recoveryIsIdempotent_staysStable
  • verifiedWithoutTerminalAck_rowMissing_keepsPublicCommitMissingBlocked
  • committedUnverifiedRowMissing_keepsExistingDebtPolicy
  • oldVerifiedExportRemoved_newReprocessExportVerified_oldMissingDoesNotOverrideNew
  • twoHistoricalMissingExports_currentVerifiedExportRemainsStable
  • externalDeletion_doesNotBlockDeleteCleanupOrReprocess
  • destructiveLocalIntents_stillBlockedByLiveOwnership
  • sidecarHistoricalMissing_doesNotEraseCurrentMainAuthority

No fallback to PUBLIC_COMMIT_MISSING for terminal-acknowledged evidence. Journal evidence is preserved (not deleted).

17. Multiple-Journal Ordering Result

PASS — GalleryExportIdempotencyTest.multipleVerifiedMainJournals_omitSameStateRewrite_doesNotReorderEvidence passed. updatedAt values are byte/value unchanged after recovery. Selected current authority remains the same.

18. Full Validation Results

Check Result
git diff --check PASS
:app:compileDebugKotlin PASS
:app:compileDebugUnitTestKotlin PASS
Focused unit tests (GalleryExportIdempotencyTest, MediaStoreExportRecoveryTest, KeplerStableFastPathTest, ExternalPublicRemovalRecoveryTest, RawSidecarJournalRecoveryTest) PASS
:app:testDebugUnitTest 1,618 tests completed, 1 pre-existing failure (KeplerRecoveryCoordinatorTest.ordinaryTerminalFinalizationFailureRemainsRecoveryFailedAndPreservesOwner — exists on baseline HEAD)
:app:lintDebug PASS
:app:assembleDebug PASS

19. Remaining Dominant Cold-Recovery Stage

MediaStore export inspection (exportRecovery / inspectMs) dominates at ~330 ms per job × 46 jobs = ~15.2 s (92% of total recovery time). All other stages are negligible (metadata temp reconcile ~1 ms, initial job read ~2 ms).

20. Next Recommendation

U2.3 MediaStore Terminal-Verified Reverification DESIGN REVIEW

MediaStore full verification (URI query, row existence, IS_PENDING, verifyGalleryExportResult with stream/signature integrity probe, bounds decode, sampled decode) is the dominant ~15 s class cost for the 46-job cohort. The same-state idempotency and fast-path optimizations have eliminated redundant writes, but the per-job MediaStore verification latency remains the bottleneck. A design review should explore whether terminal-verified MAIN_IMAGE rows can skip re-verification or use a cheaper existence-only check when durable journal evidence already proves terminal-stable verification — without weakening the verification contract for non-terminal or ambiguous jobs.