genshin-assistant/docs/LOCALIZATION_LOCAL_DELETION_TIMING_CHECKPOINT_2026-07-11.md

# Localization, Local Deletion, Timing, and OCR Corpus Checkpoint

Date: 2026-07-11

Status: implementation, offline-validation, and current-build live-smoke
checkpoint. The non-persisting 50-artifact development smoke passed on
2026-07-11; its durable evidence is in
[NATIVE_SCANNER_LIVE_SMOKE_2026-07-11.md](NATIVE_SCANNER_LIVE_SMOKE_2026-07-11.md).
An elevated packaged executable remains a separate package-runtime boundary and
is not claimed here.

## Delivered Scope

### English-default renderer with German Settings option

- `en` is the default for a fresh, missing, or invalid persisted locale.
- Settings exposes English and German; the selected app locale is persisted
  locally and keeps the document language in sync.
- Visible app-shell, navigation, Scanner, Artifacts, Review, Builds, Overlay,
  Diagnostics, Settings, and local-delete chrome use the renderer translation
  catalog. Source/domain evidence remains deliberately distinct from translated
  UI chrome.
- The setting changes application copy only. Smart Capture remains validated
  for the English Genshin UI; it does not claim German Genshin OCR/crop support.
- The package UI runner uses stable `data-*` hooks instead of translated text
  for core interaction checks.

### Local artifact removal

The delete control corrects app-local data only. It is not a Genshin action.

| Selected row | Local effect | Preserved / prohibited behavior |
| --- | --- | --- |
| Native scan result | Appends a `deleted-results.jsonl` tombstone, hides the row, and removes its crop only if the resolved PNG stays inside the selected run directory. | Raw `scan-results.json` remains replay evidence. Active capture/processing blocks removal. No input helper or game action is called. |
| Exact Store record linked to native row | Removed only when the user explicitly includes that exact local record in the confirmed removal. | No broad Store deletion and no inferred record matching. |
| Store-only row | Removes that one explicit local Store record through main-process IPC. | No native scan evidence or Genshin state is touched. |

The workflow is idempotent for an already tombstoned native row. It retains the
native run's raw result array for audit/replay and limits file deletion to
contained PNG paths.

### Shared-clock native run timing

Each native run can now write atomic `run-timing.json` evidence with version
`native-scanner-run-timing-v1`.

- Markers are first-observed and immutable after a valid value is written.
- The record covers request start, native scanner start, capture start/end,
  processing start/end, durable results, and final reconciliation.
- `requestToResultsDurableMs` and `requestToResultsReconciledMs` are derived
  directly from their endpoints and validated against timestamp ordering.
- Capture and processing remain separate metrics. They may overlap and are
  never added to create an end-to-end duration.

### Confirmed OCR corpus changes

The historical 136-result Review population was inspected against saved crops,
not treated as self-labeled training data:

| Audit result | Count | Current behavior |
| --- | ---: | --- |
| Explicit visual 1–4★ | 128 | Remain excluded/manual-inspection entries, not value-scored parser labels. |
| Confirmed `HP+4..1%` OCR punctuation defect | 6 | Exact duplicated decimal punctuation is repaired to `HP+4.1%`; the regression cases are in the confirmed corpus. |
| Confirmed 5★ partial-name case with direct slot evidence | 1 | The unsafe global alias was removed. Canonicalization is now allowed only when independent high-confidence slot evidence makes one piece valid. |
| Incomplete/malformed 5★ crop | 1 | Remains `Review`; no heuristic converts it into a clean artifact. |

This preserves the fail-closed rule: ambiguous OCR becomes unknown or Review,
not a guessed canonical artifact.

## Offline Evidence Recorded For This Checkpoint

- Focused localization tests cover English fallback, persisted locale behavior,
  and English/German catalog-key parity.
- Native deletion tests cover tombstone filtering, raw-evidence preservation,
  active-run blocking, contained crop cleanup, idempotency, and exact Store
  scope.
- Native timing tests cover first-observation behavior, ordering validation, and
  derived durations.
- Parser/corpus evaluation covers the duplicate-decimal repair and slot-safe
  name behavior. The latest saved 50-artifact evidence validation completed at
  `50/50` with zero validation issues.
- The final current-build offline gate passed after all localization, deletion,
  timing, corpus, and acceptance-runner changes: `npm run acceptance:offline`
  completed TypeScript, 72 test files / 444 tests, OCR evaluation, saved-run
  validation, package verification (27/27), both dependency audits, and the
  diff check.
- The built-renderer Builds/UI acceptance passed fresh English, German
  selection, English restoration, Builds context/focus, reduced motion,
  overflow, and native plus Store local-delete confirmation probes. It did not
  launch an elevated packaged executable, so it is not
  UAC/preload/helper/native-runtime evidence.

## Remaining Release Gates

1. If a package-runtime/UAC claim is needed, separately launch the elevated
   packaged executable and verify its bridge/helper/runtime behavior. Do not
   substitute the passed built-renderer UI acceptance for that evidence.

2. The required elevated development 50-artifact smoke now passes: 50/50
   capture, click, queue, parse, durable-result reconciliation, zero Review,
   zero processing errors, zero Store writes, and a complete valid
   `run-timing.json`. It measured `28,022 ms` request-to-reconciliation, so it
   does not claim the separate 3/s performance target.

3. Completed on 2026-07-11: generated workspace outputs were removed and the
   coherent checkpoint was committed and pushed as `8b9f948`. The running
   admin runtime's locked log was intentionally preserved.

## Follow-on Quality Phase

- Extend the corpus only from new visually confirmed crops and preserve
  ambiguity as `Review`. The historical 136-item Review population has already
  been audited; it is not an outstanding relabeling task.
- Refresh Build-Fit source profiles before their explicit expiry and add Golden
  plus adversarial fixtures for each profile expansion.
- Keep the bounded wheel reset as the production baseline; test any alternate
  inventory-entry/reset choreography as a separate guarded experiment.
- Keep Builds read-only and Artifact-only until a separately sourced product
  decision introduces character, weapon, team, or rotation inputs.