Built-in file redaction for local files

CLAUDE.md

@@ -50,6 +50,8 @@ python -m pytest tests/ -q
 
 182 tests in `tests/`. No integration tests for live M365/Google connections.
 
+**`tests/test_google_scan.py`** — 19 tests for the Google Workspace scan module. Route tests for `GET /api/google/scan/users`, `POST /api/google/scan/start`, `POST /api/google/scan/cancel`. Engine tests for `_run_google_scan` using synchronous invocation with mocked `broadcast`, `_scan_bytes`, `checkpoint.*`, `scan_engine._with_disposition`, and `gdpr_db.get_db`. The `clean_google_state` autouse fixture releases `_google_scan_lock` and clears `_google_scan_abort` after each test.
+
 **`tests/test_route_integration.py`** — 54 Flask test-client tests covering security-sensitive paths: viewer token CRUD and scope validation, `GET /api/db/flagged` role/user scope enforcement, bulk disposition isolation, viewer PIN (set/verify/rate-limit/change/clear), interface PIN gate (multi-step flows require `session["interface_ok"] = True` after PIN set — the `before_request` hook blocks the same endpoint once a PIN exists), scan lock release on `run_scan()` exception, `GET /api/db/sessions` shape and ordering, profile routes CRUD and rename (including the rename-after-copy regression). Uses a tmp-path `ScanDB` monkeypatched into `routes.database._get_db` — tests never touch the real database. Interface PIN tests manipulate the real `config.json` via `setup_method`/`teardown_method` calling `clear_interface_pin()`.
 
 **Local-file scan fixtures** — `tests/fixtures/local_files/` holds 19 files for manual/UI-level testing of the file scanner. 14 should be flagged; 5 are true negatives. All CPR numbers verified against `is_valid_cpr`. `generate_fixtures.py` (requires `python-docx`, `openpyxl`, `mutagen` — all in venv) regenerates the binary `.docx`/`.xlsx`/`.mp3`/`.flac`/`.mp4` files. Audio fixtures need 2 silent MPEG frames so mutagen can sync; FLAC uses a hand-packed STREAMINFO + Vorbis comment block; MP4 uses a minimal `ftyp`+`moov`/`mvhd` base that mutagen can tag.
@@ -111,6 +113,7 @@ Exception hierarchy (all inherit `M365Error(Exception)`):
 Large M365 tenants can generate enormous memory pressure. Key rules to preserve:
 
 - **Email body stripped at collection time** — `_scan_user_email` calls `conn.get_message_body_text(msg)`, stores the result as `msg["_precomputed_body"]`, then deletes `msg["body"]` and `msg["bodyPreview"]` before appending to `work_items`. The processing loop reads `meta.pop("_precomputed_body", "")`. Do not re-add `body` to the `$select` query without also stripping it here.
+- **`body_excerpt` — 500-char plain-text preview stored per flagged email** — just before `del body_text` in M365 email processing, `meta["_body_excerpt"] = body_text[:500].strip()`. In `google_scan.py`, a regex HTML-strip of the first 3000 bytes of Gmail body data is stored the same way. `_broadcast_card` in both engines includes `"body_excerpt"` in the card dict so the excerpt flows into `flagged_items`, the checkpoint JSON, and the DB (`body_excerpt TEXT`, migration #10). The M365 email preview route falls back to `_excerpt_page()` when Graph raises or the connector is absent. The Gmail preview shows `_excerpt_page()` as primary content with the "Open in Gmail" link appended. Do not remove the excerpt before broadcasting — that's what makes preview work on checkpoint resume.
 - **`work_items` → `deque` before processing** — converted with `deque(work_items)` and drained via `popleft()` so each item's memory is released immediately after processing. Do not convert back to a list or iterate with `enumerate()`.
 - **`del content` in file branch** — raw download bytes are deleted as soon as `content.decode()` is done (before NER/PII counting). Both the hit and no-hit paths have explicit `del content`.
 - **`del body_text` in email branch** — deleted after `_broadcast_card` call.
@@ -124,6 +127,7 @@ Large M365 tenants can generate enormous memory pressure. Key rules to preserve:
 - **Excel Summary sheet vs. per-source tabs** — the Summary sheet shows all scanned sources (even with 0 items). Per-source tabs are only created for sources with items; an empty tab has no value.
 - **ART.30 breakdown table** — iterates `scanned_sources` (not `by_source`) so Gmail, Google Drive, etc. appear with `0 | 0 | 0 | —` when the scan found nothing.
 - **Role-filtered exports** — `_build_excel_bytes(role='')` and `_build_article30_docx(role='')` accept `role='student'` or `role='staff'`. A local `_items` list is built at the top of each function and used everywhere instead of `state.flagged_items` directly — GPS sheet, External transfers sheet, and Art.30 staff/student tables all see only the filtered subset. Route handlers read `request.args.get('role', '')` and forward it. Filenames get `_elever` / `_ansatte` suffix. The `#filterRole` dropdown in the filter bar drives both the client-side grid filter and the export URL param — do not separate them.
+- **`POST /api/redact_item`** — rewrites a local file in-place with CPR numbers replaced by `██████-████` / `█` blocks, then removes the card from the grid and logs a `"redacted"` disposition. Supported extensions: `.docx`, `.xlsx`, `.csv`, `.txt` (`_REDACT_EXTS`). The file is written to a temp path in the **same directory** as the original before `shutil.move` — this avoids cross-device rename failures on mounted volumes. Uses existing `document_scanner` functions (`redact_docx`, `redact_xlsx`, `redact_csv`, `find_pii_spans_in_text`). Only works for `source_type == "local"` — SMB/cloud files are not supported (button is hidden on those cards). The button (`✂`, class `card-redact-btn`) appears in `appendCard` when `_redactable(f)` is true; hidden in viewer mode and for resolved items.
 
 ## Scan history browser — static/js/history.js + gdpr_db.py + routes/database.py
 
@@ -137,6 +141,7 @@ Allows reviewing results from any past scan session without running a new scan.
 - **History banner** (`#historyBanner`) — shown when `S._historyRefScanId` is set. Contains `#historyBannerText` (session date · sources · N items), `#historyPickerBtn` (opens `#historyDropdown`), and `#historyLatestBtn` (visible only when the viewed session is not the latest). Do not hide/show these elements from outside `history.js`.
 - **Session picker** (`#historyDropdown`) — rendered inside `[data-history-wrap]` container so the outside-click handler (`document` listener, closes on clicks outside `[data-history-wrap]`) works correctly. Do not move the picker outside this wrapper.
 - **Cache invalidation** — `_sessions` and `_latestRefScanId` are module-level in `history.js`. `invalidateHistoryCache()` clears both. All three `*_done` SSE handlers in `scan.js` call `window.invalidateHistoryCache?.()` so the picker reflects the newest scan after completion.
+- **Re-scan diff** — `loadHistorySession` fetches the immediately preceding session's items after rendering the current session. Items present in the previous session but absent from the current one (compared by `id`) are tagged `_resolved: true` and appended after a `.resolved-divider` separator. `appendCard` in `results.js` adds `.card-resolved` (opacity 0.6), a green `✓ Resolved` badge, and hides the delete button for resolved items. `_setHistoryBanner` accepts an optional `resolvedCount` parameter and appends it to the banner label. Resolved items are NOT added to `S.flaggedData` — they are grid-only and cannot be bulk-selected or exported.
 - **Auto-load on page load** — `results.js` calls `window.loadHistorySession?.(null)` once when the SSE watchdog confirms `!status.running`. `null` resolves to the latest completed session via `_fetchSessions()[0].ref_scan_id`. The `_initialStatusChecked` guard ensures this fires at most once per page load.
 - **Mode transitions** — `startScan()` calls `window.exitHistoryMode?.()` before clearing the grid, so any history banner is dismissed and `S._historyRefScanId` is reset before SSE events start arriving.