fix(semantic): Re-evaluate stored vectors and entries when privacy policy changes

[mhiro2]

Summary

• The semantic index froze each entry's capture-time sensitivity forever: adding a regex_denylist / app_denylist rule (or enabling OTP detection) after an entry was embedded neither deleted its stored vector nor stopped a manual Rebuild from re-embedding the same forbidden text verbatim
• Fingerprint the embedding privacy policy (AppSettings::semantic_policy_hash) and store it with the index metadata, so any policy edit purges the stored vectors immediately — even while the index is disabled or no embedding model is available — and gates semantic queries until the rebuild lands
• Re-assess every entry against the current policy right before embedding, over every persisted text projection (raw body, normalized text, rich-text markup, stored representations): re-assessed Secret/Blocked entries are refused and tombstoned in a new semantic_exclusions table, Private ones are redacted, and only Public text reaches the model verbatim
• Make the settings write the atomicity anchor: a policy-changing save erases the index in its own transaction, the indexer's guarded upsert re-checks the policy hash inside the write transaction, and the ranking SQL re-checks it inside the query snapshot — so no interleaving persists or serves old-policy vectors past the settings commit
• Migration 105 rebuilds the meta table with the policy_hash column and deletes pre-existing vectors outright, since they were embedded under a policy whose fingerprint predates tracking; policy-driven erases follow the hard-delete WAL-scrub contract

Changes

• 3f063cf : fix(semantic): re-evaluate stored vectors and entries when privacy policy changes
  • Track the embedding privacy policy as a fingerprint over regex_denylist, app_denylist, OTP detection, and the entry-size ceiling; include it in SemanticIndexMeta::is_compatible_with so a policy edit invalidates the index like a model change, purged before the enabled / availability / battery guards (the no-embedder loop now watches settings too)
  • Re-assess pending entries at embed time via the new SensitivityClassifier assessment API, treating the stored verdict as a floor: refused entries get a semantic_exclusions tombstone (also dropping any leftover vector) that keeps them out of the backlog until their content or the policy changes
  • Add regression tests covering the three review scenarios (a new rule erases an already-embedded Public row's vector, Rebuild never hands the raw text to the embedder, and the purge still runs with no embedding model available); update ARCHITECTURE.md and docs/privacy.md to describe the actual lifecycle
• cc87c58 : fix(semantic): close policy races and widen re-assessment to all stored payloads
  • Drop an in-flight batch when the settings watch reports a change before the model call or before persisting, so vectors shaped by the previous policy never become durable
  • Widen the re-assessment to every persisted text projection — rich-text markup and stored HTML/RTF/plain representation rows included — with the capture classifier's aggregate-size shape, and refuse bodies whose content_json cannot be parsed back (fail closed)
  • Pin the semantic ranking SQL to the validated policy fingerprint inside its own snapshot, anti-join the exclusion tombstones in search, and follow the hard-delete WAL-scrub contract (wal_checkpoint(TRUNCATE)) after policy purges, exclusions that dropped vectors, and applied migrations
• a794315 : fix(semantic): anchor policy/index atomicity on the settings write
  • Erase the semantic index (vectors, tombstones, metadata) in the same transaction as any settings save whose policy fingerprint changed, on both the plain and revision-checked save paths, so every snapshot either predates the save or sees no vectors — closing the races the watch checks and query pin could not close alone
  • Persist indexing batches through a guarded upsert that re-checks the expected policy hash against semantic_index_meta inside the write transaction, aborting when a policy-changing save erased it mid-flight
  • Decode text/uri-list (FilePaths) representations back to the newline-joined shape the capture-time classifier scanned, so anchored or quote-sensitive rules re-assess identically
• 0f310f6 : ci: bust the stale Linux apt package cache
  • Bump cache-apt-pkgs-action's version key in ci.yaml and release.yaml so the Linux package set is resolved fresh against the updated ubuntu-latest image, fixing the glib-sys / glib-2.0 build failure caused by a stale cached package set omitting transitive dev dependencies
• 6faa98f : fix(semantic): make the rebuild and exclusion writes policy-guarded too
  • Replace the worker's separate clear + set-metadata calls with semantic_rebuild_index_meta, a single write transaction validated against the committed settings row, so a pass shaped under previous settings can never resurrect the old policy fingerprint after a policy-changing save erased it
  • Apply the same in-transaction policy guard to exclusion tombstone writes, and open the guarded write transactions IMMEDIATE to avoid SQLITE_BUSY_SNAPSHOT on deferred read-to-write upgrades