Summary

Fix a regression introduced by ca483c5. When the trust-boundary privacy.enforce_write_guard was wired to POST /api/add with force_unsafe: bool = False defaulting to block, the SPA's existing privacy-warning flow in app.js was not updated to thread the bypass flag through the resulting submit.

Result: confirmed Add Memory submissions of secret-shaped content (or false positives) now get a server-side 403 instead of being saved. The confirm dialog became a no-op trap.

This PR threads forceUnsafe = true from a confirmed dialog into the POST body so the user's "I know, store it anyway" choice actually reaches the bypass branch the server already supports.

What changed

web/static/app.js — Add Memory submit handler:

+  let forceUnsafe = false;
   const patterns = STATE.privacyPatterns;
   if (patterns && patterns.some(re => re.test(content))) {
     const ok = await showConfirm({...});
     if (!ok) return;
+    forceUnsafe = true;
   }
   ...
-    const data = await api('POST', '/api/add', { content, title, tags, file, namespace });
+    const body = { content, title, tags, file, namespace };
+    if (forceUnsafe) body.force_unsafe = true;
+    const data = await api('POST', '/api/add', body);

web/static/index.html — ?v=99 → ?v=100 per feedback_static_asset_cache_bust.md.

tests-js/add-memory-force-unsafe.test.mjs — three pinned branches:

1. clean content → POST without force_unsafe
2. flagged content + confirm → POST with force_unsafe: true
3. flagged content + cancel → no POST

Mutation-validated per feedback_pin_test_mutation_validation.md: reverting body.force_unsafe = true makes case (2) fail with expected undefined to be true, so the regression cannot silently re-land.

Out of scope (follow-up)

PATCH /api/chunks/{id} and POST /api/upload accept force_unsafe server-side but the SPA has no UI for it (lines 2151 / 3621 / 3875 / 5405 in app.js). Those are new gaps, not regressions — they never had a confirm flow — and need a different pattern (detect 403 redaction_blocked, show retry-confirm dialog) plus likely a shared helper. Will track in a separate issue so this PR stays a single regression fix per feedback_one_change_per_pr.md.

Test plan

• npx vitest run — 14/14 pass (5 test files)
• Mutation-validated: revert prod fix → case (2) fails as expected → restore → all green
• Static-asset cache-bust bumped (app.js v99 → v100)
• CI green (lint / js-unit / typecheck / 3-OS python tests / Playwright / golden-path)
• Manual smoke (after merge): paste a fake sk- token into Add Memory → confirm dialog appears → confirm → save succeeds → row appears in results

References

• Server gate: web/routes/system.py:1203 (enforce_write_guard(surface="web_api_add")).
• Schema: web/schemas/memory.py:14-18 (the force_unsafe: bool = False field this PR finally exercises from the SPA).
• Companion comment cleanup that triggered the audit: docs(web): fix stale '/api/add does not invoke privacy.scan' comments #786.

🤖 Generated with Claude Code