Prerelease: 3 verified fixes — 500s, 401 retry, Firestore cache (#5441 #5442 #5443)

[beastoin]

Summary

Prerelease branch combining 3 independently verified sub-PRs:

• Fix /v1/users/people and /v1/conversations 500 errors #5441 — Fix /v1/users/people and /v1/conversations 500 errors (@yuki)
• Fix: add 401 token refresh to multipart API calls #5442 — Fix 401 token refresh for multipart API calls (@kenji)
• perf(backend): cache 3 Firestore hot paths to cut read costs ~59% (#5439) #5443 — Cache 3 Firestore hot paths to cut read costs (@Hiro)

Verification

All 3 PRs passed comprehensive verification in PR #5445 by @Kelvin:

• 72/72 unit tests passing
• Clean merge (1 trivial test.sh conflict resolved — both test lines included)
• Live API verification
• Agent-Flutter E2E
• Physical device E2E (Pixel 7a)
• Codex quality audit passed
• 1 blocking issue (BLOCKING: Freemium credits cache has no active invalidation — transcripts dropped for up to 15min after upgrade #5446 credits cache) found and fixed during review

Deploy Request

Please trigger backend deploy after merge — these fixes address active 500 errors and cost optimization. Deploy via "Deploy Backend to Cloud RUN" workflow (covers Cloud Run + GKE backend-listen rollout restart).

Merge Strategy

Each sub-PR was merged with --no-ff to preserve individual commit history.

[beastoin]

lgtm

[greptile-apps]

Greptile Summary

This prerelease combines three independently verified fixes targeting active 500 errors, a 401 token-refresh gap for multipart API calls, and Firestore read cost reduction via in-process caching.

Key changes:

• /v1/users/people 500 fix (users.py, models/other.py): Legacy Firestore person documents missing the id field or timestamp fields now handled gracefully — get_person/get_people inject doc.id via setdefault, get_people_by_ids switches from a where("id", "in", ...) query to db.get_all(doc_refs), and created_at/updated_at are made optional on the Person model.
• /v1/conversations 500 fix (conversations.py router + db): The list endpoint switches from get_conversations (loads full base64 photo content) to get_conversations_without_photos, preventing the 32 MB Cloud Run response-size limit from being exceeded. folder_id and starred filters are added to get_conversations_without_photos and correctly forwarded from the router.
• Multipart 401 retry (shared.dart): Both makeMultipartApiCall and makeMultipartStreamingApiCall now detect 401 responses, force-refresh the auth token, rebuild the multipart request (required since HTTP streams are single-use), and retry. Persistent 401 after retry triggers sign-out. Mirrors the existing retry pattern already present for non-multipart calls.
• Firestore caching (cache_manager.py, notifications.py, utils/apps.py, transcribe.py, redis_db.py): Three hot Firestore paths are wrapped with a 30-second in-memory cache (mentor notification frequency, is_tester flag + per-user app slice) and a 15-minute credit-seconds cache with active Redis invalidation on subscription changes. A pre-existing memory leak in InMemoryCacheManager._fetch_locks (locks never removed after get_or_fetch) is fixed with reference counting. A cache-mutation bug in get_available_apps (mutating cached dict objects) is fixed with a shallow copy.

All production logic changes are straightforward, well-tested (72/72 unit tests), and verified with live API + physical device E2E.

Confidence Score: 4/5

• Safe to merge — all three fixes address active production regressions with low change risk; deploy should be triggered promptly per the PR description.
• All production logic changes are straightforward, well-tested (72/72 unit tests), and verified with live API + physical device E2E. The small deductions: (1) the Dart multipart test validates logic through a standalone reimplementation rather than testing the full production singletons integration, and (2) in-memory cache invalidation is local-only in multi-pod deployments, meaning other pods may briefly serve stale tester data — both are known, acceptable tradeoffs rather than blocking issues.
• No files require special attention. All changes are correct and well-tested.

_{Last reviewed commit: cc5df0b}