chore(api): post-#101 review backlog — valid small findings from umbrella bot reviews

[leon0399]

Umbrella PR #101's bot reviews surfaced small valid findings against code that has since merged via the stacked slices (#104–#107). None are release-blocking; recorded here so the umbrella threads can be resolved without losing them. Each line links the original thread.

Correctness / robustness

• sanitizeTitle strips the Title: prefix before leading markdown, so # Title: / > Title: leak the prefix — swap strip order (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• compaction emits role: 'tool' with flattened string content for absorbed tool messages — invalid shape for the model client once tool calls are stored (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• runs.controller SSE replay: after headers are flushed, a DB error in the poll loop propagates to the exception filter which can't respond — catch, log, break (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• context-builder multi-sender detection runs on pre-compaction history — attribution noise persists after compaction reduces to one sender (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• compaction runs are not serialized per chat — overlapping fire-and-forget compactions can insert divergent rows; staleness guard only rejects committed competitors (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• PATCH /runs/:id on a queued run only stamps cancelRequestedAt — the single-flight slot stays occupied until pickup or deadman; consider settling queued runs terminal immediately (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))

Defense-in-depth (DB constraints)

• run_events RLS is FOR ALL — append-only is convention, not enforced; command-specific policies or a raising trigger (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment), Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• compaction parent_id lineage is not constrained to the same chat — composite FK/check (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment), Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• runs.user_id is not tied to chats.owner_user_id at the DB level (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• compactions.parent_id FK lacks a supporting index (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))

Cost / hygiene

• title generation forwards full turn text with no length cap — token-spike risk (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• terminal run statuses are hard-coded in several repository methods — extract one shared constant (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))
• base-URL model-client spec doesn't assert openai.chat() resolution like sibling tests (Overnight: v0.1 completion + v0.2 durable-run pipeline (umbrella) #101 (comment))

The Stop-button → PATCH /runs/:id wiring is deliberately NOT here — it's the stacked web resume slice (#49).