devlog: record the community bug sweep

devlog/_plan/260825_owner_backlog_and_bugpr_closeout/150_community_bug_sweep.md

@@ -0,0 +1,69 @@
+# 150 — community bug sweep
+
+The plan's DONE criterion was zero `lidge-jun` issues and every bug PR terminal. Both were met
+partway through. The request behind the plan was broader — no outstanding bugs — so the loop
+continued into the `bug`-labelled community backlog rather than stopping at its own inventory.
+
+Sixteen open, three left. Final `dev` at `87179e86a`: 15037 pass / 16 skip / 0 fail, typecheck
+and `privacy:scan` green.
+
+## Six were already fixed by this same run
+
+| Issue | Fixed by | Verified |
+|---|---|---|
+| #2499 Windows catalog-state latency | #2580 (`342911fec7`) | ancestry + 51 pass |
+| #2545 Antigravity thought_signature | #2577 (`4d3d2716e`) | ancestry + 150 pass |
+| #2210 Cursor stall timeout | `994e5ba87` | ancestry + 252 pass |
+| #2156 Muse Spark mid-tool-call | `1a15d6292` + `08cc2ac89` | ancestry + 54 pass |
+| #2300 Cursor slowness | #2307 | ancestry + falsification |
+| #2509 stale pool account (core) | #2515 | ancestry + falsification |
+
+Each was checked with `git merge-base --is-ancestor` **and** a covering test run before closing.
+Closing on a report that something was fixed is how a backlog acquires issues that were never
+actually resolved.
+
+## What the sweep found on its own
+
+**#2458 — the reported cause was one layer off.** A 502 on an undeclared `get_video_duration`
+tool call looked like a bridge problem. The real defect was that Chat Completions silently
+discarded `video_url` (`src/chat/inbound.ts:56`), so the model reasoned about a video it had
+never received, and the bridge refused the resulting call. OpenCodex was rejecting a symptom of
+its own dropped input.
+
+**#2509 — a second half nobody had noticed.** #2515 fixed selection to preview per candidate
+quota scope and #2623 added the entitlement filter, but the encrypted-recovery path got the
+scope and not the filter. Same stale-selection class, one layer over.
+
+**#2459 — the health check was the worst part.** A bare npm reinstall under a live proxy leaves
+`/healthz` answering 200 while every `/v1/responses` fails. A lying health check is worse than
+an honest failure, because every supervisor believes it.
+
+**A load flake, not a defect.** One full-suite failure at exactly 5003 ms — Bun's default 5s
+timeout — on a case that starts a real server while its neighbours declare `SERVER_BUDGET_MS`.
+It passed 3/3 in isolation on two machines. Fixed as a budget, not chased as a bug.
+
+## Three left open, each for a stated reason
+
+**#2221 — native main token refresh.** The safe-looking half is not separable. A refresh rotates
+the grant, `atomicWriteFile` has no compare-and-swap against an external writer, and the Codex
+CLI writes the same `auth.json`. Refreshing without a crash-safe publisher can strand the login:
+the current behavior fails one request, the naive fix can cost the credential. #2497 proposes a
+publisher and is itself held under security review with a rename-before-link crash window. This
+needs a designed protocol, not a smaller patch.
+
+**#1527 — Cursor large-context.** Both mechanisms behind the reported symptoms changed: rate
+limits are excluded from transport retry, and checkpoint continuation replaced full-history
+replay — which was the credible differentiator in the reporter's own proxy-vs-direct comparison.
+The residual claim needs a matched live probe, which only the reporter can run.
+
+**#1419 — macOS SIGTRAP.** The reported offsets symbolicate to Bun's own crash handler and
+nothing past it; the frames that would name the faulting subsystem are absent, and no upstream
+release establishes that 1.4.0 fixes that signature. The runtime is pinned so it cannot regress;
+the crash is unproven. Recorded alongside it: `ocx gui` starts an unref'd detached process, so a
+native crash there strands the proxy with no supervisor, unlike an installed service.
+
+## The pattern worth keeping
+
+Four of these closed as "already fixed" and three as "not fixable here, here is why". Neither is
+a failure to do work. A backlog that only ever accepts code changes as outcomes accumulates
+issues nobody can close, and closes issues nobody actually verified.