diff --git a/devlog/_plan/260827_bug_pr_merge_round/050_round_closeout.md b/devlog/_plan/260827_bug_pr_merge_round/050_round_closeout.md
new file mode 100644
index 0000000000..d208876743
--- /dev/null
+++ b/devlog/_plan/260827_bug_pr_merge_round/050_round_closeout.md
@@ -0,0 +1,79 @@
+# 260827 bug-PR merge round — close-out
+
+All 12 PRs dispositioned. `dev` advanced from `9b838d062` to `77e037077`.
+
+| PR | lane | state | evidence |
+|---|---|---|---|
+| #2672 | L1 | MERGED | `2a9c18dc5` in `2feffbdc3` |
+| #2674 | L1 | MERGED | `e3b136fb7` in `2feffbdc3` |
+| #2671 | L1 | MERGED | `17aadf88e` + added test `4a4df12f2` |
+| #2684 | L1 | MERGED | `2c85dd48d` in `2feffbdc3` |
+| #2639 | L3 | CLOSED, partially landed | status half in `64c6d642b`; `created_at` held back |
+| #2647 | L3 | CLOSED, landed | three ladders in `64c6d642b` |
+| #2663 | L2 | CLOSED, landed | squashed `cb9bb9b76` in `cebe005db` |
+| #2690 | L4 | CLOSED, landed | `669efd568` in `77e037077` |
+| #2694 | L4 | CLOSED, NOOP | superseded by #2663; proven by probe |
+| #2693 | L4 | **OPEN — BLOCKED** | upstream question posted |
+| #2638 | L4 | **OPEN — NEEDS_HUMAN** | auth surface, security review |
+| #2497 | L4 | **OPEN — NEEDS_HUMAN** | OAuth surface + 5-file conflict |
+
+## What the round actually caught
+
+Two PRs were not what their GitHub status claimed. #2694 was review-ready with five
+green checks, a ticked "all CI tests are green on my local testing" box, and five
+`tsc` errors including a call to a function defined nowhere. #2693 was a test-only
+diff whose test failed on its own branch. Neither was caught by CI, because the five
+checks those PRs ran compile and test nothing.
+
+That is the round's most transferable finding: **on this repository, "checks are
+green" is not evidence of health unless the list includes `ci` / `test N/4` /
+`macos`.** For draft and contributor PRs it usually does not.
+
+## What the round got wrong, and who caught it
+
+Independent reviewers audited three lanes and returned FAIL twice. Seven of my own
+claims were wrong or overstated:
+
+1. Compile-gated stale PR heads instead of merged trees (heads were 4-294 commits behind).
+2. Claimed #2684 and #2690 "do not textually conflict" — they do, provably.
+3. Cited "#2663 CI is green" — the same five non-compiling checks I had just discredited.
+4. Invented a CI mechanism ("full matrix only on maintainer PRs") — #2639 has all 27 checks and a non-maintainer author.
+5. Called #2690 an L3 cherry-pick when its fix imports the module its refactor creates.
+6. Framed the status/created_at split as principled when it survives on fixture contents.
+7. Justified unverified effort ladders by a self-correction mechanism that does not function.
+
+Number 7 is the one worth remembering. I read `refreshCommandCodeReasoningEfforts`,
+saw it re-read the profile and replace the row, and cited that as the reason unverified
+data was acceptable. The reviewer ran it against the live site: every profile page
+returns 200 with an empty `reasoningEfforts` array and no parseable prose, for every
+row in the table. **A safety net that exists in the code is not a safety net that
+functions.**
+
+The reviewers also found two real defects in code I had already verified: a `queued`
+response marking unstarted messages `completed`, and a regenerated fixture
+resurrecting a deliberately removed model behind a count-only assertion.
+
+## Sequencing paid off once, concretely
+
+Landing #2684 before #2690 was decided by `git merge-tree`, not preference. #2690
+deletes the region #2684 edits, so the 92-line host-scoped change went first and the
+larger extraction became a clean rebase the author performed themselves. The other
+order would have forced #2684 to be rewritten.
+
+## Operational notes
+
+- `dev` is push-protected; every lane travelled through a `codex/` branch and a PR.
+- The full suite belongs on `ssh lidge` via `ocx-run`, not this workstation.
+- Do NOT set `OCX_TEST_NO_QUEUE=1` to bypass a stuck lock: it leaks into the child
+  process `tests/test-runner.test.ts` spawns and fails the three machine-lock cases.
+  A stale root-owned `/tmp/opencodex-bun-test.lock` should be removed instead.
+- `cxc receipt test` flipped `core.bare` to true twice on long runs, making the main
+  checkout report as bare. Recovered with `git config --local core.bare false`; no data
+  lost. Cause not established.
+
+## Open follow-ups
+
+1. `parsedProfileEfforts` cannot read commandcode.ai profiles — dead for every row.
+2. Field backfill mutates verbatim passthrough bodies; `status` and `created_at`
+   should be decided together.
+3. #2671's "probed 2026-08-26" provenance is unverified on a merged change.