retro: 2026-05-27 memory updates

.claude/memory/MEMORY.md

@@ -15,6 +15,7 @@ Keep this file under 200 lines — anything longer is content bloat, not memory.
 - [learnings/vercel-bot-status-as-deploy-health-fallback](learnings/2026-05-26-vercel-bot-status-as-deploy-health-fallback.md) — When Vercel MCP/CLI auth is broken, read the GitHub Vercel-bot's commit status on the latest PR HEAD as a deploy-health fallback
 - [learnings/manager-agent-yaml-needs-manual-bootstrap](learnings/2026-05-26-manager-agent-yaml-needs-manual-bootstrap.md) — `manager.agent.yaml` edits do NOT auto-deploy; a human must run `npm run bootstrap` from `ai-manager/` with `ANTHROPIC_API_KEY`+`AGENT_ID` to push the prompt to the live agent (ENG-28)
 - [learnings/sentry-dedupe-must-check-closed-tickets](learnings/2026-05-27-sentry-dedupe-must-check-closed-tickets.md) — step-2 Sentry dedupe must match closed tickets too, not just open — canary errors (no Sentry write tools / ENG-21) get re-filed forever otherwise (APP-1 → ENG-17 closed → ENG-29 dup)
+- [learnings/fetch-before-merge-on-long-lived-branch](learnings/2026-05-27-fetch-before-merge-on-long-lived-branch.md) — On long-lived branches, run `git fetch origin <base>` before `git merge origin/<base>`; local origin ref goes stale during review-feedback rounds (PR #12 round 3 escalation)
 
 ## Decisions
 - [decisions/mcp-for-small-writes-checkout-for-big](decisions/2026-05-26-mcp-for-small-writes-checkout-for-big.md) — Single-file writes go through GitHub MCP; multi-file or test-needing changes use the mounted checkout + `git push`

.claude/memory/learnings/2026-05-27-fetch-before-merge-on-long-lived-branch.md

@@ -0,0 +1,27 @@
+# `git fetch origin <base>` before `git merge origin/<base>` on long-lived branches
+
+When a review-feedback round runs hours after a branch was opened — long enough for other PRs to merge to the base — the session's local `origin/<base>` ref is stale. `git merge origin/<base>` then resolves to the branch's original `base.sha`, not the current tip on the remote. The merge "succeeds" but brings in old state, dropping anything that landed in between.
+
+PR #12 (memory: pw-capture-script) hit this on round 3 (2026-05-26). The branch was opened at 09:19Z against `af2431b`. By round 3 (10:32Z) PR #13 had merged to main (`1072c9b`), adding `learnings/2026-05-26-vercel-bot-status-as-deploy-health-fallback.md` plus a `MEMORY.md` index line in the slot this branch reused. The round-3 session ran `git merge origin/main` without `git fetch` first, so it merged stale `af2431b` and the resulting branch diff showed the vercel-bot-status learning file as deleted. Mergeable state went `dirty`. Reviewer caught it on round 3, posted `AGENT_REVIEW: ESCALATE` (3-round limit), and the PR is still open the next morning waiting for a human to re-merge.
+
+## Rule
+
+Before any `git merge origin/<base>` on a branch that's been open for more than a few minutes:
+
+```sh
+git fetch origin <base>
+git merge origin/<base>
+```
+
+Or chain it: `git fetch origin main && git merge origin/main`.
+
+Applies to:
+- Review-feedback rounds (manager wakes hours after PR open).
+- Retro PRs that resolve conflicts against main.
+- Any time the session was kicked off well after the branch HEAD was last touched.
+
+Does not apply when the session JUST cloned/checked-out from the mount — the initial fetch covers it. The risk is specifically the second-and-later `merge origin/<base>` within the same long-running branch lifecycle.
+
+## Why this slipped through
+
+The escalation comment on PR #12 has the full trace. Existing `review-feedback-fanout` learning addresses the *parallel-session* race; this is the *time-passed* race on a single session. Same root cause shape ("the remote moved while you weren't looking"), different trigger.