Agent-memory index FAIL gate can be tripped by the MERGE itself, blocking a Step-10d landing

Goal

Close the gap where a Step-10d landing merge is blocked by an agent-memory index size FAIL that NEITHER merge parent violates. Either make the gate merge-aware, or give the landing session a prescribed curate-in-the-merge remedy that the .claude/rules surface names, so a session does not have to derive it live while holding a 786-file merge.

What happened (#2569, 2026-08-27)

The /issue 2569 Step-10d landing merged issue-2569 into a scratch worktree detached at origin/main. The merge commit was refused by the pre-commit hook workflow-lint-agent-memory-index-size:

.claude/agent-memory/code-reviewer-lean/MEMORY.md: 35820 bytes exceeds the
24000-byte agent-memory index FAIL threshold

Measured sizes:

ref	bytes
merge base ebfc5c0727	19,260
origin/main side	17,522
issue-2569 side	18,332
merged working tree	35,820

Neither parent is anywhere near the 24,000 FAIL threshold — both are under even the 20,000 WARN. Both sides had added ~110 DISTINCT index rows since the merge base (main via other sessions' reviewers; the branch via this task's own Step-5 review ensemble, four commits), and git's textual merge correctly unioned them. The union is what violates.

Why this is a workflow-surface gap, not a task bug

1. The failure is structural and recurring. Every agent-memory index is an append-mostly list of one-line rows written by many concurrent sessions with no natural committer. Two branches that both ran reviewers will BOTH have appended rows. So any long-lived issue branch landing onto a busy main is exposed, and the exposure grows with branch age and fleet size. #2015 already identifies this file class as the fleet's dominant standing armer for a different reason (the pre-commit stash race); this is a second failure mode on the same class.
2. The gate is index-content-only and merge-blind. It measures the resulting file. It cannot distinguish "an agent let its index grow unbounded" (the behaviour #1891 set out to stop) from "two disciplined indexes were unioned by a merge". The first deserves a FAIL; the second is a mechanical consequence of landing.
3. The remedy is not written down anywhere reachable from the landing step. .claude/skills/issue/steps/18-step-10d.md says nothing about it, and the hook's own message prescribes curation ("trim each index hook to ~1 line") without saying who owns that during a merge, or that the merge is the trigger. The #2569 session had to derive the diagnosis and the fix live.
4. The failure mode of NOT having a rule is bad. The tempting escapes are --no-verify (forbidden) or dropping one side's rows (silent lesson loss). Both are worse than the gate.

What #2569 did (the candidate remedy, already validated once)

Curated the merged index in the merge commit, preserving every row:

• verified all 218 pointed-to per-entry .md files exist FIRST, so the full lesson text is safe and only the index label is being shortened;
• re-trimmed from the MERGE-STATE text (git show :<path>), never from an already-trimmed copy — re-trimming trimmed text double-truncates hooks;
• trimmed each row to a <=95-char pointer label at a word boundary, byte-measured (the indexes contain multi-byte characters — em dashes, ellipses, ≠, ⇒ — so a character count under-reads the gate by ~2%);
• result 35,820 -> 18,700 bytes, clear of both FAIL (24,000) and WARN (20,000), 218 of 218 rows preserved;
• recorded the whole thing in the merge commit message.

Acceptance criteria

1. A landing session hitting this failure finds the diagnosis and the remedy in the rules surface, without deriving them. Concretely: .claude/rules/ names the merge-union trigger, and the Step-10d step file points at it.
2. The prescribed remedy is byte-measured, preserves every row, verifies the per-entry targets exist before trimming, and trims from the merge state.
3. --no-verify and one-side row-dropping are both explicitly named as forbidden escapes.
4. Decide (and record the decision either way) whether the gate itself should become merge-aware — e.g. skip or downgrade to WARN when both merge parents are under the threshold and the path is an agent-memory index, with the landing commit still owing the curation. A deliberate "no, keep it strict and document the remedy" is an acceptable outcome; leaving it undecided is not.
5. Whatever ships is pinned by a test.

Provenance

Surfaced by the /issue 2569 Step-10d landing (2026-08-27). Merge commit bc1d4a8f19, pushed as 0260a47a98. Sibling rules already covering this file class for other reasons: .claude/rules/repo-root-uncommitted-state.md (#2015), #1891 (the index size ratchet itself).