curate .claude/agent-memory/code-reviewer-lean/MEMORY.md under the 24KB agent-memory index FAIL cap (41KB on main; reds worktree spec-freshness sync commits + silently truncates the lean twin's loaded lessons)

The defect

workflow_lint.py's agent-memory index size check (#1891) FAILs any MEMORY.md over 24,000 bytes (the loader truncates the always-loaded index at ~25,000 bytes, silently dropping the newest lessons). The copy of .claude/agent-memory/code-reviewer-lean/MEMORY.md currently ON MAIN measures 41,371 bytes — 72% over the FAIL cap — so:

1. The code-reviewer-lean agent's always-loaded index is silently truncated at ~25 KB: roughly the newest 16 KB of its lessons never load. The lean twin exists precisely for context-constrained respawns, so an oversized index is self-defeating.
2. Every worktree spec-freshness sync (the /issue Step 5a family-atomic block and its Step 9c pre-gate re-run) that pulls the file from origin/main FAILs its sync commit on the workflow-lint-agent-memory-index-size pre-commit hook. Observed on #2610's Step 9c pre-gate re-sync (2026-08-27): the sync commit died rc=1 on exactly this file and the orchestrator had to exclude it by hand (revert to branch-era copy) to land the other ~20 synced files.

Two more indexes are in the WARN band (>20,000 bytes) and will cross the FAIL cap on their current growth path: .claude/agent-memory/codex-code-reviewer/MEMORY.md (21,545 B) and .claude/agent-memory/experiment-implementer/MEMORY.md (21,235 B).

The fix (per the #1891 curation recipe the lint message names)

Curate .claude/agent-memory/code-reviewer-lean/MEMORY.md to under 20,000 bytes: trim each index hook to ~1 line (≤~150 chars), move the detail into the pointed-to per-entry file (creating per-entry files where a row has none), and merge duplicate/sibling rows. No lesson content is deleted — detail moves to per-entry files; only the always-loaded index shrinks. Opportunistically apply the same pass to the two WARN-band indexes above if cheap.

Also worth checking (secondary): how a 41 KB index landed on main past the pre-commit hook — likely appended by sessions whose commits bypassed the hook path or accreted before the FAIL threshold existed. If a durable gap is found (e.g. a commit path that skips the lint), name it in the clean-up commit message; no separate enforcement change is in scope here.

Acceptance

• .claude/agent-memory/code-reviewer-lean/MEMORY.md < 20,000 bytes on main, with the displaced detail present in per-entry files linked from the index.
• The no-flags workflow_lint.py run PASSes with zero agent-memory-index FAIL rows.
• A worktree spec-freshness sync including the file commits cleanly (reproduce: stage the file in any worktree and run the pre-commit hook).

Provenance

Observed while driving task #2610 through /issue (Step 9c pre-gate spec-freshness re-sync, 2026-08-27). Grounded by the failing hook output (rc=1, workflow-lint-agent-memory-index-size) and byte counts from that run.