chore(memory): record #394 lessons and compact the orchestrator memory index

[drmoisan]

chore(memory): record #394 lessons and compact the orchestrator memory index

Summary

• Adds one new orchestrator memory: never commit a helper script under a feature's evidence/ tree.
• Updates the pr_context.summary.txt memory with a newly observed stale-content failure mode.
• Compacts MEMORY.md from 19.6KB to 13.9KB. All 77 entries preserved; none dropped or merged.
• Touches only .claude/agent-memory/orchestrator/**. No production code, no policy files, no rules.

Why

Follow-up to #394 (merged as PR #533). Two findings from that run are worth persisting, and the SessionStart hook flagged the orchestrator memory index as approaching its 24.4KB read limit.

Finding 1 — helper scripts under evidence/. feature-review's Get-ChangedLanguageSet matches on file extension alone and is completely path-blind. A 27-line duplicate-sweep.ps1 that the executor retained under <FEATURE>/evidence/baseline/ "for reproducibility" therefore put PowerShell into the branch's changed-language set, which triggered the mandatory coverage-verification rule with no Pester test and no coverage artifact behind it. That cost a full remediation cycle (planner, executor, reaudit) on a branch whose only real change was deleting one line from a .csproj. The plan had only ever required capturing the script's output.

Finding 2 — stale PR context. collect_pr_context returned ok:true and listed the worktree artifact paths, but did not overwrite an existing hand-authored pr_context.summary.txt. The stale file still described the .ps1 that remediation had already deleted, so authoring a PR body from it would have described the wrong diff.

What Changed

• New feedback_no_helper_scripts_under_evidence.md — the rule, why it fires, and the proportionate remediation (git rm the script; the sibling .md artifact already carries its logic and output verbatim).
• Updated pr-context-summary-unreliable-gh-and-classification.md — adds the stale-content variant alongside the existing "gh unavailable" and misclassification variants.
• Compacted MEMORY.md — one line per entry with shortened hooks. Detail already lives in the topic files the index points at.

Verification

• Entry count before and after: 77. Verified with grep -c "^- \[".
• Size: 19.6KB → 13.9KB, under the 17.1KB target the hook asked for.
• Every index link target still exists as a file in .claude/agent-memory/orchestrator/.

Not verified in this PR: no build or test gate applies. The diff contains zero source files.

Backward Compatibility / Migration Notes

None. Memory files are agent-facing documentation.

Risks and Mitigations

The compaction rewrites every line of MEMORY.md, so a concurrently in-flight epic sibling that appends its own index line will conflict on this file rather than merging cleanly. This is the known .claude/agent-memory/** fan-in conflict, and the resolution is unchanged: take both sides' entries. To keep that window as small as possible this PR is deliberately tiny and is being merged immediately after creation. A resolver should keep the compacted form and re-append the sibling's line rather than restoring the long-hook version.

Review Guide

Read feedback_no_helper_scripts_under_evidence.md first; the MEMORY.md diff is mechanical hook-shortening.

Follow-ups

None.

GitHub Auto-close

None.