chore: track fixes from the 2026-10-04 Windows performance audit #6244

Problem

A performance audit of one Windows 11 workstation (SETWORKS-KSDEV1, Claude Code 2.1.289, Git for
Windows 2.55) on 2026-10-04 traced the user's typing lag and part of the steady per-call overhead to
plugins in this repository. This issue tracks the nine plugin fixes it produced, highest value
first.

Headline measurement: four dead-parent rm/du processes left by the testing plugin's
test-judge.test.sh cleanup had run since 03:08 and burned 26.1 CPU-hours. Stopping them dropped
machine CPU from 58% to a median of 31% and ended the lag (one trial).

Evidence

Evidence labels used in every child issue:

• E1: measured on this host and attributed to a cause.
• E2: aggregate measurement here; the cause is inferred.
• E3: structural inference (from code, config or docs), not measured.
• E4: suspicion or judgment.

Measured process start cost on this host (medium confidence): PowerShell 7 541 ms (887 ms with a
profile), bash 53 ms, node 95 ms, python 85 ms. The audit's reading of the Claude Code docs: all
matching hooks run in parallel, plugin copies of a hook are not deduplicated, an if condition that
does not match starts no process, and mod (in-process) hooks need no process from 2.1.287 but fail
open without a .catch.

The audit's local evidence files are not durable; each child carries the facts it needs.

Proposed approach

1. testing: gate test-judge.test.sh link fixtures on a real symlink so cleanup cannot spin (E1).
   fix(testing): gate test-judge.test.sh link fixtures on a real symlink so cleanup cannot spin #6248
2. harness-ops: rank dead-parent processes by CPU time in audit-performance; its census missed
   all four (E1). feat(harness-ops): rank dead-parent processes by CPU time in audit-performance #6249
3. testing: test-judge Stop hook blocked the turn on "0 FLAG, N PASS" (E1). Already fixed on main
   by fix(testing): do not block Stop when the judge has nothing to decide #6092 (testing 0.22.5, closing fix(testing): skip temp-copy tests and stop blocking Stop on judge runs with nothing to decide #6037). Blocks still seen after it are 0 FLAG with an UNKNOWN,
   which blocks by design; the UNKNOWN cause seen is fix(testing): test judge marks a linked-worktree test UNKNOWN because its quotes are grounded against the wrong repository #6206. Message noise is fix(testing): test-judge Stop hook floods task end with every deferred test name and relays all-pass verdicts #6226.
4. animation, explainer-video, speech: pass pydeps.py to native Python as a Windows path. Low:
   did not reproduce outside the audit's own MSYS_NO_PATHCONV=1 claude -p launches.
   fix(animation, explainer-video, speech): pass pydeps.py to native Python as a Windows path #6250
5. Edit/Write hook chain: Edit wait p50 5 to 7 s against Read under 1 s (E2).
   perf(guardrails): skip the verify rows before bash when all verifiers are off #6252
6. guardrails Bash PreToolUse bundle: p50 1.3 to 1.6 s, two interpreters per call; security class,
   consolidation is Kyle's call (E2). perf(guardrails): decide the route to one guard process per shell call #6245
7. SessionStart burst: about 75 to 80 processes on every start, resume, /clear and compaction
   (E3 count, E2 time). perf(hooks): skip node-notice and probe SessionStart rows on compaction #6251
8. Hooks that spawn and then do nothing (E3).
   perf(testing): skip test-scan-bash before bash when no bashEditDiff #6253, perf(harness-ops): decide whether the session event log moves into a mod #6246, perf(autonomy, disk-hygiene): decide how Stop rows skip bash when idle #6254
9. Stale plugin paths in long-running sessions (E1).
   fix(harness-ops): name stale hook config and suggest /reload-plugins #6247

Acceptance criteria

• Every child issue above is closed, or closed as not planned with a reason.
• No child removes a guard or a security test case; each one consolidates or gates instead.

Constraints and gotchas

• Kyle's standing rule: guards change only when Kyle says so. Consolidate or gate, never remove.
  Security-class items (fix 6 and any guard change) are Kyle's call.
• Every number comes from one host on one morning, during heavy agent activity; re-measure before
  claiming a win.

Context

Source: local handoff item 20261004-150216-claude-perf-audit-plugin-fixes.md (retired into this
issue and its children).