steps.json history-rewrite step-3 regex misses hyphenated git-filter-repo invocation

[toejough]

Problem

dev/eval/cumulative/runbook_vs_skill/phase2/fixtures/history-rewrite/steps.json step 3's detection regex is:

git\s+(filter-branch|filter-repo)\b

This only matches the git-subcommand form git filter-repo ... (space-separated). It does not match the equally-valid hyphenated standalone-script invocation git-filter-repo ..., which is how git-filter-repo is commonly installed/invoked (e.g. via pip/brew, directly on PATH as git-filter-repo).

Steps 4/5/6 are gated "after": 3, so when an agent uses the hyphenated form, steps 3-6 are ALL scored false in followed_steps/followed_all, even when the agent genuinely ran the rewrite, fetched, force-pushed, and verified -- confirmed via done_when_checks.sh passing all 6 real checks and via reading the full bash command transcript.

Evidence

Found during runbook-shim-follow-frame task 4.3's 3rd rerun (2026-09-16), results/4.3_stage1_rerun3_sonnet5_shimonly_history-rewrite.jsonl. Trials history-rewrite-R-0 and history-rewrite-R-1 both ran:

git-filter-repo --path secrets.env --invert-paths --refs main --force

-- hyphenated form, not matched by the regex -- followed by git fetch origin and git push origin main --force-with-lease, both genuinely run but scored false because they're gated behind the unmatched step 3. history-rewrite-R-2 happened to use git filter-repo and scored followed_all=True. All 3 trials passed the real done_when_checks.sh end-state check (which passed all 6 checks silently and printed the final PASS line), confirming this is a scorer/regex gap, not a genuine behavior difference between trials.

Fix

Broaden step 3's pattern to match both invocation forms, e.g.:

git[\s-]+(filter-branch|filter-repo)\b

(only filter-repo ships a standalone hyphenated script; git-filter-branch isn't a real binary, so no new false-positive class).

Scope note

Out of scope for runbook-shim-follow-frame task 4.3 (no shim/runbook amendments authorized this round) -- filed for later harness maintenance. Does not change the 4.3 D8 verdict, which is based on done_when_checks.sh's real end-state result, not followed_all.