Bug: same-domain git over-fire — push-specific rule still marks git status/git pull successes as broken #815

Description

simulate() still returns broken for same-domain successes that merely rhyme with a rule's trigger. The #723 domain gate compares only coarse tool groups (git -> vcs), so it cannot distinguish git push from git status/git pull, and BROKEN_MARGIN is a fixed absolute score offset. When a successful trajectory in the same domain happens to contain push-failure wording anywhere in its haystack (task or any step input/output/error), the rule is counted as having broken a success.

This is the structural gap flagged in a reader comment on the "Killed by the Word 'git'" article (dev.to comment 3f3ob, article https://dev.to/debashish_ghosal/killed-by-the-word-git-one-token-of-coincidence-40-points-of-pass-rate-140f):

"did you consider keying the domain on the trigger command itself rather than lexical similarity? Even after the margin, I'd worry a push-specific rule still over-fires on git status runs in corpora where push and pull failures both show up often."

The #723 fix (domain-gate broken + match-strength margin) is necessary but not sufficient: the gate is keyed on the trigger token (git), not the trigger command (push), so a git status success is "same domain" and the only remaining guard is the 0.10 margin.

Steps to Reproduce

1. Build the F-001 rule from field-test/corpus/curated/failures/positive/F-001-git--001-git-push-non-ff.jsonl (the near-verbatim extracted candidate).
2. Replay it against a successful git trajectory whose haystack mentions an earlier push failure (e.g. a runbook/notes step), as happens in corpora where push and pull failures co-occur.
3. Call simulate(candidate, success_trajectory).

Expected Behavior

no_effect (or near_miss): a successful git status/git pull/documentation run is not interference from a push-specific rule.

Actual Behavior

broken — a spurious broken success that demotes a correct rule (the same failure class #723 was meant to eliminate, just within-domain instead of cross-domain).

Environment

• CauterRule version: 0.3.1 (pyproject.toml version = "0.3.1")
• Python version: 3.11+ (reproduced on 3.14)
• OS: macOS
• LLM provider: None (deterministic replay)
• Model: n/a

Minimal Reproducible Example

from cauterule.models.candidate import CandidateRule
from cauterule.models.rule import RuleWhen, RuleDo
from cauterule.models.trajectory import Trajectory, Step
from cauterule.replay.matcher import match_score, extract_trigger_domain
from cauterule.replay.simulator import simulate

c = CandidateRule(
    when=RuleWhen(trigger="when git push fails with non-fast-forward"),
    do=RuleDo(directive="pull latest changes before pushing"),
    confidence=1.0,
)

s = Trajectory(
    id="S-git-status-rhymes",
    timestamp="2026-09-16T00:00:00Z",
    task="Check git status, then document the earlier push failure for the runbook",
    steps=(
        Step(1, "bash", "git status", "nothing to commit, working tree clean"),
        Step(2, "bash", "cat notes.md",
             "Earlier: git push origin main was rejected non-fast-forward; resolved by pulling."),
    ),
    success=True,
    domain="git",
    failure_class=None,
)

assert extract_trigger_domain(c.when.trigger) == "git"
assert match_score(c, s) == 0.70          # threshold 0.60 + BROKEN_MARGIN 0.10
assert simulate(c, s) == "broken"          # expected "no_effect"

Trajectory or Corpus (if applicable)

• Rule: field-test/corpus/curated/failures/positive/F-001-git--001-git-push-non-ff.jsonl
• Same-domain successes that expose the gap: field-test/corpus/curated/successes/S-023-git-status.jsonl, S-022-git-pull.jsonl, NM-044-git-commit-hook.jsonl

Output

trigger domain      : git
match_score         : 0.7
rule_matches(0.6)   : (True, 0.7)
simulate ->         : broken (expected: no_effect)

Relevant code:

• src/cauterule/replay/simulator.py:135-138 — check_domain_mismatch then score < threshold + BROKEN_MARGIN
• src/cauterule/replay/matcher.py:749-787 — extract_trigger_domain / check_domain_mismatch operate on the coarse _KNOWN_DOMAINS/_DOMAIN_GROUPS token, not the command
• src/cauterule/replay/matcher.py:438-450 — _build_haystack includes every step input/output, letting unrelated wording raise the score

Severity

• Critical — crashes, data loss, or security vulnerability
• High — incorrect results or broken core feature
• Medium — unexpected behavior with workaround
• Low — cosmetic or minor inconvenience

Acceptance Criteria

• Bug reproduced and root cause identified
• Fix implemented
• Test added that catches the regression
• Code review passed
• Lint strict clean (ruff check . + mypy src/ tests/)
• Code coverage > 95%
• All necessary documents updated

Suggested direction

Key the domain on the trigger command (e.g. derive push/pull/status sub-domains for git), or gate broken on a signature/grounding check (as is_grounded already does for prevented) so a same-domain success must share the failure signature, not just the tool token, before it counts as interference.