ci: gate PR merges on tri-comparison validated-precision regression

[squid-protocol]

Summary

Makes the tri-comparison system (GitGalaxy vs. tree-sitter vs. ctags) CI-enforced for any PR
touching core-engine parsing logic, instead of skill/human-driven only. Two parts, following the
already-proven tree-sitter-accuracy-audit.yml/-history.yml split rather than inventing a new
mechanism:

• tests/tools/tri_comparison_chart.py --ci/--regenerate: baseline-gated regression check
  on GitGalaxy's own validated precision (func_precision/class_precision, read after
  apply_verified_adjustments() applies any ledger verdict — never a raw, unvalidated
  disagreement count, per this repo's "comparative-correctness claims require verification" rule).
  New per-language baseline files, tests/tri_comparison_baseline_<lang>.json, populated
  incrementally via --regenerate (none bundled in this PR — that's a deliberate follow-up once
  this scaffolding lands, not bulk-generated here). Recall is deliberately not gated — the
  chart's own module docstring already explains why that ratio's cross-tool denominator isn't
  trustworthy enough to rank on (a bug in either other tool can inflate it without GitGalaxy
  doing anything wrong); gating CI on it would reintroduce the exact problem the chart already
  designed around.
• .github/workflows/tri-comparison-audit.yml (PR-time, blocking, same core-engine paths
  tree-sitter-accuracy-audit.yml already uses) runs the --ci check. It only ever measures —
  it never requires a contributor to regenerate and commit the chart/ledger/report themselves.
• .github/workflows/tri-comparison-history.yml (push-to-main companion) does that
  regeneration automatically after a relevant change lands, opening an auto-merged PR only when
  the numbers actually moved (overwrite-in-place files, so "no diff" already is the adaptivity —
  no separate dedup check needed the way the tree-sitter history CSV needs one).
• Both workflows hard-fail if ctags --version doesn't print Universal Ctags before doing
  anything else, addressing the PR Fix #1892: COBOL func_start regex reserved word shield #2111 (2026-08-22) incident where a shadowed ctags binary
  silently degraded every language's comparison with no error anywhere in the run.
• Updates docs/self_scan/tri_comparison_README.md (new "CI enforcement" section), CLAUDE.md's
  Differential Scan section, and the ci-push-checklist/tree-sitter-accuracy-sweep skills to
  reflect that the chart/ledger regen is no longer a manual pre-push step, and that a sweep fix
  touching a baselined language now needs tri_comparison_chart.py --all --ci clean too.

Verified end-to-end locally (real ctags, real language-crucible corpus, real
tree-sitter-language-pack): generated a real cpp baseline (func_precision: 99.71%,
class_precision: 100% — matches the already-documented cpp numbers), confirmed --ci passes
clean against it, then simulated a regression by editing the baseline and confirmed --ci fails
with the correct diff message. That test-generated baseline file was removed before committing,
per the "start empty, populate incrementally" decision above.

Test plan

• tests/tools/tri_comparison_chart.py --languages cpp --regenerate produces correct real
  numbers against the pinned language-crucible corpus.
• --ci passes clean against a matching baseline.
• --ci fails with a correct, readable diff against a deliberately-inflated baseline, and
  recovers cleanly once reverted.
• python tests/tools/audit_check.py (ruff/mypy/dead-key/ast-accuracy) — clean except 2
  pre-existing mypy findings in gitgalaxy/core/network_risk_sensor.py, unrelated to this
  change (not touched by this PR).
• Both new workflow YAML files parse cleanly (yaml.safe_load).
• CI itself (this PR doesn't touch any baselined language's tri_comparison_baseline_*.json,
  so tri-comparison-audit.yml should report "0 languages checked" / pass trivially — first
  real population of a baseline is the deliberate follow-up mentioned above).

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab