Break tri-comparison precision ties by validated count; codify as a system-wide rule

[squid-protocol]

Summary

Found while investigating why rust's tri-comparison chart shows zero badges despite having the highest, ledger-validated accuracy: _winner_or_tie() in tests/tools/tri_comparison_chart.py had no tie-break at all -- any 2+-way tie on rate_pct returned "tie", blank badge, full stop.

Rust's real case: Func/Class Precision now ties 3-way at 100% (GitGalaxy 1927/1927, tree-sitter 1775/1775, ctags 1774/1774) once GitGalaxy's macro-body-only claims were ledger-validated (#1941). Each tool is simply never wrong about what it itself claims, at very different claim counts -- so a rate-only comparison awards nobody, despite GitGalaxy having demonstrably found more of the validated-real total.

Fix

Break a rate tie using each tied tool's absolute matched_consensus (its count of validated-correct occurrences), not the raw rate alone. This is the same "more evidence should count for something" principle as the existing sample-size fix (a 2-sample 100% cell no longer silently outranks an 80-sample 98.75% cell) -- just pointed the other direction: a 1927-sample validated 100% shouldn't lose to a 1774-sample validated 100% just because both cleared the same bar.

Only ever reached after the caller's has_open_question() check (already gated before _winner_or_tie is called), so every count used in the tie-break is already either unquestioned or ledger-validated -- this is not a reversion to the "just trust the bigger number" anti-pattern the verification system exists to prevent.

Verified

• Direct unit check against rust's real numbers: gitgalaxy now wins the tie.
• Guard case 1: a genuine unbreakable tie (equal rate AND equal count) still returns "tie".
• Guard case 2: the original sample-size-bug scenario (2/2 @ 100% vs. 79/80 @ 98.75%) is unaffected -- it was never a tie to begin with, so this path isn't touched.
• Full 45-language chart regenerated (tri_comparison_chart.py --all --write, using tree-sitter-language-pack + a locally-built universal-ctags): diff is exactly rust's two new badges (Func Precision, Class Precision) plus the summary tally moving 4→6 GitGalaxy wins / 18→16 ties. No other language's badges changed. tri_comparison_ledger.json's diff is routine last_reconciled_at timestamp churn from the same regen run, no verdict/status content changed.

Also: codifies this as a system-wide CLAUDE.md rule

The "badges require verification, not just the higher number" principle (originally proven necessary on 2026-08-19, landed via PR #1868) was only ever documented inside tri_comparison_chart.py's own docstring and the tri-comparison-ledger-sweep skill -- never promoted to a repo-wide instruction. Added a new CLAUDE.md section generalizing it to any comparative-correctness claim in this repo (a badge, a README claim, a PR description), plus this PR's new tie-break rule, so future sessions apply the same standard without re-deriving it from a specific tool's source.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab