feat(core-engine): language-strictness table + measured fidelity coefficients replace the three-bucket tiers (#2718)

[squid-protocol]

Child of #2716 (design pass approved 2026-09-04). Same equations; the two constants they read now come from a documented table and a measured table instead of a three-bucket lookup on the language's name.

What changed

analysis_lens.LANGUAGE_STRICTNESS replaces _get_tier + TIER_VARS (Irc / Ot). One row per language, four yes/no columns checkable against the language's documentation — static types · enforced errors · memory safe · no implicit globals. Irc = number of gaps (0–4), Ot = 1 + 0.1·Irc. None rows (yaml, json, css, html, …) have no runtime and carry no language-level term; an unknown language also gets none — the opposite of the old fall-through to the harshest bucket. LANGUAGE_FAMILY resolves dialects first, so embedded_python reads python's row (#2653). The dead FIDELITY_TIERS (one occurrence repo-wide, disagreeing with the live sets for ~20 languages) is deleted.

gitgalaxy/standards/fidelity_table.py replaces the single per-language Fc. Generated by tests/tools/fidelity_table.py from keyword-rosetta: fc(lang, signal) = min(1, planted / measured), where planted is the SPEC's plant (the corpus-wide median of the manifests) and measured is what the rule found. A safety rule that fires 3 times on 2 planted constructs now credits each hit at 2/3; under-firing stays at 1.0 on purpose (a rule to fix, never compensated). The engine reads it per signal — safety, test, doc, ownership — at each defence site. One deliberate consequence in _calc_documentation: only the rule hits (doc, ownership) carry the coefficient; doc_loc and the GuideStar umbrella are structural evidence, not rule hits, and are no longer scaled (they used to sit inside the × Fc bracket). --check fails when the committed module is stale against the sibling corpus (test_committed_fidelity_table_is_fresh_against_the_corpus, skipped without the corpus); --docs renders both tables into 08-03.

_calc_safety's systems_buffer_ratio is retired (closes #2717). Keyed on fc < 1.0 it made every non-tier-1 file score safer than tier 1 above 6/15 attack-weighted hits — 592 crucible files. The over-firing it compensated for (C's pointer syntax) is now scaled at the source by the fidelity coefficient.

Docs. 08-03's tier paragraph is now the chain — what a regex can't see → the three inputs and where each is sourced → the invariants → the pilot that will validate the strictness weight — with the strictness and fidelity tables rendered from the data files between <!-- generated:… --> markers, never typed. 03-02's hand-written "Tier" column is labelled as the retired 2025 buckets and points at the live table; 08-08 / 08-09 / 08-15 / 08-16 say where Irc and Fc come from.

What moved, and why it's the expected shape

Golden master regenerated in both venvs and attributed with golden_diff.deep_compare (uncapped): 8,803 substantive diffs, 0 topological — risk columns only, no structural or coordinate movement.

• Zero diffs for rust, java, yaml, json, batch, nix, proto, plaintext, td, mlir, hlo, tabular — every language whose constants did not change. swift: one file, +0.01 (float ordering in the per-signal defence sum).
• Former tier-3 languages move down, because the 40% defence cut and the flat irc 5 are gone: cobol −2.9 mean over 1,700 cells, shell −2.5, lua −3.0, jcl −4.9, html −9.6, zig −7.9, embedded_python −3.3 (its bldc_driver.py Documentation −45.7, vibration_driver.py −41.6 — the risk formulas: audit language-tier membership (embedded_python lands tier3) + decide how rosetta bias reporting treats documented Fc/Irc tier constants #2653 case).
• Former tier-1/2 languages move slightly up: typescript +2.1 (measured safety fidelity 0.67), javascript +1.4 (Irc 2→3: sloppy-mode globals), python +0.9 and c/cpp +1.0 (_calc_safety: systems_buffer_ratio keyed on fc<1.0 makes tier-2/3 files score SAFER than tier 1 above 6/15 attack hits #2717: the 0.75 attack discount is gone; python test files with heavy safety_bypasses are the +31 Error & Exception movers).
• The largest per-file swings are the same files the flat-tier experiment on risk formulas: replace language tiers with a measured fidelity table + a language-strictness table (child of #2560) #2716 flagged, now moving the other way: zig constants.zig Documentation −81.5; cobol/zig/lua Testing −77.6 (Ot 1.40 → 1.20 / 1.10 / 1.30 at a sigmoid knee — pre-existing sensitivity, symmetric with the +77.5 the old constants produced).

Gauntlet

• audit_check.py: all clear (ruff baseline regenerated for 5 pure line-shifts; mypy 2-error baseline; dead-key; ast-accuracy).
• tree_sitter_accuracy_audit.py --ci --all: 30/30 OK. tri_comparison_chart.py --all --ci: 3/3 OK.
• rosetta_audit.py: 46/46 PASS, 0 regressions — the corpus gates §2 raw signals, which this PR does not touch, so no corpus re-bless is owed; the risk_* rows of the bias report move on the next bias-history regen, and keyword-rosetta#50's within-tier banding can then be retired in favour of printing the strictness offset (risk formulas: replace language tiers with a measured fidelity table + a language-strictness table (child of #2560) #2716 checklist).
• Full suite: 7,562 passed, 3 skipped, 9 xfailed (the 4 pre-UEF small-file normalization isn't achieving its intent — redesign scoring so identical intent scores identically regardless of file size (design pass for Fable) #2655 pins at fc<1.0 for safety and documentation are exempted, see test comment).
• Golden master: blessed in both venvs (crucible_check.py --update --yes), scoped diff above.

Not in this PR

• The flat irc is still summed into six formulas in six units — engine PR 2: per-file dynamism count replaces the flat irc term (child of #2716) #2719 swaps it for the per-file dynamism count, one formula at a time.
• The strictness weight (1 Irc per gap, +0.1 Ot) is provisional; tooling: outcome-label pilot — do strictness / dynamism predict later fix commits? (child of #2716) #2720 is the outcome pilot that decides whether it deserves more or less.

Closes #2718
Closes #2717
Closes #2653

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab