fix(recorders): floor func_internal_density's denominator at 12 lines per function (#2705)

[squid-protocol]

Closes #2705

What

func_internal_density = avg_func_complexity / max(avg_func_loc, FUNC_EVIDENCE_MASS_FLOOR) with FUNC_EVIDENCE_MASS_FLOOR = 12 beside EVIDENCE_MASS_FLOOR = 50 in analysis_lens.ENGINE_CONSTANTS. One constant, one recorder line, one test module, one wiki paragraph.

Why

#2655 floored every per-file density; the recorder-side per-function density was left without one, so it read function length where structure was held equal. keyword-rosetta (the same 13 functions at 46 lengths) measured it at Spearman −0.96 against coding_loc with branch/args/func_start held. In the golden-master scan, five find*.sql files whose functions average one line and one branch each read a flat 1.00 — the densest files in the corpus — for being the tersest.

The floor follows #2655's own derivation: 50 sits at the golden-master median coding_loc (51); 12 sits at the golden-master median avg_func_loc (12.0) and floors the same 49% of the population. Below it the column is an exact rescale of avg_func_complexity (branch structure); at or above it, byte-identical to before. Design pass and decisions on #2705 (options 2 / 4 / separate bug).

Effect: rosetta leak rho −0.96 → 0.000, consistency 54% → 67%, no --gate verdict changes (the cells were already derived). Real code (scan DB): 574 / 2074 files with functions move, median drop 0.060, 72.3% untouched.

Blast radius

• Not in either pre-trained vector (neither the 5-dim per-function archetype vector nor the file-level raw_vector), so no Remove AST-dependent Security & Vulnerability metrics (Injection, RCE, Memory) #1020-style archetype drift.
• Its only other consumer has always read 0.0: security_auditor.py:385 takes the key from telemetry_payload, which signal_processor never writes. Filed as security_auditor reads func_internal_density from telemetry, but signal_processor never writes it — a permanently-zero feature in the security frame #2714, deliberately not fixed here — populating a security-frame input is a different blast-radius question and must not share a PR with this.
• No golden-master bless. crucible_check.py passes in both venvs: the audit fixtures carry Raw Cognitive Density but not this recorder column, so the 574-row figure in the design pass (measured on file_data) never reaches tests/golden_master_*.json.

Verification

• tests/tools_recorders/test_func_internal_density_floor.py pins the four 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 properties for the per-function analog through RecordKeeper.record_mission: invariance below the floor, no cliff at it, parity with avg_comp / avg_loc above it, zero with no functions; plus the motivating one-line-one-branch shape and the 12/50 derivation pin.
• audit_check.py all clear (ruff baseline regenerated for pure line-shifts; mypy clean via cast).
• crucible_check.py: full_precision PASS, zero_dependency PASS.

Cross-repo

keyword-rosetta: no manifest changes (the column is a derived measure, not a planted signal); bias-history.yml regenerates the report against engine main after merge and the func_internal_density row should leave the length-leak table. control_flow_ratio stays per #2705 decision 2 (ledger already exists; leak-check label rides with F.3). Label rosetta:rebless-owed not needed.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab