[squid-protocol]

Closes #4106.

Problem

Five per-file archetype fields were placeholders reported as measurements: constant N/A / 0.0 / {} on all 3,003 crucible code files. #3061 moved file classification into record_keeper, which only overwrites the label. Nothing has computed the distance or fingerprint since, and nothing applies the per-language model at all.

Fix

Computed: the general file archetype (k=18).

• record_keeper._classify_file_archetype_detail returns the nearest archetype, its Euclidean distance to that centroid, and the distance to every centroid. The nearest-centroid loop already computed all of them and threw them away. They're now written to global_drift / archetype_fingerprint, which revives:
  • the GPU recorder's secondary-archetype a_ids / a_dists (the site reads only a_ids[i][0], so a secondary id is safe);
  • the LLM brief's "Top Global Matches" and its mixed-responsibility section.
• Audit key Repository Drift (Z-Score) → Repository Drift (Centroid Distance). It's a distance, never a z-score. Nothing outside the recorder read the old key.
• Mixed-responsibility is recalibrated. The old absolute delta <= 0.9 was tuned for a feature space retired in fix(archetypes): classify file archetype from a self-describing brain, post-assembly #3061: on today's fingerprints (median distance 6.1) it would flag 33% of crucible code files. It's now relative, ENGINE_CONSTANTS["MIXED_RESPONSIBILITY_GAP"] = 0.02 (second-nearest within 2% of the nearest), which flags ~5% (149 / 3,003). Both the LLM brief and the GPU recorder use the same constant.

Retired: the per-language "local micro-species" model.

• SPECIFIC_FILE_INFERENCE_MODEL (~3,600 lines) and LANGUAGE_INFERENCE_MODELS are removed. Nothing applied them since fix(archetypes): classify file archetype from a self-describing brain, post-assembly #3061, and they can't be applied: there's no FEATURE_NAMES, and 74 dims match no feature space the engine computes (83 at fix(archetypes): classify file archetype from a self-describing brain, post-assembly #3061, 84 today). Reviving them means retraining as a self-describing brain.
• local_archetype / local_drift / local_fingerprint are removed from:
  • telemetry;
  • the audit (File Archetype, File Drift (Z-Score), File Fingerprint);
  • the LLM brief, including the "Severe Anti-Patterns" drift-ratio check, which could never fire;
  • the LLM graph DB columns.
• docs/wiki/08-19 no longer documents the drift-ratio multiplier. The page also flags that _calc_obscured_payload itself is no longer in signal_processor.py and needs re-verification.

Golden master

Re-blessed against crucible v1.6.0. Only 3_architectural_profile parts move in each fixture:

• the four retired keys are deleted;
• repository_fingerprint is now populated;
• repository_drift_centroid_distance is added.

No label, signal, risk or summary value changed.

Tests

• test_signal_processor: the detail call returns the correct distance and full fingerprint, and the nearest label matches the minimum distance.
• test_audit_recorder: the distance is reported under its new key, and the retired keys are absent.
• core_engine + tools_recorders: 2,407 passed; audit_check all clear. Every changed file compiles on Python 3.9.

Merge order

This touches llm_recorder.py / audit_recorder.py / record_keeper.py, as does #4113. Whichever merges second needs a merge from main (import-line and adjacent-hunk conflicts only).

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab