Repository navigation
Dead per-file archetype fields: File Archetype / drift z / fingerprints are never computed but reported as measured #4106
Description
Problem
Five per-file archetype fields in the audit are placeholders that nothing computes any more, but they're reported as measured values. In the committed golden master they are constant across all 3,003 crucible code files:
| Audit field | Telemetry key | Value everywhere |
|---|---|---|
File Archetype |
local_archetype |
"N/A" |
File Drift (Z-Score) |
local_drift |
0.0 |
File Fingerprint |
local_fingerprint |
{} |
Repository Drift (Z-Score) |
global_drift |
0.0 |
Repository Fingerprint |
archetype_fingerprint |
{} |
Why: #3061 moved file classification out of signal_processor.calculate_risk_vector into record_keeper (post-assembly). signal_processor still emits these as placeholders, with a comment saying record_keeper overwrites them. record_keeper only overwrites the label (tel["archetype"], shown under the legacy key "Repository Archetype"). It never computes the distance or fingerprint, and nothing applies SPECIFIC_FILE_INFERENCE_MODEL (the per-language models for python/js/ts/c/cpp) at all.
Also: the general file archetype is reported under the key "Repository Archetype", but it's a file-level label (k=18, file_cluster_N).
Why it matters
A 0.0 drift z-score reads as "textbook fit", and the LLM brief and DB carry the same placeholders (llm_recorder global_drift/local_drift/local_archetype; DB columns of the same names). Golden can't catch a regression in a field that can't move.
Options
- Compute them: distance + fingerprint in
record_keeper._classify_file_archetype, and apply the per-language model (or retire it). - Retire them: drop the dead fields and
SPECIFIC_FILE_INFERENCE_MODEL, or report them asnull/"not computed".
Either way, rename or alias "Repository Archetype" honestly (schema change, so it needs a golden re-bless and DB migration notes).
Activity
fix(#4106): compute the file archetype's distance + fingerprint; reti…