File-level K-Means archetype classification silently truncates against its own trained models (115-dim global vs. 83-dim live vector) #1158

🛑 Problem Statement

file_data.file_archetype (telemetry["archetype"]) is assigned by _classify_archetype()
(signal_processor.py:139-166) against two separate pre-trained K-Means models, and both are out
of sync with the live feature vector actually fed into them.

The live vector (signal_processor.py:609-633, raw_vector) is built by iterating
SIGNAL_SCHEMA (skipping indent_tabs/indent_spaces/hardware_bridge/cryptography/sec_*
keys) and appending 7 engineered features (cfr, log_logic_loc, log_imports_out,
log_popularity_in, log_max_func_comp, log_avg_func_args, log_churn). Against the current
SIGNAL_SCHEMA, that's:

SIGNAL_SCHEMA length: 93
filtered length: 76
raw_vector length: 76 + 7 = 83

Global model -- GENERAL_FILE_INFERENCE_MODEL (analysis_lens.py:2155, SCALER_MEDIANS /
SCALER_IQRS / 18-cluster ARCHETYPES_K18 centroids) is 115-dimensional:

SCALER_MEDIANS (file) length: 115
GLOBAL_ARCHETYPES cluster count: 18
file_cluster_0 centroid len: 115

_classify_archetype's distance loop (signal_processor.py:156) uses
range(min(len(scaled_vector), len(centroid_vector))), so it silently compares only the first 83
of each centroid's 115 dimensions -- 32 trained dimensions (~28%) are dropped on every file.

Per-language ("local") models -- SPECIFIC_FILE_INFERENCE_MODEL (analysis_lens.py:4501) --
are 74-dimensional and only exist for 5 of the 59 languages this engine parses (python,
javascript, typescript, c, cpp; confirmed via LANGUAGE_DEFINITIONS). Every other language's
files never get a local_archetype at all (signal_processor.py:654's lang_brain lookup
misses), and even the 5 covered languages compare 74-of-83 live dimensions against their own
centroids -- same truncation bug, smaller magnitude.

This also breaks a shipped, published claim. docs/wiki/03-05-claim-5-file-archetypes.md
("Claim 5") states the file archetype model was trained on 74 dimensions across 10
clusters. The shipped model is now 115 dimensions and 18 clusters (file_cluster_0..17) -- the
model has clearly been retrained since that doc was written, but neither the live raw_vector
construction, the doc, nor (per the sibling repo-level issue) the downstream repo-clustering
feature vocabulary were updated to match.

🎯 Objective

Bring the live per-file feature vector and both trained models (global + per-language) back into
dimensional agreement, and make future drift fail loudly instead of silently truncating.

🛠️ Proposed Implementation

1. Reconstruct what the 115-dim global / 74-dim local training vectors actually contained (likely
   a superset of today's filtered SIGNAL_SCHEMA -- e.g. sec_* keys currently excluded from
   raw_vector construction may have been included at training time, or SIGNAL_SCHEMA has
   grown new keys since). Either update raw_vector construction to match what the models
   expect, or retrain the models against today's 83-dim vector.
2. Same loud-failure guard as the function-level issue: assert/log when
   len(raw_vector) != len(SCALER_MEDIANS) instead of relying on min()-truncated comparisons.
3. Decide + document the per-language model coverage scope: is 5-of-59 languages intentional
   (highest-volume languages only), or a stalled rollout? If intentional, no code change needed
   beyond documenting it; if not, track expanding it as a separate follow-up (not in this issue's
   scope -- retraining models needs a real corpus and offline tooling that doesn't currently live
   in this repo).
4. Update docs/wiki/03-05-claim-5-file-archetypes.md to reflect the model actually shipped
   today (dimension count, cluster count, and the cluster descriptions/DNA -- all presumably
   stale if the model was retrained since that doc's numbers were captured).

✅ Acceptance Criteria

• Live raw_vector length matches GENERAL_FILE_INFERENCE_MODEL's trained dimensionality,
  verified by a real test.
• Per-language coverage scope (5-of-59) is a documented, deliberate decision, not a silent gap.
• A live/model dimension mismatch fails loudly instead of silently truncating via min().
• docs/wiki/03-05-claim-5-file-archetypes.md matches the currently shipped model's real
  dimension/cluster counts (or is explicitly marked stale/historical).