Commander override classifier tool with fixed taxonomy

Add classify_override.js: reads pending rows from the existing overrides table (NOT the hypotheses table — overrides is a sibling table), calls Sonnet 4.6 with a fixed 6-bucket taxonomy {shape-rejection, fluency-trap, missing-context, wrong-vertical, scope-shift, other}, writes label + 1-line rationale to a new override_class column on the overrides table. Migration adds the column only (single non-null-defaulted text column on overrides, NOT on hypotheses). Tool runs as a one-shot script + a per-cycle backfill hook. The bucket labels then become an aggregable signal that other engine moves and proposals can index against.

After classifying the last 200 overrides, shape-rejection + fluency-trap combined account for >40% of overrides. This validates whether the fluency-trap gate proposal is targeting the dominant failure mode, or whether 'missing-context' / 'wrong-vertical' would be a higher-yield intervention.

If >50% of overrides land in 'other' on first pass, the 6-bucket taxonomy fails to span the actual override space — tool is useless until taxonomy is redesigned. Specifically: a sample-of-20 hand-classification by Commander disagreeing with the auto-classification at >30% rate also invalidates.

Rejected by filter_score. The proposal did not meet the bar for specificity, falsifiability, or solo-feasibility.