Add kill-reason taxonomy classifier tool

Today kill reasons are free-text strings written by Commander or council moves; aggregating them in Corpus E requires manual interpretation. Add a single-shot classifier (Sonnet 4.6, temp 0) that maps each raw kill_reason to one label from a controlled vocab: {off_thesis, unreachable, candy_not_pain, redundant_with_existing, complexity_blown, low_revenue_ceiling, provider_displacement_risk, time_moat_negative, other}. Persist normalized label alongside raw text. Add a nightly cron that classifies any unclassified KILL rows. No schema migration: add column normalized_kill_reason TEXT NULL via additive ALTER.

Weekly kill-reason distribution chart becomes computable without manual aggregation; the 'other' bucket stays under 20% of classifications after 14 days, indicating vocab covers observed reality.

If 'other' exceeds 30% after 14 days of operation, the vocab is missing categories — i.e. either Commander reasoning is more diverse than the v2 axes assume, or the classifier is failing to map to existing labels. Either way, the kill-reason aggregations in Corpus E are not the trustworthy meta signal the engine has been treating them as.

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