Log non-convergent council transcripts to data/non_convergent/ JSONL

After the existing non-convergence detection path in council_verdict.js (the branch that produces a Commander-escalation outcome after 3 failed rounds), append a JSONL record to hypothesis_engine/data/non_convergent/YYYY-MM-DD.jsonl before returning. Record schema: { hypothesis_id, title, round_count: 3, model_a_final_verdict, model_b_final_verdict, model_a_reasoning_excerpt (first 400 chars), model_b_reasoning_excerpt (first 400 chars), timestamp_iso, cost_usd }. Add a single-pass redaction function pre-write that strips OAuth tokens and API keys via regex. No DB schema changes. No changes to the convergence logic, escalation status, or any other path — purely additive instrumentation on the failure branch. Add a deploy-time smoke test: inject 2 synthetic non-convergent model response objects and assert 2 JSONL lines are produced before merging.

After 30 days of active council cycles (current rate ~9/week, empirical non-convergence rate 22%), directory accumulates 8-12 non-convergent transcripts. This corpus enables first structured analysis of split-reason taxonomy: whether splits cluster by hypothesis type, ICP, or model-specific reasoning patterns. Currently 2 of 9 recent verdicts produce zero structured split-reason data.

After 30 days, if hypothesis_engine/data/non_convergent/ contains fewer than 5 JSONL lines while hypothesis_engine/moves/council_verdict.js internal logs show ≥10 Commander-escalation events during that window, the sink is silently dropping payloads. Validate by diffing JSONL file line count against the count of hypotheses with council_non_convergent=true status in engine.db.

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