Vocabulary-bypass detector for closed-thesis archetype at council_verdict

Today closed_thesis_distance is only required when proposer's title/core_description uses one of the listed vocabulary tokens (genesis.js line 41). This is bypassable by paraphrase. Add a pre-council gate in council_verdict.js: at function entry, compute cosine similarity (use existing local-embed path from llm.js — bak_pre_s182_localembed shows it's already available) between the new proposal's title+core_description and the 19 archived oracle hypotheses (pool_status='archived' in hypotheses table). If max similarity ≥ 0.72, force closed_thesis_distance evaluation regardless of vocabulary triggers, and require a `defense` field. If max similarity ≥ 0.85, hard-reject with reason `semantic_closed_thesis_match` and log which archived id matched.

Of the next 50 genesis proposals, the gate fires on 2-5 (4-10%) that previously would have slipped past the vocabulary trigger. Manually inspecting the matched archived hypothesis for each fired proposal confirms a true archetype match in ≥70% of fires (precision check on a 5-fire sample).

Run the gate retroactively on the last 100 genesis outputs in engine.db. If zero proposals match an archived oracle hypothesis at ≥0.72 similarity, the bypass risk being addressed is hypothetical and the gate is unnecessary. If ≥20% match, threshold is too low and gate is over-eager.

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