Add fluency-trap pre-graduation gate

Add a new gate file fluency_trap.js exporting check(candidate) -> {pass: bool, reason: string}. Gate runs a regex+rubric scan over candidate.title, candidate.thesis, and candidate.beachhead for known fluency-trap markers: (a) LLM-output-as-product patterns ('summarize X', 'rewrite Y', 'AI assistant for [content vertical]'), (b) content-as-moat patterns where the buyer is supplying the corpus, (c) generic 'agent for [profession]' framings without a falsifiable workflow specified. Returns pass=false + reason when any marker fires. Wire into graduate.js between the existing reachability veto and the layer4 handoff write — block + write reason into existing kill_reason channel; no schema change.

On retro pass against the last 30 graduated candidates, ~20-30% trip the gate (matching the manifesto-v4 deflation rate). On forward pass, graduation volume drops 20-30% with the dropped candidates concentrated in shape-rejection / fluency-trap kill_reason buckets.

Run the gate retroactively against the subset of graduated candidates that were ever commander-APPROVED. If commander-approved candidates trip the gate at >10% rate, the markers are over-broad and the gate is mistuned — revert.

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