Add NBJ 5-question describability pre-check at start of argument.js

At the start of argument.js, before the adversarial debate loop begins, add a describabilityPreCheck(hypothesis) function that makes a single cheap model call (Gemini Flash or equivalent) with the five S157 NBJ questions as a structured prompt: Q1 inputs nameable (can buyer write down exact artifacts before engaging the tool?), Q2 outputs nameable (can buyer name deliverable shape in pre-AI terms?), Q3 exception classes named (can buyer name upfront cases where workflow should not fire?), Q4 accept/reject criteria for one canonical case (can buyer state correct output for a specific input without invoking model confidence?), Q5 describable without AI as a primitive verb (does the workflow form survive without the LLM being the verb?). Each question binary pass/fail. If composite ≤ 1/5, set hypothesis status to 'pre_argument_kill' with kill_reason 'describability_gate_fragile', persist to DB, and return without running the argument debate. Log the per-question breakdown as a JSON field in the hypothesis record. Estimated cost per check: ~$0.001. The gate fires only on hypotheses that have already passed filter_score, so it sees a pre-filtered population.

Applied to 43 S157 graduated candidates: hyp-2026-05-06-847f7e (0/5 on S157 manual review) is killed before argument; none of the 25 ROBUST candidates (4-5/5) are killed. Per the move cost rollup, argument + council_verdict + 7 deep moves average approximately $0.12-0.18 per hypothesis reaching that stage. Each structurally fragile candidate caught here saves that full downstream cost. At current council volume (9/week), prevents ~1 structurally fragile hypothesis from consuming full debate budget per week.

Apply gate retroactively to 43 S157 candidates using stored hypothesis descriptions. Gate must kill 847f7e (0/5) and must NOT kill any of the top 5 ROBUST S157 candidates (ec4507, 24a849, ec4507, e9cb5c — 5/5 scorers). If any 5/5 ROBUST candidate is killed (gate score ≤1/5 despite S157 rating it 5/5 manually), evaluator is misinterpreting well-specified hypotheses. Threshold: zero false kills among ROBUST; at minimum 847f7e caught.

Accepted with status: accepted with revision shadow mode. Awaiting or completed implementation by Architect.