Prompt genesis proposer to reason about exception classes inline

In the PROPOSER_SYSTEM prompt of genesis.js, after the existing mandatory-field instructions (structural_veto, resolution_event, manual_entry_path), add a reasoning block that does NOT introduce a new JSON schema field: 'Before writing your hypothesis description, explicitly reason: what classes of cases does this workflow explicitly exclude or route to human review? Name at least 2 distinct exception classes — not failure modes of the tool, but structural boundaries of the workflow (e.g., tier of customer, data-quality condition, type of event that falls outside scope). Embed this reasoning into the description field using phrases such as: The workflow excludes..., Human review is required when..., or This tool is not designed for cases where... A hypothesis description that names zero exception classes signals an idealized workflow rather than a deployable one. Do not add a new JSON field for exceptions — embed the reasoning in the description.' This is reasoning pressure, not a schema mandate, specifically to avoid the boilerplate-bait failure of a mandatory exceptions field (P2 was killed for this reason in S158).

After 20 genesis runs post-patch, at least 30% of hypothesis descriptions contain explicit exception-class language ('The workflow excludes...', 'Human review is required when...', or equivalent). Current S157 baseline: 1/43 graduated candidates (2.3%) explicitly named exception classes in their descriptions. The shadow describability gate (B.v2, targeting Q3) becomes more useful as exception-class language starts appearing in genesis output.

After 20 genesis runs, if fewer than 5% of descriptions contain exception-class language — statistically indistinguishable from the 2.3% S157 baseline — the reasoning instruction is being overridden by other PROPOSER_SYSTEM instructions or dropped under token pressure. Counter-falsifier: if 80%+ of descriptions contain the exact same boilerplate exception phrase across different hypotheses (indicating the LLM is pattern-matching to satisfy the instruction without genuine reasoning), P2's boilerplate-bait failure mode has been reproduced in prose form — roll back and move the check to an evaluator-side rubric instead.

Rejected at the council verdict. The two-judge council did not find the case strong enough to advance to Commander review.