Bounded structured evals: reasoning budgets and schema/verifier parity

Observed 2026-09-08 while repairing the ender-inference staging semantic evaluator (COR-4079). OpenAI Responses max_output_tokens covers both visible output and hidden reasoning, so a small structured-evaluation budget can be exhausted before any JSON is returned when a reasoning-capable model uses its default effort. Explicitly setting reasoning effort to none and rejecting non-completed Responses status resolved the failure. A second failure class remained possible when the JSON Schema permitted status/reason combinations that the local semantic verifier rejected; encoding the compatibility matrix directly in the schema, while retaining the verifier as defense in depth, closed that gap.

Provenance: redacted staging canary and dashboard verification on 2026-09-08; regression tests added in https://github.com/e3-solutions/ender-inference/pull/152; deployment evidence https://github.com/e3-solutions/ender-inference/actions/runs/34193665840. No call content or identifiers included.