Story-3112 Walkthrough

Outcome: Blake can emit an auto classify-field intent only after the distinct-roster critic passes at or above the configured 0.70 bar. Missing, failing, and below-bar critic verdicts emit suggest.

classify proposal
  → build field intents as suggest
  → run distinct-roster classify critic
  → aggregate verdict using registrar.yaml confidence.critic_pass_min
  → pass: restore the earned manifest policy
  → absent / fail / below bar: keep suggest
  → emit typed broker-mediated intent

Boundary preserved

The change only stamps intent policy. Blake still writes no canonical truth and holds no execution credentials; the broker remains the sole executor. Exception flags and vocabulary intents keep their existing policies.

Executable proof

Critic result	Emitted field policy
absent	`suggest`
fail at 0.90	`suggest`
abstain after runtime error	`suggest`
pass at 0.69	`suggest`
pass at 0.70	`auto`

The regression asserts the serialized wire intent from runtime._turn_event, not only the critic decision object.

Verification

uv run ruff check agents/registrar tests evalkit       → passed
uv run --package registrar pytest agents/registrar/tests -q → 521 passed
uv run pytest tests -q                                  → 444 passed