feat(cli): add a per-stage accept-all to curate and honor review (#385)

[jasonssdev]

Closes #385. Follow-up filed as #508.

A test run needed 74 individual confirmations in the Structure stage alone, while openkos.yaml had carried review: true all along and this verb never read it. Measuring first: cfg.review is honored in 16 sites in main.py and zero in curate.py.

The deeper cause is an asymmetry. curate --auto consents to model spend, never to a per-item write (design rule D3), and its escape hatch is the unattended_hint pointing at the standalone verb — where --auto does write. So suggest-relations --auto has always written 74 relations unattended while curate --auto wrote none.

What ships

curate --accept structure,metadata answers those stages' per-item prompts without asking. review: false does the same for every acceptable stage when no flag is passed.

Identity is excluded structurally — three independent guards

A merge absorbs one concept into another and deletes the absorbed file. One guard is not enough when the failure mode is a silently destroyed concept:

1. auto_acceptable is a field on the Stage descriptor and Identity keeps the False default. A future writing stage cannot inherit accept-all merely by being added to _STAGES.
2. --accept identity is refused with exit 2, and review: false only ever populates acceptable stages.
3. Identity's walk calls _confirm directly rather than through _confirm_item, so the acceptance path cannot reach it at all.

This is load-bearing, not belt-and-braces. Mutating guard 1 alone does not make merges silent; mutating guards 1 and 2 together does, and the test catches it. There is also a fourth surface: the non-TTY write refusal is lifted by accepted_stages membership, so if Identity ever became acceptable it would gain unattended merge authority — concepts deleted by a config line and a pipe, with no prompt anywhere to decline at. That path has its own test, and mutating it produces an exit-1 crash rather than a silent merge.

Other decisions

• --accept IS write consent, so an accepted stage passes the non-TTY refusal. curate --auto --accept structure writing on a pipe is parity with suggest-relations --auto, not new authority.
• An explicit --accept narrows rather than widens review: false. A flag that could only ever add would leave an operator no way to re-review one stage without editing the config mid-session.
• Refusals run before the workspace gate, so curate --accept strcture outside a workspace reports the typo as itself — the ordering list's TYPE and set-volatility's tier already use.

Confidence thresholds are out of scope, and the issue's plan was inverted

#385 proposed auto-accepting above 0.9. That is not implementable on the stages that hurt: EdgeSuggestion carries edge/suggested_type/rationale and TierSuggestion carries type_name/current_default/suggested_tier — neither has confidence. The only confidence in the project lives in adjudication.py, i.e. Identity, the destructive stage. Filed as #508, which has to add confidence to the suggesters first (a prompt change, so it needs a cap-harness A/B, not just a green suite).

Verification

• Suite 4040 → 4048, all passing. ruff and mypy clean over the whole repo.
• Mutation-verified, including the composite invariant above and the non-TTY path. One test passed on first write; rather than trust it, it was mutated until it discriminated, which is how the fourth surface was found in the first place.

Disclosure: no review receipt

Same lens defect as PRs #503, #506 and #507: the review agents are configured with only Read/Grep/Glob and cannot run the read-only Git commands the contract requires, so inspection returns incomplete. Shipping without a receipt is the maintainer's standing decision this session.