feat(critic): Definition-of-Done gate chain + Finalize Critic (M3.5 /…

[wusijian007]

… §4)

Generalize the single structural-verify gate (M3.3) into a DoneGate chain run at the completion path: when the model stops calling tools, each gate must pass; the first failure injects a reflective user turn and the loop continues, bounded by ONE shared bounce budget. With a single gate this is byte-for-byte the old verify behavior (zero behavior change; M3.3 tests stay green).

Add a second gate, the Finalize Critic: a read-only model call (no tools given, so read-only by construction; core cannot import the Agent tool, so a tool-less critic call is the layer-safe realization of the read-only verifier idea) that judges whether the final answer satisfies the root task. APPROVE completes; REJECT injects a reflective revise turn. It runs in its own child context (a single synthesized message), so it does not pollute the parent prefix until the reflection is appended (invariant #3). Gate order is verify-then-critic (fail-fast: no model call if it does not even compile).

• core: QueryOptions.critic (CriticConfig), CriticEvent, gate-chain runner, createVerifyGate/createCriticGate, runCritic/parseCriticVerdict.
• cli: myagent agent --critic [--critic-instructions ""], [critic] event printing, help/usage text.
• eval: new finalize-critic task (reject -> revise -> approve) on a separate scripted critic FakeModel; gate fingerprint updated to tasks=9 turns=19 in=13150 out=765.
• docs: CLAUDE.md query-loop step 8 rewritten as the gate chain; roadmap M3.5 section + decisions anchored.

210 tests pass (+6 critic query tests).