feat(verify): structural verification gate with edit->test->fix loop …

[wusijian007]

…(M3.3 / §4)

Third v3 milestone (§4 Self-Correction). Upgrades verification from model-opt-in (the verifier sub-agent, which the model may or may not spawn) to a structural loop gate. Includes the written §4 design section in docs/v3-kernel-roadmap.md (design-before-code per the roadmap's blast-radius tiering).

M3.3a -- verification gate (query.ts):
QueryOptions.verify = { command, args, when?, maxBounces? }. At the
completion path (the model emits no tool_uses -- "I'm done"), the loop
no longer returns immediately. It runs the verify command via
ToolContext.executor (the M2.1 seam, NOT the whitelisted Bash tool --
arbitrary npm test / tsc --noEmit is allowed). Exit 0 -> completes.
Non-zero -> the failure is injected as a reflective user turn
(reflectiveVerifyFailure: command + exit + truncated output + "locate
and fix; I'll re-run") and the loop continues an edit->test->fix cycle.
Bounded by maxBounces (default 2); exceeding it (or running out of
turns) ends with a verification_failed terminal state rather than a
silent completed. Each check yields a verification LoopEvent;
bounces also emit a query.verify_bounce profile mark.

M3.3b -- CLI + eval:
myagent agent --verify "<command>" parses into QueryOptions.verify
(whitespace split). The CLI prints a [verify] line per check. A 7th
eval task "self-correction" drives edit -> verify-fail -> fix ->
verify-pass through an injected scripted mock executor (exit codes
[1, 0]) -- fully deterministic, offline. The eval gate fingerprint
updated (tasks 7, turns 15, in 10800, out 625).

Determinism (invariant #2): the gate runs through ToolContext.executor, so the eval + the 4 new query-loop tests inject a mock CommandExecutor instead of spawning real processes. New TerminalState status "verification_failed" and a VerificationEvent added to the LoopEvent union.

finalize critic pass stays deferred to a §4 follow-up.

Local: 201 tests, 3/3 green.