refactor(aidd-orchestrator): the check zone says when to stop, and reviews its axes in one round

Read plugins/aidd-orchestrator/skills/01-sdlc/references/03-check.md alongside the issue's acceptance criteria, focusing on the loop exit rule and review-axis dispatch. Done means the file states the reachable-input exit condition, gives out-of-scope findings a destination, combines parallel findings, and preserves regression re-entry after a fix.

Problem

The Check zone says to re-enter after every new candidate and never says when to leave. It also reviews with one checker, sequentially, while its own behavior section says to parallelize independent work.

Measured on a real run (#250, PR #885): six review rounds, ten subagents, all sequential.

Rounds 1 to 4 each found a defect a real edit would hit. Rounds 5 and 6 found only shapes nothing in the repository can produce, and cost roughly an hour between the reviews themselves and the fixes they prompted.

Total: 2h12 of agent wall-clock and ~1.4M subagent tokens, for work where the axes were independent the whole time.

Scope

A stated exit condition. Check leaves the loop when the reviewer returns no finding that an input reachable from the repository can trigger. A finding about a shape nothing produces is recorded — as a spec limit or a follow-up issue — and does not re-enter the loop. The zone says this, so it is a rule and not a judgement call made differently each run.

Review axes run in parallel. Code, contract conformance and adversarial inputs are independent: none needs another's verdict. One reviewer per axis, dispatched together, their findings merged before the loop decides. A regression pass after a fix stays single, because it only re-checks what changed.

Acceptance criteria

• plugins/aidd-orchestrator/skills/01-sdlc/references/03-check.md states when the loop ends, in terms a reader can apply without knowing this run.
• A finding the exit condition excludes has a stated destination, so nothing is dropped by being out of scope.
• The zone dispatches its review axes in one round rather than one after another, and says how their findings combine.
• A run that finds a defect in round one still re-enters Check after the fix — the exit condition narrows the loop, it does not remove it.

Out of scope

• Model tiering per step. That is #677.
• Measuring what a run costs. aidd-telemetry already does it and is simply not enabled on this repository.