brainstorming D1: fresh-context spec review — investigate sprint-only dispatch

[SUaDtL]

Spun out of #597 (D1), implemented slice landed in #599.

The Phase 5 adversarial pass is same-agent. The stronger form dispatches a fresh-context reviewer against the written spec — the never-grade-your-own-homework discipline code already gets here.

Maintainer steer (2026-08-04): strongly leaning SPRINT-ONLY. Under attended /feature the user is themselves the second reviewer at the approval STOP, so a dispatched reviewer is ceremony on top of a human gate. Under /sprint there is no human at approval — SMARTS auto-approves — which is exactly where an independent reviewer earns its cost.

Investigate before implementing:

• Reuse an existing agent (decision-challenger pointed at specs, or design-quality-reviewer) vs. a purpose-built spec-reviewer definition; Make INDEX/routing-surface consistency a CI-checked invariant, not an authoring discipline #592 makes every new agent a routing-surface liability.
• Wiring: brainstorming Phase 5 under /sprint dispatches the reviewer BEFORE SMARTS scoring; does the verdict feed SMARTS as evidence, or gate it?
• Failure mode: a reviewer BLOCK on an auto-approvable spec — does it escalate to the user as a sprint true stop, or downgrade the spec to attended approval?
• Cost: one subagent per feature spec per sprint; measure on a real sprint before defaulting on.

Done when: a ruling (ADR or decision-log entry) fixes the scope to sprint-only or otherwise, and the chosen wiring ships with a test pinning that /feature attended runs do NOT dispatch it.