refine: back design-doc behavioral assumptions with evidence

[shinpr]

Summary

Design docs frequently rely on unstated behavioral or factual claims — a framework/library default, a capability assumed already provided, or a feature assumed already implemented — that then flow into implementation unverified. The existing dependency check only confirms that a component exists, not that it behaves as the design assumes.

This adds an evidence-backed assumption check that runs end-to-end across the generator, the template, and the reviewer, so each relied-upon behavioral claim is either backed by evidence at design time or explicitly bound to a follow-up.

What changed

• Generator (technical-designer, backend + frontend variants) — a Behavioral Claim Verification step under Existing Code Investigation. For each relied-upon behavioral claim, attach one evidence source (codebase file:line, command result, or authoritative doc) and record it with Confirmed: Yes. A claim without locatable evidence is recorded Confirmed: No plus a matching Risks and Mitigation row that restates the claim and names how it will be verified during implementation (a command, test, or code-inspection point) or guarded by a fallback.
• Template (design-template) — an Assumed Behaviors slot in the Agreement Checklist that reuses the existing Evidence / Confirmed pattern. The restated claim is the shared lookup key linking the slot row to its Risks and Mitigation row.
• Reviewer (document-reviewer) — a behavioral-claim evidence check in the quality gate. It uses declarative phrasing (already, by default, defaults to, handled by) as scan starting points (a hint set, not exhaustive) and flags, as an important / feasibility issue, any relied-upon claim that is absent from the slot, marked Confirmed: Yes without evidence, or marked Confirmed: No without a matching Risks and Mitigation row.

Design notes

• Reuses existing structures — the Agreement Checklist evidence pattern and the Risks and Mitigation table — rather than introducing a new table, enum, or ID scheme, keeping the added surface proportionate.
• Reuses the existing feasibility issue category; no new reviewer enum value.
• All instructions are positive-form with measurable if-then predicates.

Validation

• pnpm sync + pnpm sync:check clean across all plugin distributions.
• pnpm check:skills-index consistent.
• claude plugin validate --strict passes for the marketplace manifest and all four plugins.
• Patch version bumped to 0.22.3.

🤖 Generated with Claude Code