Align skill metadata and headings with honest evidence lanes

[BaseInfinity]

Follow-up from #115

The bounded final review for #115 certified the candidate with no P0/P1 blockers, but found small terminology and coverage cleanup that should not force another broad proof/review cycle on the already-certified delivery.

Scope

• Replace stale TDD-only discovery metadata with the four honest evidence lanes: writable RED, observable EVAL, mechanical assertion, and judgment review.
• Align the README AGENTS description and the remaining TDD Workflow heading.
• Make the baseline list RED explicitly alongside EVAL/assert/review.
• Decide whether the active-surface regression should cover PROVE-IT.md, START-SDLC.md, and skill metadata, and remove any dead assertion alternatives.

Exclusions

• Do not reverse Scope TDD RED to where a RED mutation is writable — an unqualified mandate produces fake tests #115's deliberate implement-first rule.
• Do not rerun broad proof repeatedly for wording-only corrections; use affected checks and one completion proof.

Evidence

#115 final gate: Sol High CERTIFIED at 94 confidence; Opus 4.8 xhigh fallback CERTIFIED at 85 confidence; initial agreement, zero reconciliation rounds.

Related: #115