docs: surface anti-patterns earlier in TEMPLATE.md + add PROPOSAL/REVIEW templates

[indykish]

Summary

Three structural improvements to the spec-authoring templates based on observed drift in cache-kit.rs/plans/.

• TEMPLATE.md — Document Types Map at the top, anti-patterns moved up + tabulated, Provenance field added, pseudocode litmus test surfaced, prologue minimum (3-5) made explicit, 300-line cap visible.
• PROPOSAL_TEMPLATE.md (new) — for design exploration before commitment. Banner: "do not implement directly." Forces ≥2 options + a do-nothing baseline. Conversion-to-spec tracking.
• REVIEW_TEMPLATE.md (new) — for code/audit reviews with many findings. Banner: "each actionable finding becomes its own spec — don't bundle." Findings table with Severity (P0/P1/P2/Note) + Status (OPEN/SPEC'D/REJECTED/OUT_OF_SCOPE) + Spec column.

Motivation — observed failure modes

Doc	Failure mode
cache-kit.rs/plans/comprehensive-code-review.md	Code review masquerading as a spec — 12 action items in one file, none actionable as-is. "Add circuit breaker" with no scope, no API, no tests.
cache-kit.rs/plans/error-codes-improvement.md	Proposal masquerading as a spec — 850 lines of inlined Rust pseudocode (pinned enum values, full match arms, sample test bodies). All of it will drift the moment implementation lands.

The fix isn't more discipline — it's making the correct shape obvious at template-pick time, before the author drifts. Both docs would have been forced to a different shape if the Document Types Map had been on line 1.

What changed concretely

TEMPLATE.md

• Added Document Types Map (Spec / Proposal / Review / Postmortem) at the top.
• Pseudocode litmus test (one-line callout) inside Spec Registry.
• Moved 6 anti-patterns from line 173 → near top, converted to table form.
• Added `Provenance:` field to the spec header (human / LLM-drafted / agent-generated) so executing agents calibrate trust.
• Made prologue minimum explicit (3-5 pointers).
• Promoted "no test code in this section" to a top-of-section callout in Test Specification.
• Surfaced the 300-line cap as a hard upper bound.

PROPOSAL_TEMPLATE.md (new — 93 lines)

• Banner + clear contrast with TEMPLATE.md.
• Required: ≥2 options + a "do nothing" baseline. If you only have one option, you have a spec, not a proposal.
• Open Questions section forces explicit blockers before conversion.
• Conversion-to-Spec section links the resulting spec(s).

REVIEW_TEMPLATE.md (new — 97 lines)

• Banner + clear contrast.
• Findings table — one row per finding, with Severity + Status + Spec link columns.
• Severity guide: P0 (must spec now), P1 (spec when capacity allows), P2 (polish), Note (record-only).
• Closed only when every P0/P1 has either a spec or a REJECTED row.

Invariance

Audit script passes (18 gates, fixtures clean). `AGENTS.md` untouched; questionnaire scenarios 1-20 unchanged. Sign-off refreshed for 79aba99.

Test plan

• Read TEMPLATE.md top-to-bottom; the Document Types Map should reroute you before reaching the spec body.
• Try drafting a spec for one of the cache-kit recommendations ("add circuit breaker") against the new TEMPLATE.md and verify the Anti-Patterns table catches the drift.
• Try porting `cache-kit.rs/plans/comprehensive-code-review.md` into REVIEW_TEMPLATE.md and confirm the structure surfaces each finding as a discrete row.
• Confirm `audit-agents-md.sh` continues to pass on this branch.

🤖 Generated with Claude Code