Extract invest-gate subagent

[gringolito]

What

Extract the INVEST validation logic into a dedicated subagent at agents/invest-gate.md. Update every command that currently inlines INVEST validation to delegate to this agent.

Why

INVEST validation appears verbatim across add-backlog-item.md, migrate-backlog.md, refine-backlog-item.md, and validate-backlog.md. Any wording drift between copies is a real bug — refinement and creation can disagree on what "ready" means. Pulling the prompt into a subagent collapses 4 copies into 1, gives the orchestrator commands cleaner context, and lets us pin a cheap model (haiku) since the check is mechanical (six pass/fail letters with one-line justifications).

Sub-issue of #34.

In Scope

• Create agents/invest-gate.md with frontmatter:
  • name: invest-gate
  • model: haiku
  • effort: low
  • disallowedTools: Write, Edit (read-only validator)
  • description: Returns PASS/FAIL per INVEST letter for a given issue body, with one-line reasoning per criterion. Invoke when a backlog item needs gating.
• Body of the agent file: clear input contract (issue body text + optional metadata), explicit output schema (per-letter PASS/FAIL + reasoning + overall verdict), and the canonical INVEST table from the project docs as the rubric.
• Update callers:
  • commands/add-backlog-item.md — replace inline INVEST check with delegation to invest-gate.
  • commands/migrate-backlog.md — same.
  • commands/refine-backlog-item.md — same. The pre-removal validation gate becomes an invest-gate call.
  • commands/validate-backlog.md — same. The INVEST-related part of the audit calls the agent per item.
• Cross-reference: see Re-validate effort label in /refine-backlog-item pre-removal gate #12 (re-validate effort label in pre-removal gate) and Expand /refine-backlog to include issues with incomplete metadata #19 (refine-backlog incomplete metadata) — the contract here may need to flex to satisfy those callers.

Out of Scope

• Other agents (separate sub-issues).
• Adding new INVEST criteria or relaxing existing ones.
• Refactoring the issue body shape (### What/### Why/etc.).

Acceptance Criteria

• agents/invest-gate.md exists with the frontmatter above.
• All four caller commands delegate to invest-gate and contain no inline INVEST prompt logic.
• Output schema is documented in the agent body and consistent across callers.
• In a scratch repo, /add-backlog-item for a deliberately under-specified item produces a per-letter FAIL report from invest-gate.
• claude --debug shows the agent registered.
• Existing CLAUDE.md consistency greps still pass.

INVEST Notes

• Independent — agent design is self-contained; callers update mechanically.
• Negotiable — model pinning may move from haiku to sonnet if smoke testing shows gaps.
• Valuable — eliminates the largest single source of drift across command files.
• Estimable — S.
• Small — one new file, four small caller edits.
• Testable — scratch-repo smoke test in Acceptance Criteria above.