feat(goal): cheap Haiku inner-review tier + selectable model + optional Copilot

[JFK]

Summary

/goal already implemented the "short invocation, policy in the file, DoD over micro-steps" design well, so this is additive, not a rewrite. Three changes, all scoped to the review pipeline and its cost:

1. Cheap inner-review tier (step 5b). Before /ship's heavy gate2 cascade, a reviewer subagent reviews the working-tree diff (git diff — correct for the pre-PR stage, no PR exists yet) on the goal.inner_review.model tier (default haiku). Fix findings → re-check, capped at max_rounds (default 2). Implementation stays on the normal session model; only the reviewer downshifts. gate2 remains the authoritative gate — the Haiku pass is a fast filter, never a replacement.
2. Tighter TDD wording (step 5b). Red→Green→Refactor framed as an invariant, not a forced march: drive test-first where there's a test surface, keep cycles tight, and explicitly skip the ritual (no token tests, no mechanical "refactor") when there's nothing to assert.
3. Cost/selectability controls (reusing /ship's existing contract — no new Copilot keys):
   • PR review (the expensive Copilot loop) is optional via the inherited review.provider, or the per-run no-copilot flag (forwarded to /ship).
   • Inner-review model is selectable via goal.inner_review.model or the per-run review-model=<tier> flag.

Files

• commands/goal.md — step 5b rewrite, flag parsing (no-copilot, review-model=), config resolution, 5c forwarding, plan block, deps table, frontmatter.
• commands/config.md — goal.inner_review.* defaults + goal.* table rows + cost-control note.
• README.md / README.ja.md — /goal flag-table parity.

Config (new keys, backward-compatible defaults)

"goal": {
  "autonomy": "red-only",
  "max_issues_per_run": 20,
  "inner_review": { "enabled": true, "model": "haiku", "max_rounds": 2 }
}

Defaults keep Copilot on and inner-review on — the cost levers exist but aren't forced. Flip review.provider / inner_review.enabled for a cheaper default posture.

Test results

All existing suites green, no new drift:

• enum-sync-check — 33/33 in sync
• jq-sync-check — 5/5 in sync
• test_state_schema — 60/60 pass
• copilot-detection, test_verdict_parser — pass
• Built-in-defaults JSON re-validated with the new goal.inner_review block

Notes / follow-ups

• No CHANGELOG entry or version bump — those are tag-time, owned by /tag (the CHANGELOG is a release index).
• No automated test added for the new flags/config (the command files are prose specs; existing tests cover schema/enum sync, which still pass). A fixture exercising no-copilot forwarding + review-model override could be added if you want coverage there.
• Three review layers now exist (Haiku inner pass → optional /code-review fallback → gate2 cascade); the inner tier is deliberately the cheap filter so this doesn't become review bloat.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR updates the /gh-issue-driven:goal milestone driver to add a low-cost “inner review” pass before delegating to /ship, along with new configuration defaults and README flag documentation to control review cost and model tier selection.

Changes:

• Add an inner-review loop in /goal step 5b (default Haiku tier) with max-rounds and per-run model override.
• Make PR review cost controllable by forwarding no-copilot to /ship and documenting inherited review.provider.
• Document new goal.inner_review.* defaults in /config and keep README flag tables in sync (EN/JA).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
README.md	Updates /goal command synopsis to document inner-review and new flags.
README.ja.md	Japanese parity update for /goal synopsis and flags.
commands/goal.md	Adds inner-review tier, tighter TDD guidance, new flags (no-copilot, review-model=), and forwarding behavior to /ship.
commands/config.md	Introduces goal.inner_review.* defaults and adds cost-control documentation for /goal runs.