feat(build): loop-until-verified Phase 4

[ryzizub]

Problem. Phase 4 runs the suite once then ships; the fix loop caps at 3 attempts then asks the user. Nothing re-runs criteria until all pass.

Proposal. Re-run each criterion's verify: command until green; ship only when every box is checked by real output; escalate only when a criterion is un-runnable or self-contradictory. Run the final VERIFICATION COMMAND before /create-pr.

Acceptance criteria.

• Build re-runs failing criteria without re-prompting on objective fails.
• Ship is blocked until all criteria pass.
• Final verification command runs before PR creation.

Depends on #203 (machine-checkable success-criteria gate).

[omartinma]

Additional context: the Flutter plugin's new green-gate skill (VeryGoodOpenSource/vgv-ai-flutter-plugin#118) already does this loop for Dart/Flutter, analyze then format, test, coverage, rerun until green. So Phase 4 shouldn't rebuild that generically. It should delegate to it.

Simplest way to link them: add a verificationSkill field to the existing hooks/recommendations/very-good-ai-flutter-plugin.json, pointing at very-good-ai-flutter-plugin:green-gate. Phase 4 reuses the detection we already have there. If the project matches and the skill is installed, call it to drive to green. Otherwise fall back to the generic lint+test loop.

The verify: commands from #203 line up with green-gate's gates on a Flutter project, so the two stay consistent.

This keeps Wingspan tech-agnostic and the plugin name stays in the registry file, not in every skill.