Fix start skill re-prompting and deploy lightweight tag failure

[sebastientaggart]

Two skill reliability fixes observed on consumer projects:

1. Start skill — Agents re-ran milestone detection and re-prompted the user after the Step 2 human gate. Added an explicit execution-order directive to the Parsing section so labels/milestones are resolved before Step 1, and updated Step 3 to reference already-resolved values instead of re-running resolution.

2. Deploy skill — Lightweight tags created by bump commands silently fail when tag.forceSignAnnotated = true is set in git config. Added a tag verification step after bump commands with an annotated-tag fallback (git tag -a -m).

Issue #145

[sebastientaggart]

Review Summary

Verdict: APPROVE

Findings

No issues found. Code looks correct and follows project conventions.

Both fixes are well-scoped instruction changes to skill templates:

• Start skill: The execution-order directive correctly moves label/milestone resolution before Step 1, preventing re-prompting after the Step 2 human gate. Step 3 is updated consistently to reference already-resolved values.
• Deploy skill: The tag verification + annotated-tag fallback handles the tag.forceSignAnnotated = true edge case cleanly.

All adapter outputs (.claude/commands/, .cursor/rules/, .agents/skills/, .gemini/skills/) are consistent with their source templates and have updated sync hashes.