Clarify /start skill gate messages with intent classification

[sebastientaggart]

Both human gate messages in the /start skill now suggest "go" as a fast-path shorthand and include agent-side intent classification rules — affirmative (proceed), affirmative with conditions (discuss first), questions (discuss), and abandonment (stop). This replaces the previous yes/no framing with clearer next-step options for Case A (approach confirmation) and Case B (resume confirmation).

Issue #96

[sebastientaggart]

Review Summary

Verdict: APPROVE

Findings

• [WARNING] Nested double quotes in gate messages create ambiguous quoting — "Does this approach sound right? Type "go" to create..." has inner "go" breaking the outer "..." pair. Consider using single quotes or backticks for the inner token (e.g., Type 'go' or Type `go`) so the sentence boundary is unambiguous when an LLM parses it. This appears in both Case A (Step 2) and Case B (Step 2) gate messages across all 5 adapter files. Not blocking because LLMs will likely infer the intent correctly from context, but it is technically malformed quoting.

No other issues found. The intent classification rules are consistent across all adapter files, the changes match the PR description, and the sync hashes are updated. No secrets, no security concerns.

[sebastientaggart]

Review Summary

Verdict: APPROVE

Findings

No issues found. The previous WARNING about nested double quotes has been resolved — go is now backtick-quoted consistently across all 10 gate messages (5 adapter files x 2 gates each). Intent classification blocks are identical across all adapters, the changes match the PR description, sync hashes are updated, and no secrets or security concerns are present.