Make ask_user question cards idempotent on retry (CL-7248)

[TheGreatAxios]

Summary

A crash between postQuestion and suspendOnGate used to mint a new question id and post a second card. ask_user now derives questionId from the tool-call id, and POST /participants/messages treats a re-post of the same id as a no-op that returns the existing card.

What changed

• questionIdForCall(call.id) gives a stable q_<hex32> id across retries of the same tool call.
• postQuestion stamps a caller-supplied questionId instead of always minting.
• The workflow-participant messages route looks up an existing question block in the workbench and returns that message when the id matches.

Tests

• Retrying ask_user for the same call reuses the question id.
• Re-posting the same question id returns the existing card and does not insert a second message.
• A different question id still posts a second card.

Linear: CL-7248

Fixes CL-7248

[TheGreatAxios]

Fixes CL-7248

[TheGreatAxios]

Self-review

Verdict: the idempotency mechanism is correct — deriving questionId from the tool-call id and treating a re-post of that id as a no-op closes the crash window described in CL-7248, and the new tests exercise both the retry and the different-id paths. No defects required a fix.

• CONFIRMED-FIXED: none — no defects found.
• DISMISSED: existingQuestionMessage paginates the workbench's full message history to check for an existing question id, on every ask_user post (not just retries). This is an O(n) scan per question rather than an indexed lookup, but it is a pre-existing pattern in this route (listMessages is the only read primitive available here) and bounded by one workbench's message count; not a correctness defect and out of scope for a minimal fix.
• DISMISSED: the route returns 200 instead of 201 on the no-op retry path — postQuestion's caller only checks response.ok, so this doesn't break the client.
• DISMISSED: questionIdForCall truncates a SHA-256 digest to 32 hex chars (128 bits) — collision risk is negligible and matches the existing q_<hex32> shape.
• DISMISSED: no exactOptionalPropertyTypes, bare-catch, as T, or vendoring issues found in the diff.
• CONFIRMED-OPEN: none.

Targeted tests run:

• bun test src/client.test.ts src/tool.test.ts (packages/interaction-tools) — 16 pass
• bun test test/workflow-participant-routes.test.ts (packages/chat) — 20 pass
• bun run typecheck in both packages — clean

[linear-code]

CL-7248