fix(coding-agent): reject empty advisor stop completions

[roboomp]

Repro

A focused advisor-runtime regression models a provider returning HTTP 200 with finish_reason: stop, content: [], zero usage, no advise tool call, and no state.error. On the pre-fix HEAD worktree, bun test packages/coding-agent/src/advisor/__tests__/advisor.test.ts -t "treats an empty stop turn without advise as a failed advisor turn" failed with Expected length: 3, Received length: 1, showing the first empty stop was accepted as success instead of retried.

Cause

packages/coding-agent/src/advisor/runtime.ts only inspected agent.state.error after AdvisorAgent.prompt(). Empty successful-looking assistant turns with stopReason: "stop" and no content/tool call left state.error undefined, so AdvisorRuntime.#drain() marked the batch successful, reset consecutive failures, and decremented backlog.

Fix

• Added advisor-turn validation after each prompt to reject empty stop responses or turns that append only the user batch without any assistant result.
• Reused the existing advisor failure path so empty responses trigger rollback, onTurnError, retry, and the three-failure backlog drop + notifyFailure warning.
• Added a regression test covering the reporter's empty stop response shape, retry count, failure hooks, rollback behavior, and no-advice output.
• Added an Unreleased changelog entry for the advisor failure handling fix.

Verification

bun test packages/coding-agent/src/advisor/__tests__/advisor.test.ts -t "treats an empty stop turn without advise as a failed advisor turn" passed with 1 test and 36 assertions. bun test packages/coding-agent/src/advisor/__tests__/advisor.test.ts passed with 73 tests and 299 assertions. bun run fix passed after adding the exhaustive callback return. The pre-publish gate in gh_push_branch completed and pushed 8608395eefb2. Fixes #5212