test(coding-agent): cover provider-error goal recovery

[realsigridjin]

Summary

• Add a real AgentSession regression that joins terminal provider failure, persisted blocking, next-input reactivation before agent_start, and completion.
• Prove deliberate model/user blocks remain blocked, including an unchanged numeric blockedAt.
• Add a hermetic source-CLI RPC scenario using the existing localhost fake provider, auth guard, and sandbox helpers.

Context

The stale-goal behavior was reproduced on published Senpi 2026.8.3-3. Production code is already fixed by #688 and included in newer releases; this PR closes the remaining end-to-end coverage gap rather than duplicating that fix. Existing coverage manually invokes goal handlers or tests direct input separately, so it would not catch a break in the full provider-error -> input disposition -> agent_start ordering.

RED -> GREEN

• Pre-fix classifier mutation (remove the provider-error reason): the next real agent_start observed [null, "blocked"] instead of [null, "active"].
• Deliberate-block mutation (classify Waiting on a user decision as mechanical): the real direct prompt observed active instead of blocked.
• Restored current main: goal-provider-error-recovery.test.ts passes 3/3.

Manual QA

node .agents/skills/senpi-qa/scripts/goal-provider-error-recovery.mjs \
  --self-test \
  --evidence goal-provider-error-recovery

Result: 14/14 checks passed. The real source CLI made exactly seven localhost provider requests, reactivated the mechanical block on the next direct prompt, preserved the intentional block on a later prompt, exited RPC with code 0, left real auth unchanged, stopped the fake server, and removed its sandbox.

Validation

• npm --prefix packages/coding-agent test -- test/suite/goal-provider-error-recovery.test.ts test/suite/goal-abort-extension.test.ts test/suite/goal-extension.test.ts test/suite/goal-turn-usage.test.ts — 53/53 passed
• npm run check — passed
• node --check .agents/skills/senpi-qa/scripts/goal-provider-error-recovery.mjs — passed

---

Summary by cubic

Added an end-to-end test and a hermetic RPC QA script to cover goal recovery after a terminal provider error. Confirms the next direct prompt reactivates the goal before agent_start, while intentional user blocks remain with the same blockedAt; runs fully against a fake localhost provider with no real auth or traffic.

^{Written for commit 30fce90. Summary will update on new commits.}