fix(config-reload): retire watcher after provider scope closes

[choi138]

Summary

• treat only Provider scope is closed from scope-bound config-reload callbacks as a terminal lifecycle signal
• retire the orphaned watcher while preserving existing error propagation and non-blocking teardown behavior
• add a regression test for a debounced settings event delivered after the owning provider scope closes

Reproduction

On main, the new regression test fails from ConfigReloadWatchEngine.#evaluatePending with an uncaught Error: Provider scope is closed. With this patch, the callback closes the watcher and its subscriptions drain on the existing deferred teardown tick.

Verification

• bun run --cwd packages/coding-agent test test/suite/config-reload-lazy-teardown.test.ts (4 passed)
• config-reload test suite (7 files, 107 tests passed)
• bun run check
• isolated senpi QA: common.mjs --self-check (10/10), rpc-drive.mjs --self-test (4/4), mock-loop.mjs --with-tool --api openai-responses (4/4)

Changelog

This fixes a user-visible RPC host crash and prompt timeout. Per CONTRIBUTING.md, external contributors must not edit CHANGELOG.md; a maintainer changelog entry is still needed for the changelog gate.

---

Summary by cubic

Fixes a crash in the config-reload extension when a debounced filesystem event fires after the owning provider scope closes.

• Treats only Provider scope is closed as a terminal signal and retires the orphaned watcher.
• Preserves error propagation for all other callback errors and retains non-blocking teardown.
• Adds a regression test covering a debounced event delivered after scope closure.

^{Written for commit 4061523. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}