Unwind runner resources in reverse order on shutdown (#37)

[hsliuustc0106]

Closes #37.

What

Registrations are effects (adapter-seam.md admission discipline 2), applied to the native runner:

• new core.teardown.Teardown — a registry: resources register a named disposer in creation order; run() pops them in reverse, runs every step even when one raises, reports failures by name only (never exception text, which can carry private data), and is idempotent across repeated stops. Not thread-safe by contract: register and run from the orchestrating thread.
• TaskLoop.close() — drains an in-flight summary bounded by the summary budget (a hung provider costs the wait, not the shutdown), then drops the per-task single-flight locks. Safe to call twice.
• _wiring(teardown) registers task-store → activity-log → memory-store → task-loop in creation order; _run_runner unwinds them in a finally inside the RunnerLease, for both the serve and --once paths. Previously closing the SQLite connections was left to garbage collection after serve returned — nondeterministic and unordered (any lingering reference, e.g. an exception traceback or the summary worker, could keep a store open past the stop window); now they close explicitly, in reverse creation order, while still under the lease.
• runner_control.py is deliberately unchanged: its __exit__ already unwinds in correct reverse order (watcher join → file cleanup → lock release) and its behavior is pinned by its own test suite.

Cooperative-stop semantics are untouched: stop_runner still waits for real lock release and reports a pending stop on timeout; teardown completes before ownership is released, so a replacement runner never meets a half-closed store.

Deliberate edges (documented, not coded)

• If prepare() itself fails mid-wiring, partial registrations are not unwound — the runner is a whole process exiting on the spot, ownership is still released by RunnerLease's exception path, and the OS reclaims the rest.
• A second SIGINT during the teardown window could interrupt it; the first SIGINT already triggers the graceful stop path, and the process exits regardless.

Tests

• tests/test_teardown.py: reverse order + idempotence; every step runs on failure with names-only reporting; close() bounded under a hung provider and draining a finishing one; the real wiring unwinds all four registrations in reverse on runner --once; unpatched pass leaves no shutdown warnings and readable data.
• test_runner_loads_policy_under_lock_before_ready updated to the new _wiring(teardown) signature (same intent: configuration loads under ownership).

545 passed in 22.83s

Full offline suite (dead-proxy), plus examples/first_pr_watch.py: all 8 scenarios pass, including the real background-runner cooperative-shutdown scenario.