Fix hung invoke freeze and reap OS children on chat stop

[wu1w]

One hung WASM invoke or audit write could freeze the whole host, and Stop left isolation_spawn_os children running until a 600s dispatcher tick. This PR keeps the personal-desktop fail-open: no new court asks, capability gates, permission prompts, or tighter allowlists.

What changed

1. A stuck invoke cannot wedge the host

• wasm_invoke snapshots the compiled module, drops the kernel RwLock, runs the guest, then applies status under a short write lock.
• wasmtime epoch interruption is enabled. Host TCP/stdio 30s dispatch timeout increments the engine epoch so the guest actually traps (spawn_blocking cannot be cancelled).
• Audit JSONL append is a bounded background queue. A stuck disk no longer holds the kernel write lock. Full queue drops the line (fail-open) instead of stalling mediate/charge/court.
• Host also starts a 5s isolation poll loop so exited children are reaped even when the Python dispatcher is idle. max_age stays 600s so long-running intended children are not killed early.

2. Chat stop reaps OS children

• Chat stop already calls llm_release_by_process (not end_process). That RPC now isolation.drop_process and SIGKILLs the process group.
• end_process already called drop_process; kill() no longer wait()s on the calling thread, so a wedged child cannot hold the write lock.
• No new confirm. Default tool permissions unchanged.

3. kernel-ci stale test

• Confirmed on current main: loop_guard::tests::orch_window_force still expected a trip after ~3 orch rounds. Production window is ≥6 of last 8.
• Test updated to require allow for 5 orch-heavy rounds and orch_window_thrash on the 6th. Production thresholds were not loosened.

What I verified

• Static review of current main (cfccbbb) against the C1 / H5 / H7 hypotheses. They held.
• cargo test -p tevarn-kernel --locked --all-targets: 117 lib + 4 abi_v1 passed (including orch_window_force, wasm_invoke_does_not_hold_kernel_lock, epoch_increment_traps_infinite_loop, end_process_reaps_os_children, llm_release_by_process_reaps_os_children).
• cargo check -p tevarn-kernel-host --locked succeeded.

Deliberately skipped (smoothness)

• No court ask list expansion (file_edit / remote_exec / shell_session).
• No mcp_* allowlist tightening, extra_roots rewrite, steward/_workforce checks, or per-call policy confirmation.
• No “refuse new WASM until lock free” gate (that would idle-spin the agent). Isolation is unlock + epoch trap.
• No Python/frontend changes. Chat stop is wired through the existing llm_release_by_process RPC.
• No test_court_authority.py added to kernel-ci (out of scope; would not fix the cargo-first red).
• No new crates or drive-by refactors.

How to tell this is done

• One PR against main
• Touched kernel tests pass, especially orch_window_force
• Stop / end_process SIGKILLs isolation_spawn_os children without a new prompt
• One hung wasm_invoke does not block later mediate/charge/court