fix(hooks): keep a strong reference to fire-and-forget hook triggers

[LHMQ878]

Fixes #2564.

Problem

asyncio holds running tasks in a WeakSet. Writing

_hook_task = asyncio.create_task(self._hook_engine.trigger(...))
_hook_task.add_done_callback(lambda t: t.exception() if not t.cancelled() else None)
return ToolResult(...)   # <- _hook_task goes out of scope here

lets the GC collect the still-pending task, so the hook subprocess is never
awaited and the hook silently does not run. It fires non-deterministically,
depending on when a collection happens.

HookEngine.fire_and_forget_trigger() already exists for exactly this pattern —
its docstring describes this bug — and keeps the task in
_pending_fire_and_forget until it completes. Only SubagentStop had been
migrated to it.

What changed

Five call sites still rolled their own create_task:

file	hook events
soul/toolset.py	PostToolUse, PostToolUseFailure
soul/kimisoul.py	StopFailure, Notification, PostCompact

#2564 reports the two in toolset.py. The three in kimisoul.py are the same
defect, so they are fixed here too.

Two side benefits of using the helper: the task is discarded from the pending set
on completion (no leak), and a failing task is reported through
logger.warning("Fire-and-forget hook task failed") instead of having its
exception fetched purely for the side effect and dropped.

SessionEnd in cli/__init__.py:697 is deliberately left as-is — it runs under
await asyncio.wait_for(..., timeout=5), so it is not fire-and-forget.

Tests

fire_and_forget_trigger had no test coverage at all, which is plausibly how
the migration got missed. New tests/hooks/test_fire_and_forget.py:

• the task survives a gc.collect() when the caller keeps no reference, and its
  command still runs to completion (asserted via a file the hook touches)
• the strong reference is released once the task finishes
• a hook whose command cannot be executed does not leak its reference
• an AST guard over toolset.py and kimisoul.py asserting that no
  create_task call wraps a hook trigger(...) call, so a future fire-and-forget
  site cannot silently reintroduce this

Control experiment — reverting only src/ and rerunning:

FAILED tests/hooks/test_fire_and_forget.py::test_no_bare_create_task_around_hook_triggers[kimi_cli.soul.toolset]
FAILED tests/hooks/test_fire_and_forget.py::test_no_bare_create_task_around_hook_triggers[kimi_cli.soul.kimisoul]
2 failed, 3 passed

With the fix applied: 5 passed.

ruff check and ruff format --check are clean on all three files.

Note on the pre-existing failures

pytest tests/hooks/ tests/core/test_dynamic_injection_hooks.py reports
5 failed, 38 passed on this branch. Those 5 (test_integration.py::test_pre_tool_use_block_flow,
::test_stop_hook_feedback, ::test_user_prompt_submit_block,
test_runner.py::test_exit_2_blocks, ::test_json_deny_decision) fail
identically on unmodified main — they rely on POSIX shell semantics and I ran
this on Windows. Baseline is 5 failed, 33 passed; the only difference is my 5
new tests passing.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.