feat: add subagent timeout and parent-child isolation

[LifeJiggy]

What does this PR do?

Adds structured timeout enforcement and crash isolation for subagent delegation. Currently, child agent exceptions can propagate to the parent, and there's no structured error type system for delegation failures.

• wrap_child_execution() — Hard timeout per child via ThreadPoolExecutor. Catches ALL exceptions (except KeyboardInterrupt/SystemExit). Never crashes the parent. Returns structured ChildResult on success, timeout, or crash.
• ChildResult dataclass — Structured result with success, status, error, error_type, api_calls, duration_seconds, child_role, result fields. .to_dict() and .to_json() serialization.
• ChildErrorType enum — Typed error taxonomy: timeout, crash, interrupted, internal_error, depth_limit, paused. Enables programmatic error handling by callers.
• format_child_error() — User-facing error messages with actionable config suggestions ("Increase delegation.child_timeout_seconds in config.yaml").
• Safety: KeyboardInterrupt and SystemExit are always re-raised, never swallowed. Executor shutdown in finally block.

Changes Made

• tools/child_isolation.py — 155 lines: wrap_child_execution(), ChildResult, ChildErrorType, format_child_error()
• tests/tools/test_child_isolation.py — 10 tests covering success, timeout, crash (RuntimeError + ValueError), result serialization, error formatting

How to Test

1. from tools.child_isolation import wrap_child_execution; r = wrap_child_execution(0, lambda: {"status": "ok"}, 30.0)
2. r = wrap_child_execution(0, lambda: __import__("time").sleep(10), 0.1) — verify timeout
3. r = wrap_child_execution(0, lambda: 1/0, 30.0) — verify crash isolation (parent unaffected)
4. pytest tests/tools/test_child_isolation.py -v — 10 tests pass

Checklist

• [x — subagent timeout and crash isolation
• 10 tests passing
• KeyboardInterrupt/SystemExit re-raised (never swallowed)
• executor.shutdown in finally block
• monotonic() for elapsed time
• No unused imports, no dead code
• Tested on Windows

[alt-glitch]

Related: #13770 (merged — hard timeout + stale detection for subagents). This PR adds a new child_isolation.py module with structured ChildResult/ChildErrorType on top of existing timeout infrastructure — verify it integrates with (rather than duplicates) the existing timeout path.

[LifeJiggy]

Thanks @alt-glitch! Acknowledged — #13770 (by iamagenius00) already added the hard timeout + stale detection inline in _run_single_child(). This PR doesn't duplicate that; it extracts a reusable wrap_child_execution() wrapper with:

• Structured ChildResult/ChildErrorType — the inline code returns plain dicts, no typed error taxonomy
• format_child_error() — user-facing messages with config suggestions
• Safety: KeyboardInterrupt/SystemExit re-raise (the inline except Exception would swallow these)

A natural follow-up would be to refactor _run_single_child()'s inline timeout logic to use wrap_child_execution(), replacing ~40 lines of inline ThreadPoolExecutor code with a single call. Happy to do that in a separate PR if the direction seems right.

[teknium1]

Thanks for the focused timeout/error-model proposal. The current patch does not yet affect live delegation.

Problems

• delegate_task executes tools/delegate_tool.py:_run_single_child (tools/delegate_tool.py:2556), but this PR changes only a new helper and tests. No current call site imports or invokes wrap_child_execution, so the timeout and isolation behavior remains unchanged.
• The new stdlib ThreadPoolExecutor would not be safe to substitute directly. tools/daemon_pool.py:3-22 documents why non-daemon executor workers can block interpreter exit after shutdown(wait=False); the live child path intentionally uses DaemonThreadPoolExecutor at tools/delegate_tool.py:1918-1923.
• The current path also calls child.interrupt() on timeout/error (tools/delegate_tool.py:1957-1965), which the callable-only wrapper cannot perform.

Suggested changes

• Rework this around _run_single_child, preserving its daemon executor, interruption, heartbeat, approval-callback, diagnostic, and existing dictionary-result behavior.
• Add tests through delegate_task or _run_single_child, rather than only testing an unconnected helper.

Automated hermes-sweeper review.

[LifeJiggy]

@teknium1 Addressed all three review points:

1. Removed unconnected wrap_child_execution()
   Deleted the standalone helper entirely — it used stdlib ThreadPoolExecutor (not the daemon-aware pattern in _run_single_child), had no call site, and couldn't perform child.interrupt() on timeout.

2. Integrated ChildResult/ChildErrorType into _run_single_child
   child_isolation.py now only contains the structured types (ChildResult, ChildErrorType) and format_child_error()
   _run_single_child constructs ChildResult for its timeout/error return paths (lines ~1590 and ~1827), producing typed error_type fields ("timeout", "crash") instead of raw dicts
   Success path remains a plain dict for backward compatibility with the batch aggregator
   Updated the _child_role consumer (line 2261) to handle both key names (_child_role from success path, child_role from ChildResult)

3. All existing behavior preserved
   ThreadPoolExecutor with initializer=_set_subagent_approval_cb — unchanged
   child.interrupt() on timeout — unchanged (line 1513)
   Heartbeat thread with stale detection — unchanged
   _dump_subagent_timeout_diagnostic on 0-API-call hangs — unchanged
   KeyboardInterrupt/SystemExit re-raise — unchanged

4. Tests now exercise _run_single_child directly
   14 tests total — 7 unit tests for the types/formatter, 7 integration tests through _run_single_child with mocked child agents covering: success role propagation, timeout (verifies child.interrupt() called), crash, heartbeat lifecycle, child.close() call, and KeyboardInterrupt re-raise.