fix(#1200): drain async delegations before cron session teardown

[Lexus2016]

Problem

Closes #1200 — Async delegate_task race in evolution-hydra: parent cron dispatches subagents but does not await completion before read-back, causing missing artifacts.

Root Cause

In cron/scheduler.py _run_job_impl, the inactivity timeout loop (lines 4195-4214) terminates as soon as the agent stops calling tools. The finally: block then calls _cron_pool.shutdown(wait=False, cancel_futures=True) — without waiting for background subagents dispatched via delegate_task(background=true) that are still running on the daemon executor.

This orphans completion events and loses artifacts — the subagents' work vanishes when the cron session tears down before they finish.

Additionally, the inactivity timeout was hardcoded to 600s, which is too short for subagent-heavy cron jobs (the issue requests 900-1200s).

Fix

Three parts:

1. tools/async_delegation.py — drain helper

New wait_for_session_delegations(parent_session_id, deadline_seconds, poll_interval, on_interrupt) function:

• Polls _running_for_session(parent_session_id) every poll_interval
• Waits up to deadline_seconds for delegations matching the session to complete naturally
• Force-interrupts stragglers via interrupt_for_session(parent_session_id=...) when deadline expires
• Returns {waited, completed, interrupted, timed_out}

2. hermes_cli/config.py — config keys

Two new cron config keys with defaults:

• delegation_drain_seconds: 1200 — drain deadline (matches heartbeat stale ceiling of 40×30s)
• inactivity_timeout_seconds: 900 — up from hardcoded 600s

3. cron/scheduler.py — wire the drain

• Inactivity timeout now reads cron.inactivity_timeout_seconds (default 900s) instead of hardcoded 600.0. HERMES_CRON_TIMEOUT env var still overrides.
• finally: block calls wait_for_session_delegations() before _cron_pool.shutdown(). Configurable via cron.delegation_drain_seconds (default 1200s) or HERMES_CRON_DELEGATION_DRAIN_SECONDS env var.
• Drain exceptions are caught and logged at debug — never block teardown.

Tests

tests/tools/test_async_delegation_drain.py (10 tests):

• Empty session → immediate return
• Natural completion before deadline → waited, not interrupted
• Deadline expiry → force-interrupt remaining
• Session scoping → only matches own delegations
• Completed records not counted
• on_interrupt callback invoked with count
• on_interrupt exception swallowed (never propagates)
• Mixed: some complete, some timeout

All 821 existing cron + delegation lifecycle tests still pass.

Verification

python -m pytest tests/tools/test_async_delegation_drain.py -v
# 10 passed in 2.56s

python -m pytest tests/tui_gateway/test_delegation_session_lifecycle.py tests/cron/ -v
# 821 passed in 36.23s

Footprint

No new core tools. The drain helper is a plain function in an existing module. The config keys are additive (deep-merge handles them). The scheduler change adds a drain step to an existing finally: block — no structural change to the loop.