fix(cli): print exit summary before cleanup so watchdog can't swallow it

[adurham]

What does this PR do?

Fixes a bug where the CLI's exit summary (cost report + --resume <id>
hint) can be silently swallowed if cleanup takes long enough to trip the
exit watchdog.

Root Cause

_run_cleanup() was called BEFORE self._print_exit_summary() on both
interactive-exit paths in cli.py's run(). _run_cleanup() includes a
memory provider's on_session_end hook (can be a network-bound call for
external memory providers) and shutdown_mcp_servers() (can block during
MCP server teardown), both guarded by a 30s exit watchdog
(_arm_exit_watchdog) that calls os._exit(0) if cleanup hasn't returned
in time. If the watchdog fires while still inside _run_cleanup(), the
process is killed before any code written AFTER _run_cleanup() ever
runs — silently swallowing the cost report and resume hint with zero
user-visible error.

Changes Made

• Extracted a shared HermesCLI._finish_interactive_exit(*, release_session=False)
  helper (print summary → cleanup → optionally release the session) and
  wired both interactive-exit call sites in run() to it — the
  stdin-unavailable early return and the main finally-block exit path —
  so the ordering can't drift out of sync between the two sites again.
• Bumped the exit watchdog's default timeout from 30s to 60s for
  headroom against realistic worst-case cleanup time.
• Fixed a second, separately-hardcoded 30s default in
  _arm_exit_watchdog_on_shutdown_signal (the SIGTERM/SIGHUP-armed
  backstop) that would otherwise leave its "2x headroom over the normal
  cleanup timeout" computed from a stale base — caught by the new
  regression test's explicit assertion on the doubled value (120s, not
  60s).
• Replaced a pre-existing regression test that read cli.py's raw source
  text and pattern-matched call order via regex with a real behavioral
  test exercising _finish_interactive_exit() directly (mocking
  _print_exit_summary/_run_cleanup/_release_active_session and
  asserting invocation order), matching the established pattern already
  used in test_exit_watchdog_signal_arm.py in the same directory.

Related Issue

No existing issue found (searched gh search issues for "exit watchdog
fired cost report missing", "resume hint swallowed exit", "cli exit cost
summary missing" — no hits).

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

How to Test

1. Set HERMES_EXIT_WATCHDOG_S=1 and configure a memory provider whose
   on_session_end hook takes longer than 1s (or just observe the
   ordering change directly).
2. Before this fix: a slow-enough cleanup step could trip the watchdog
   before the cost report ever printed.
3. After this fix: the cost report and resume hint always print first,
   regardless of how long cleanup takes afterward.
4. pytest tests/cli/test_exit_summary_before_cleanup_ordering.py tests/cli/test_exit_watchdog_signal_arm.py -q — 10 passed.
5. Also ran adjacent exit/cleanup test files to check for regressions:
   test_cli_active_agent_ref_wiring.py, test_cli_new_session.py,
   test_cli_shutdown_memory_messages.py, test_session_boundary_hooks.py,
   test_single_query_session_finalize.py,
   test_tui_terminal_reset_on_exit.py — 51 passed, 0 failed.

Checklist

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs/issues (none found)
• My PR contains only changes related to this fix (the
  _finish_interactive_exit extraction is coupled to this fix
  because it's what makes the ordering invariant testable without a
  source-regex test — flagging this explicitly since it's a small
  refactor riding along with the bugfix, not unrelated scope)
• I've run the relevant test suite and all tests pass
• I've added tests for my changes
• Tested on macOS

[teknium1]

Thanks for the focused exit-path fix. The premise remains valid on current main: both interactive paths invoke cleanup before printing the summary (cli.py:17157-17158, cli.py:17318-17320), while cleanup arms a watchdog (cli.py:1169-1172) whose timer can terminate the process with os._exit(0) (cli.py:1103-1108).

The shared helper in PR head 6c46afa0b3dd changes both affected paths and its behavioral tests cover summary-before-cleanup, optional lease release, and print failures. The single-query sibling already prints its summary before final cleanup (cli.py:17936-17938), so no matching unaddressed path was found.

Automated hermes-sweeper review.