…e checking exit

Three bugs found from log analysis:
- Claude CLI p_comm is its version string (e.g. "2.1.50"), not "claude", so direct
  p_comm matching failed; fix: run KERN_PROCARGS2 scan for any unrecognized p_comm
  and check exec_path (full binary path) in addition to argv[1]
- Gemini argv[1] doesn't contain "gemini" in all installs; exec_path check covers it
- drainOutputForMs exited immediately after SIGTERM because getForegroundPgrp saw
  the shell as foreground before we polled for output; fix: remove the early fg_pgrp
  check so we drain until waitpid, POLLHUP, or the 1.5 s deadline

…t-specific prefix

Searching for the last RFC 4122 UUID in the drained PTY output is simpler and more
robust than matching per-agent resume command prefixes. It works regardless of how
an agent formats its exit message, and doesn't need updating when agents change their
CLI syntax. The last UUID in the output is taken, so earlier UUIDs in scrollback
don't shadow the session ID printed just before exit.

…t leak

Issue: After SIGTERM teardown, codex detection failed silently and no UUID
was found in terminal output, with no visibility into why. WaitContext
allocations also leaked on quit because the xev loop stops before callbacks
fire, leaving the completion.state() != .dead check always skipping the free.

Solution: Added debug logging in parseArgv to show exec_path and argv[1]
so codex's actual install path is visible in logs. Added a 300-char tail
log of the terminal text after drain so we can verify whether the UUID is
present at all. Fixed deinit to free WaitContext unconditionally: deinit
runs after the event loop stops, so processExitCallback will never fire
for pending completions and the old completion.state() guard caused a
guaranteed leak for every session reaped via drainOutputForMs.

…ture

Issue: Agents never printed their session UUID at quit because SIGTERM
causes an immediate unclean exit. Claude only outputs "To resume, run:
claude --resume <uuid>" when it processes the /exit command.

Solution: Write Ctrl+C + agent-specific exit command to the PTY master fd
before draining. Ctrl+C (0x03) interrupts any in-progress generation, then
the exit command (/exit for claude and gemini, q for codex) triggers a
graceful shutdown that includes the resume UUID. Drain window extended to
2s to give the agent time to print the UUID and terminate. If the agent
doesn't exit within the drain window, SIGTERM is sent as a fallback with
an additional 500ms drain. Also added argc/raw-bytes logging in parseArgv
to diagnose why codex's KERN_PROCARGS2 exec_path is empty.

…erlay (#237)