fix(droid-control): kill full true-input process tree on close

[factory-ain3sh]

Description

tctl close on a recorded true-input session killed only the wf-recorder and deleted the session state dir, leaving the entire cage/ghostty/script/CLI process tree running orphaned; one dev host had accumulated 9 stale compositors and 48 leaked temp workspace dirs. Root cause: launch_true_input persisted session meta with CAGE_PID="" before starting cage, and for --record launches (the apps/cli true-input harness always records) start_true_input_recording's load_meta clobbered the in-memory PID back to empty, so cmd_close skipped the compositor kill entirely. Restoring the PID alone would not fix teardown either: script(1) moves its children into new sessions, so even a correct kill $CAGE_PID left ghostty/script/CLI behind.

This PR persists CAGE_PID immediately after the wayland socket wait (before the recording branch can reload the meta), launches cage under setsid, and replaces the bare kill with a full-stack teardown: descendant enumeration plus process-group kill with TERM->poll->KILL escalation, INT->poll->KILL for the recorder so recordings finalize, and a stray sweep on the session's run-*.sh argv that also closes pre-fix sessions whose metas carry an empty CAGE_PID.

Related Issue

Closes CLI-1171

Reviewer Guide

Read order: plugins/droid-control/bin/tctl only -- launch_true_input (PID persistence, setsid, socket-timeout cleanup) > terminate_true_input_stack + collect_tree_pids > terminate_session_strays + pid_is_self_or_ancestor > cmd_close.
Review depth: Standard -- signal-sending code with a self-kill guard deserves a careful read, but it is one file with no external contract.
Open for pushback: the fixed grace budgets (2s tree, 3s recorder, 1s per stray) mean close can block a few seconds worst case; chosen over a tunable to keep the CLI surface unchanged.

Risk & Impact

The new paths send signals, so mis-scoping would kill unrelated processes. Containment: the group kill targets only cage's own setsid group; enumeration walks descendants of the recorded cage PID; the stray sweep matches only the session's run-*.sh runner argv (not the bare session dir path, which an inspecting shell or editor can carry) and skips self/ancestors, so a close invoked from inside a session cannot kill its caller. close now blocks up to the escalation deadlines instead of returning instantly.

Verification

Behavior verified. verified @ adff9fa on a host with cage/ghostty/wtype/grim/wf-recorder: recorded launch -> type -> close leaves zero processes referencing the session, session and runtime dirs removed, mp4 finalized (ffprobe duration 1.83s); a session whose meta was hand-edited to the pre-fix CAGE_PID="" shape (5 live processes) sweeps clean on close; tuistory launch/snapshot/close unaffected. Standalone record start/stop and the self/ancestor guard (a close whose caller argv contains the runner path survives) were verified on the same host during fix development.
Regression coverage. None automated: the repo has no shell test harness (CI's check-skills.yml validates skill metadata only). The manual matrix above plus the Repro Recipe stand in.
Not tested. The KILL-escalation branch under a process that traps and ignores TERM -- genuinely contrived to set up; the branch only runs when the grace poll expires.
Standard validators. bash -n clean; shellcheck not installed on the verification host; no formatter/linter configured in this repo.

Repro Recipe

On master, a recorded session leaks its whole tree; on this branch the same steps leave nothing:

tctl launch bash --backend true-input -s leak-test --record /tmp/leak-test.mp4
tctl -s leak-test close
pgrep -fa 'tctl-sessions/leak-test'   # master: cage/ghostty/script/bash survive; this PR: empty
ffprobe -v error -show_entries format=duration -of csv=p=0 /tmp/leak-test.mp4  # this PR: valid duration (recorder got INT, container finalized)