feat: human-readable action logs for GUI harnesses

[stephenleo]

Summary

Running openmicro codex-app, every button press logged the raw osascript payload ("key down control / key code 46") — gibberish to an end user. Logs now show the physical button pressed (named per pad family) and the action it executes:

● triangle → push-to-talk
● right stick flick right → thinking depth up
● R3 → send ctrl+shift+m
● d-pad up → send up arrow

Changes

• src/labels.ts (new): controlLabel maps logical controls to physical names per controller family (cross/circle/square/triangle for DualSense/DS4, A/B/X/Y for Xbox/GameSir/generic) and actionLabel describes actions, decoding keys bindings (arrows, ctrl chars, CSI-u chords) into readable key names.
• src/router.ts: LayerRouter.lastControl exposes the ControlId matched by the last successful route(), so stick-gesture actions are nameable at the cli.
• src/cli.ts: track the connected pad type, announce each dispatched action via guiStatus, and drop the raw payload dump.

Testing

• New test/labels.test.ts (6 tests) covering per-family face names, stick gestures, and keys decoding.
• Full gate: typecheck, lint, format:check, 261 tests — all pass.

🤖 Generated with Claude Code

[stephenleo]

Follow-up commit 392f82e: fixes the system-wide stuck-modifier hazard (Ctrl stuck → left click acts as right click, keyboard unusable until reboot).

Root cause was not the logging change — synthetic key down events are machine-wide, and three paths could leave them unbalanced:

1. push_to_talk holds Ctrl+Shift+D by design; process exit mid-hold stranded it
2. concurrent osascript chords could interleave their down/up steps
3. a failed/hung osascript left key down without its key up, with no recovery

Fix: osascript runs are serialized through a queue with a 5s timeout; any failed key-down chord triggers a synchronous release of every key the harness can hold; and Harness.dispose() releases all keys on Ctrl+C/SIGTERM. 4 new tests.