feat(agent-manager): add WezTerm terminal support

[codeaholicguy]

Summary

Adds support for discovering, focusing, and sending input to agents running inside WezTerm, alongside the existing tmux / iTerm2 / Terminal.app flows.

• TerminalFocusManager: new WEZTERM terminal type. Discovers the owning WezTerm pane from wezterm cli list --format json by matching the agent PID's TTY → pane id; focuses via wezterm cli focus-pane --pane-id. Probed after tmux, before the macOS AppleScript probes, so non-macOS hosts short-circuit cleanly (tmux-inside-WezTerm still resolves to tmux).
• TtyWriter: sends text via wezterm cli send-text --pane-id over the child stdin (text, then Enter, 150 ms apart) — the same two-step convention used for the other emulators so a bracketed-paste-aware TUI still sees Enter as a submit.
• CLI-only integration (no AppleScript) → cross-platform (macOS/Linux/Windows). Missing or non-running wezterm degrades like the tmux probe (null / false), so non-WezTerm users are unaffected.
• No CLI changes: agent open, agent send, and channel bridges pick up WezTerm automatically through the existing findTerminal / focusTerminal / TtyWriter.send abstract API.

Out of scope (deliberate): spawning new agents into WezTerm (that remains TmuxManager's responsibility).

[codeaholicguy]

Updated per review feedback. New HEAD: `38ffcda` (force-pushed, single amended commit).

Changes addressing the three review notes:

1. tty_name field — WezTerm's `cli list --format json` exposes the TTY as `tty_name`, not `tty`. Fixed `WeztermPaneEntry` and the match in `findWeztermPane` to read `pane.tty_name`. Tests/fixtures updated.

2. Focus command — `focusWeztermPane` now uses `wezterm cli activate-pane --pane-id ` instead of `focus-pane`. Tests updated.

3. `agent open --debug` — added a `--debug` option to the `open` command. `TerminalFocusManager` now accepts an optional `debug` logger; when `--debug` is set, it's wired to the existing `ai-devkit:terminal` debug logger (`createLogger('terminal')` + `enableDebug()`, same convention as `agent start`/`channel`). It emits one line per discovery/focus step — pid/tty, each emulator probe's match/no-match (tmux → wezterm → iTerm2 → Terminal.app → UNKNOWN), and the focus result — so the matching/focus decision path is inspectable.

Validation (fresh):

• `agent-manager` + `cli` `tsc --noEmit` clean; `eslint` 0 errors.
• vitest: agent-manager 442, cli 843, monorepo 1079 total via the pre-commit hook (`nx run-many -t lint test`).
• New WezTerm + debug branches covered (TtyWriter 100%; new `findTerminal`/`focusTerminal` debug paths + `tty_name`/ `activate-pane` all hit). Pre-existing uncovered iTerm2/Terminal.app focus methods left untouched to avoid unrelated refactors.
• `npx ai-devkit@latest lint --feature wezterm-support` passes.

Not merging.

[codeaholicguy]

Updated per the WezTerm send-behavior feedback. New HEAD: `1e3fc77` (force-pushed, single amended commit).

`TtyWriter.sendViaWezterm` now makes two explicit CLI calls (no stdin):

1. Text — `wezterm cli send-text --pane-id <pane_id> `, with the message passed as a positional argv element via `execFile` (no shell). Because `execFile` spawns the process directly, the bytes are delivered verbatim regardless of shell metacharacters — there is no injection surface (same approach as tmux's `send-keys -l`).
2. Enter — `wezterm cli send-text --pane-id <pane_id> --no-paste `, where the carriage-return byte is `0x0d` (shell `$'\x0d'`), passed as a discrete argv element. `--no-paste` delivers the CR literally rather than wrapped in paste brackets. (150 ms pacing between the two calls, matching the other senders.)

The user message is never built into a shell command or a shell string.

Code: removed the previous stdin-based `ExecFileInputOptions` workaround entirely (net simplification); added a `CARRIAGE_RETURN = '\x0d'` constant.

Tests added/updated (`TtyWriter.test.ts`):

• two-call contract: message as last argv element of call 1, then `--no-paste` + CR as call 2, no options/stdin arg;
• Enter byte is exactly char code `0x0d` (not `0x0a`);
• shell-injection safety: a hostile message (`$(rm -rf /)`, backticks, `;`, newlines) is passed verbatim as one argv element.

Validation (fresh): `agent-manager` + `cli` `tsc --noEmit` clean; `eslint` 0 errors; pre-commit hook `nx run-many -t lint test` green — agent-manager 444, cli 843. Docs (design/implementation/testing) updated to describe the two-call, argv, CR-byte send. Not merging.

[codeaholicguy]

Notation corrected per review. New HEAD: `f8e8b65` (force-pushed, single amended commit).

Fix: wherever the shell form of the Enter command is written, it now uses dollar-single-quoted CR — `$'\x0d'` with the leading `$` — i.e.

```
wezterm cli send-text --pane-id <pane_id> --no-paste $'\x0d'
```

Implementation is unchanged and still correct: `execFile` passes the actual CR byte as an argv element (JS `'\x0d'` = char `0x0d`), as the review confirmed it should. The distinction is now explicit everywhere:

• argv value (implementation/test): the CR byte, JS `'\x0d'` — labelled "JS char / argv element" in comments.
• shell command (docs/comments): `$'\x0d'` (ANSI-C quoting) — bare `'\x0d'` would be passed by a shell as the four literal characters `\ x 0 d`.

Corrected in: `design`, `implementation`, `testing` docs, and the `TtyWriter.ts` + `TtyWriter.test.ts` comments. A grep confirms every shell-command occurrence uses `$'\x0d'`; remaining bare `'\x0d'` are only JS-source argv values (the constant `CARRIAGE_RETURN = '\x0d'`, the argv assertion, and prose explicitly labelled "JS char").

(I considered adding a unit test that asserts the shell-form string, but it would only check a constant defined in the test itself — a tautology. The behavioral test already pins the real contract: the Enter argv element is exactly char code `0x0d`, which is what `$'\x0d'` denotes.)

Validation (fresh): pre-commit hook `nx run-many -t lint test` green — agent-manager 444, cli 843; `tsc --noEmit` clean; `eslint` 0 errors; `lint --feature wezterm-support` passes. Not merging.