fix(cli): restore terminal from cbreak mode after run exits

[hertznsk]

Summary

Fixes #290 — Conductor could leave the invoking terminal in cbreak mode (no echo / no line editing) after a workflow exited, requiring a blind reset or stty sane to recover.

Root cause

KeyboardListener.start() captured the tty settings on every call. Under specific interleavings — a second listener instance in the same process, or a second start() on an already-active listener — the captured "original" settings were the already-broken cbreak state, which stop() then faithfully restored.

Changes

Two atomic commits, each leaving the suite fully green:

Commit 1 — prevent cbreak capture (e5205b5)

• Capture the tty baseline once per process into a module-level cache before the first tty.setcbreak(); every later listener instance reuses it instead of re-snapshotting the live tty.
• Make start() idempotent while the listener is truly active (baseline held AND reader thread running), so a duplicate call can't overwrite the baseline or spawn duplicate threads. Start-after-suspend() still restarts correctly.
• Restore with TCSANOW instead of TCSADRAIN so a blocked output drain can't delay the restore.
• Clear the saved baseline only after a successful tcsetattr, so a transient restore failure can be retried by atexit/SIGTERM/stop().

Commit 2 — SIGTERM correctness (8ebd92f)

• The SIGTERM cleanup handler restored the terminal but then swallowed the signal, leaving the process alive instead of terminating. It now delegates to the previously-installed disposition after restoring: SIG_DFL resets and re-raises via os.kill, SIG_IGN stays ignored, a callable previous handler is invoked.
• The previous disposition is captured into the handler closure at registration time (not the mutable instance field), and each instance stores its own handler, skipping re-registration only when that exact handler is still installed — so a stopped listener's stale closure never blocks a new listener from protecting its terminal, and a handler can never capture itself as "previous" (infinite recursion).

Tests

• New real-pty integration tests (tests/test_interrupt/test_listener_pty.py) reproduce both corruption scenarios on a genuine pseudo-terminal: two listener instances in one process, and double-start() on one instance. Both fail on the old code and pass on the new. Gated skipif(win32); CI runs them on ubuntu-latest.
• New mocked unit tests for the baseline cache, idempotent start, TCSANOW restore paths, restore retry semantics, and all SIGTERM delegation cases (DFL re-raise, IGN no-op, callable delegation, self-recursion guard, cross-instance registration).
• Full suite: 3670 passed, 13 skipped (excl. integration/perf/real_api/install_scripts); make lint and make typecheck clean.
• End-to-end pty smoke: conductor run examples/wait-smoke.yaml on a real pty leaves ICANON|ECHO intact.

Scope

Only src/conductor/interrupt/listener.py plus its tests. No CLI/engine/web/dashboard changes, no new dependencies, no new CLI flags, no behavior change for suspend()/resume() semantics. Windows is unaffected (the listener is already Unix-only).

Note for the release-prep PR: this change intentionally does not add a CHANGELOG entry — please backfill it under [Unreleased] at release time.