fix(audio-input): open-source returns the actually-bound device (read-back)

[byrongamatos]

What & why

The companion core half of the macOS wrong-mic fix. The desktop half (stable name-based input identity + fail-loud open, so picking BlackHole can't silently fall back to the internal mic) is already on `main` (src/renderer/screen.js). This adds the matching core contract so a caller can see which device actually bound.

openInputSource() (static/capabilities/audio-session.js) now surfaces the provider's bound device on the command's return value as payload.bound = { type, name }, for the trusted in-process caller — enabling an honest "Now listening to: " readout and detection of a silent substitution.

How

• A source.open handler may return payload: { boundType, boundName } (the merged desktop handler already does). openInputSource threads it onto the _handled(...) return only.
• PII-safe: the raw device name is kept out of the emitted source-opened event and the diagnostics snapshot (both stay redacted) — mirroring how list-sources returns the device label verbatim to the UI but pseudonymizes it in diagnostics.
• Purely additive — the open-session summary shape is unchanged, and bound is omitted when the provider reports nothing.

Scope note

This PR is read-back contract only — no input-setup meter. The live "we can hear you" meter that would consume bound is intentionally not included here: it surfaced an unrelated audio-quality regression on re-enable during tester runs and is being reworked separately. So bound currently has no in-tree consumer yet; it's groundwork + makes the desktop handler's read-back reachable.

Verification

• node --test tests/js/audio_session_input.test.js → 39/39, incl. a new test asserting the read-back reaches the caller but is absent from the event + snapshot.
• (Org Actions billing is blocked, so CI checks won't run — verified locally.)

Notes

• Please do not merge without byron's sign-off.

🤖 Generated with Claude Code