Task: Fix Code fresh-conversation voice input

Identity

Enable AI Programming voice capture after the user selects an Agent but before the lazy first OpenCode session exists.
Add a component-level regression that exercises the real fresh-contact state and verifies transcription populates the composer draft without creating an empty session.
Preserve existing-session voice behavior and the current busy/transcribing guards.

Preserve lazy contact selection: choosing an unused Agent or starting voice capture must not create or start an OpenCode session; only message submission may create the first session.
Keep speech capture and transcription on the existing Renderer -> Host API -> Electron Main -> Works Square path.
Do not change Canvas, Learning, Agent templates, provider configuration, or unrelated composer behavior.
Keep the implementation surgical and limited to the ChatPanel predicate and its focused unit coverage.

Updated the AI Programming composer voice-availability predicate so a selected Agent is sufficient before the lazy first session exists.
Kept the existing loading, transcription, recording, and session-busy guards unchanged.
Added a component regression covering a fresh Agent conversation: voice capture transcribes into the draft while the selected session remains lazy and no session-creation request is sent.

Regression-first evidence: the new fresh-Agent test failed before the predicate change because the voice button was disabled, then passed after the one-line fix.
pnpm exec eslint src/pages/Chat/OpencodeChatPanel.tsx tests/unit/opencode-chat-panel.test.tsx passed.
Five focused lazy-session and voice-input tests passed (5 passed, 80 skipped).
The complete opencode-chat-panel.test.tsx file passed (85/85) during implementation.
pnpm test passed (175 files, 2048 tests).
pnpm run typecheck passed.
pnpm run build:vite passed for Renderer, Electron Main, and Preload (3921, 133, and 1 modules transformed respectively); only pre-existing chunk-size and mixed static/dynamic import warnings were reported.
Final read-only Sol review: PASS, with no blocking findings across spec and standards checks.
Electron microphone E2E was not added because the shared fixture does not provide a deterministic MediaRecorder/microphone seam; the component regression exercises Agent selection, recording start/stop, transcription, and lazy-session behavior directly.

None. The change restores the documented lazy-session behavior and does not alter an architectural or product contract.