Chat mode v1 — follow-ups

[kwannoel]

Tracking follow-ups surfaced during the chat-mode v1 implementation on branch session-1-07ce2e (design: docs/plans/2026-04-22-chat-mode-design.md, plan: docs/plans/2026-04-22-chat-mode-plan.md).

Blockers for full e2e coverage

• Expose read_daemon_token on the Axum test server (src-tauri/src/bin/server.rs). The browser-mode Playwright harness uses this server; without a /api/read_daemon_token route, the daemon-reachable test in e2e/specs/chat-mode.spec.ts cannot pass and is currently marked test.fixme. Removing the fixme and extending the spec with a "+ New chat" + streaming-reply scenario is the payoff.
• After above lands, implement the third e2e scenario — create a chat via the "+ New chat" dialog, send a user message, assert user bubble + streamed agent bubble appear.

Deferred UX / error handling

• 401 re-read-token + retry-once in the daemon error classifier (src/lib/daemon/errors.ts / ChatView / ChatInput). Today 401 shows a toast; the design doc calls for re-reading ~/.the-controller/daemon.token and retrying the request once before giving up.
• Make "+ New chat" rows keyboard-walkable. HotkeyManager.getVisibleItems' chat branch intentionally excludes the "+ New chat" rows from j/k navigation. Enter on a walked "+ New chat" row should open NewChatDialog.
• Dedicated Chat help section in HotkeyHelp. The mode label already switches to "Chat"; a full section (send / interrupt / navigate) requires registering chat-specific commands in the keymap (src/lib/commands.ts + wherever getHelpSections lives). Today users only see the global shortcuts in chat mode.

Pre-existing svelte-check errors

Not introduced by this work but worth clearing to keep the suite clean:

• src/lib/IssuesModal.svelte:297 — MouseEventHandler type mismatch (pre-existing on main).
• src/lib/daemon/client.test.ts:26 — two tuple-index errors (const [url, init] = fetchMock.mock.calls[0]). Introduced by Task 6 of the chat plan. Doesn't fail tests; just svelte-check noise.

Quality polish (non-blocking)

• Error toast on failed tool-approval submit. ToolApprovalBlock currently try/finallys the send — errors are silently swallowed. A toast via classifyError + showToast would match how ChatInput handles failures.
• Trim whitespace in initial_prompt / tool approval reason. Current check is empty-string only; " " still gets sent.
• Prune orphaned transcripts in loadSessions. Today it only initializes emptyTranscript() for new sessions; sessions removed on the daemon don't have their transcripts GC'd.
• Virtualize Transcript.svelte once transcripts grow large. Plain scroll for v1 per the design doc.

Architecture follow-ups

• Auto-start the daemon from the Tauri app (the "like a language server" option from the brainstorm). v1 is manual; auto-start requires process supervision, bundling, and upgrade plumbing.
• Tracking issue for the fixme test — once /api/read_daemon_token is exposed, un-fixme in e2e/specs/chat-mode.spec.ts and tune the 1500ms daemon-readiness sleep (likely flaky on slow hardware).