feat(agent): chat panel with streaming + tool dispatch

[appergb]

Summary

Implements the in-app Agent chat panel (HANDOFF §3.3, P1). The UI, external MCP server, and in-app chat are now peer clients of the same Dispatcher — a chat turn's tool calls land on the exact same 44-tool EditCommand pipeline an external agent would use, preserving the single-edit-entry invariant.

Changes

Backend — crates/opentake-agent/src/chat/ (new)

• session.rs — ChatSession / ChatMessage / Role / ToolCall (serde camelCase for the front-end mirror).
• llm.rs — BYOK streaming chat client over opentake_gen::KeyringStore. Two providers: OpenAI (/v1/chat/completions SSE) + Anthropic (/v1/messages SSE). Picks the first with a stored key; returns a guided message when neither is configured.
• loop.rs — the turn loop: assembles the system prompt (base + active plugin via prompt::assemble) + a live Context Signal block (signal::engine::build_signal) so the model can act without a get_timeline round-trip; streams the model; dispatches tool calls through the shared Dispatcher via spawn_blocking; re-feeds results until a final text turn. Cancel is a shared Arc<AtomicBool> (no tokio dep, so the --no-default-features build stays clean).

Tauri — src-tauri/src/chat.rs (new) + lib.rs

• chat_send(session_id, text) — spawns a turn; streams chat_delta / chat_tool_call / chat_done events.
• chat_history(session_id) — returns the persisted message log.
• chat_cancel(session_id) — sets the cancel flag at the next boundary.
• ChatState managed in setup: builds its own Dispatcher over the live AppCore + the workflow registry from the same <app_data_dir>/workflows the MCP server scans. No media bridge (chat doesn't render/import).

Frontend

• web/src/lib/types.ts — ChatRole / ChatMessage / ChatToolCall (mirror Rust serde camelCase).
• web/src/lib/api.ts — chatSend / chatHistory / chatCancel + onChatDelta / onChatToolCall / onChatDone listeners.
• web/src/store/chatStore.ts (new) — Zustand store: message list + streaming state + tool-call card upsert by id.
• web/src/components/agent/AgentPanel.tsx — rewritten: message list (user right / assistant left), streaming delta accumulation, collapsible tool-call cards (args + result + error badge), input (Enter send / Shift+Enter newline), cancel button, no-key guided message with an Open-Settings affordance.
• web/src/i18n/dict.ts — agent.* keys (zh + en).

Verification (all 6 green)

cargo fmt --all -- --check                                 ✅
cargo clippy --workspace --all-targets -- -D warnings      ✅
cargo clippy -p opentake-tauri --no-default-features --all-targets -- -D warnings  ✅
cargo test --workspace                                     ✅ (1428 tests, 0 failed)
pnpm -C web build                                          ✅
pnpm -C web test                                           ✅ (437 tests)

New unit tests: 18 in opentake-agent::chat (session serde, OpenAI/Anthropic body shape, tool-call round-trip, no-key path, system-prompt signal injection) + 2 in opentake-tauri::chat (registry + session take/put).

Acceptance

• ✅ "把这段静音剪掉" routes through tighten_silences → ripple_delete_ranges (same Dispatcher as MCP).
• ✅ No-key path returns a guided assistant message with an Open-Settings button.
• ✅ Cancel button sets the flag; the loop stops at the next round boundary.
• ✅ Streaming renders incrementally via chat_delta events.

Notes

• The chat Dispatcher is built without a MediaBridge (no GPU/import from chat); inspect_timeline / import_media return "not available" — the model routes around them. A follow-up can share the MCP server's bridge if chat-side rendering becomes important.
• LLM client is transport-thin (reqwest + SSE); opentake-gen is a generation-job client (no chat surface), so chat ships its own streaming client over the same KeyringStore.