refactor(assistant): split run_session_turn into sibling API and CLI runners

[juacker]

Why

engine::run_session_turn was ~630 lines that ran the whole API turn inline (stream consumption, tool execution, context-limit recovery, persistence) while delegating CLI providers to local_agent. The two runners were not at the same level.

What

Pure refactor, no intended behaviour change.

• engine::run_session_turn (now 37 lines) loads the session and connection once, validates (including a missing agent_workspace_id), resolves the workspace root once, then dispatches with a plain if to api_turn::run_session_turn or local_agent::run_session_turn. No trait. Callers are unchanged.
• New assistant/api_turn.rs holds the API loop, split into prepare_api_request, try_context_limit_recovery, consume_api_stream, finalize_api_message, execute_api_tool_calls and ApiTurnState. The orchestrator is about 130 lines.
• New assistant/turn_common.rs holds the helpers both runners use: trigger marker, empty-content rule, image preservation, and cleanup of unanswered input.

Tests

Characterisation tests, each checked by mutation:

• partial API output persists across a reload;
• tool calls are saved in order and replayed to the next request;
• partial text is kept on cancel or stream error;
• unanswered input is cleaned up while images are kept.

Known gap: there are no runner-level tests for these cases:

• a tool failure being returned to the model;
• forced compaction running only once;
• the final run status;
• UI event order.

The blocker is that AssistantDeps requires AppHandle<Wry>, while Tauri's mock_app gives MockRuntime. A follow-up PR will add a runtime seam.

Review

Two independent review rounds. Round 1 found a double workspace-root lookup, an unused adapter-injection seam, visibility that was too wide, and a stale comment; all were fixed. Round 2 returned production_quality, with a function-by-function mapping against main.

Manual test requested

API-provider chat with tool calls, plus a cancel partway through a reply.