v0.3.3 — run_code dispatch path + clean thinking placement

v0.3.3 — ride the run_code dispatch path; stop mid-text thinking chips

Fixes the two issues reported on 0.3.2 in the live GUI:

• unknown tool "agy_tool" on dispatch: this DSH deployment only lets a
  model call run_code directly; every other tool must be invoked from
  inside a run_code program. Span cuts now address run_code with a
  generated one-line program that calls the registered agy_tool mirror —
  the inner dispatch replays the recording and renders the native tool
  card (terminal / diff / read / search). Cursor-only (run, step)
  invocations are first-class: the mirror schema makes tool/input optional
  and presenters enrich them from the recording.
• thinking chip wedged mid-text: agy streams answers in fragments and
  attaches thinking_tokens to the DONE tail; v0.3.2 annotated at that
  point, splitting the reply. The annotation now fires only while the step
  has not yet streamed text — one-shot answers keep their leading thinking
  line, streamed answers stay clean.

Also: tool cards read agy's PascalCase argument keys (CommandLine,
AbsolutePath, SearchDirectory, WorkingDirectory, …) instead of falling
back to raw JSON titles.