fix(backend/blocks): orchestrator EXTENDED_THINKING emits final-answer only (#13188)

### Why / What / How

**Why:** Discovered while debugging a debate-style orchestrator agent on
dev preview. The two execution modes were producing structurally
different `finished` outputs for the same block:

| Mode | `finished` content | What's wrong with it |
|---|---|---|
| `BUILT_IN` (agent loop via `tool_call_loop`) |
`response.response_text` from the last LLM call (i.e. the model's final
text-only response after it stopped calling tools) | **Correct** — this
is the agent's composed answer |
| `EXTENDED_THINKING` (Claude Agent SDK) | `"".join(every TextBlock from
every AssistantMessage)` | **Buggy** — accumulates every narration
emitted between tool calls; the answer gets buried in working-log noise,
and worse, *masks* prompts that never compose a real answer |

The original-direction commit (`0f4d348d0d`, now force-replaced) tried
to fix the divergence by making BUILT_IN match EXTENDED_THINKING's
accumulation. That was the wrong direction: BUILT_IN's last-answer
behaviour is the correct contract, and the accumulation in
EXTENDED_THINKING was hiding composition bugs that dry-run / autopilot
should be able to detect.

**What:** Replace `response_parts.append(...)` (every TextBlock) with
`final_response_parts = list(text_parts)` only when the assistant
message has **no tool calls** — i.e. the model has stopped calling tools
and is emitting its composed answer. The last such message wins.

**How:**

```diff
- response_parts: list[str] = []
+ final_response_parts: list[str] = []   # only the final-answer message's text
  ...
  if isinstance(sdk_msg, AssistantMessage):
      text_parts = []
      tool_use_parts = []
      for content_block in sdk_msg.content:
          if isinstance(content_block, TextBlock):
              text_parts.append(content_block.text)
-             response_parts.append(content_block.text)
          elif isinstance(content_block, ToolUseBlock):
              ...
+     # Capture the agent's final answer: last text-only message wins.
+     if text_parts and not tool_use_parts:
+         final_response_parts = list(text_parts)
  ...
- response_text = "".join(response_parts)
+ response_text = "".join(final_response_parts)
```

**Edge cases:**

- *Agent never stops calling tools (e.g. hits max iterations)*:
`final_response_parts` stays empty → `finished` = `""`. This is the
**correct** signal for dry-run / autopilot diagnostics: the prompt
didn't compose a final answer, so the agent author needs to fix it.
Previously the over-accumulation returned a transcript-shaped string
that masked this failure mode.
- *Agent emits multiple text-only messages over the course of the run*
(rare): the **last** one wins, matching "the model's most recent
composed answer."
- *Agent emits text + tool calls in the same message*: that message's
text is treated as narration belonging to `conversations`, not the
answer (the `not tool_use_parts` guard).

**Existing tests:** all 14 in `test_orchestrator_execution_mode.py`
still pass. No new SDK-mocking infrastructure exists in the test suite
for this code path, so behavioural verification will happen on the next
dev preview redeploy (combo branch redeployment to follow).

### Changes 🏗️

- `_execute_tools_sdk_mode` captures only the last assistant message's
text-only content into `final_response_parts`.
- `finished` output yields `"".join(final_response_parts)` — empty
string when the agent never composed a final answer (useful diagnostic
signal).
- No change to `conversations` output (still carries the full
back-and-forth).

### Checklist 📋

#### For code changes:
- [x] I have clearly listed my changes in the PR description
- [x] I have made a test plan
- [x] I have tested my changes according to the test plan:
- [x] `poetry run pytest
backend/blocks/test/test_orchestrator_execution_mode.py` — 14 tests pass
  - [x] `poetry run black` + `poetry run ruff check` clean
  - [ ] Manual verification on dev preview after combo redeploy

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>