fix: unpack StreamingThinkScrubber tuple in streaming delta paths

[lld1995]

Every API call through the chat-completions transport failed with can only concatenate str (not tuple) to str. StreamingThinkScrubber.feed()/flush() return a (visible, reasoning) tuple but two call sites in run_agent.py assigned the whole tuple to text; the tuple later hit a str+ concat. Fixed by unpacking at _fire_stream_delta (hot path) and _reset_stream_delivery_tracking (end-of-stream flush). Verified locally and against the updated container image.

[liuhao1024]

This PR unpacks think_scrubber.flush() and think_scrubber.feed() as returning (visible, reasoning) tuples, but on current main both methods return str, not a tuple:

# agent/think_scrubber.py on main:
def feed(self, text: str) -> str:     # line 106
def flush(self) -> str:               # line 204

Merging this PR without the scrubber change would cause ValueError: not enough values to unpack at runtime on both call sites.

I see that your PR #23638 modifies agent/think_scrubber.py to return tuples — this PR appears to be stacked on top of that one. The diff alone doesn't include the scrubber change, so this would break if merged independently.

Suggestion: either:

1. Merge feat(api_server): stream inline <think> as delta.reasoning_content (unify A/B-class reasoning) #23638 first, then this PR can land cleanly, or
2. Combine both PRs into a single self-contained PR so reviewers can see the full picture (scrubber return type change + call-site unpacking).

Also, the PR body says "StreamingThinkScrubber.feed()/flush() return a (visible, reasoning) tuple" as if it's current behavior — but it's not on main. Adding a "depends on #23638" note would help reviewers understand the stacking.

[lld1995]

Closing this PR. As the reviewer correctly pointed out, this change only includes the run_agent.py call-site unpacking but not the agent/think_scrubber.py return-type change (on main, feed()/flush() still return str, so merging this alone would raise ValueError: not enough values to unpack at runtime).

The fix is already fully and self-containedly covered by #23638, which changes think_scrubber.py to return (visible, reasoning) tuples and updates both call sites in run_agent.py (_reset_stream_delivery_tracking + _fire_stream_delta) plus the api_server.py reasoning channel and tests. Consolidating there instead of stacking. Thanks for catching this!