fix: flush Coze stream tail when no done event is emitted

[WODE25500]

What

Self-hosted Coze does not emit a terminating done event, but the runner detects the
terminal event with event_type.split('.')[-1] == 'done', so is_final can stay False
for the whole stream. Since the runner only flushed on a message_idx % 8 == 0 boundary:

• tail text that did not land on a boundary was never yielded;
• a stream ending exactly on a boundary never delivered is_final at all, leaving the
  upstream stream open -- adapters that only deliver on is_final (e.g. WeCom) hang
  until timeout.

The runner now emits one final cumulative snapshot after the loop whenever no terminal
event was seen. The guard is not is_final rather than a content-length comparison: on
a boundary ending the content length is unchanged, yet the terminal signal still has to
be sent. Re-sending the same cumulative snapshot is safe -- every reply_message_chunk
implementation treats chunk content as the full message text (replace semantics, never
append).

Tests

tests/unit_tests/provider/runners/test_coze_tail_flush.py drives the real runner
(_chat_messages_chunk) with a mocked Coze client:

• stream lengths 1..40 with no done event -> the last chunk is complete and is_final;
• boundary endings (7/15/23/31/39) -> exactly one terminal snapshot;
• done at every position, plus the self-hosted end event
  (conversation.chat.completed), which the runner does not match;
• an error event and a mid-stream exception are never reported as success.

uv run pytest tests/unit_tests/provider/runners/test_coze_tail_flush.py -q
uv run ruff check src/langbot/ tests/

[RockChinQ]

@dadachann Check if this PR is okay for merging, if not, submit your review

[dadachann]

👋 I’m dadachann, the LangBot community maintenance agent. I’ve picked up the review/task on this pull request and will share findings and next steps here.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[dadachann]

• Your tail flush helps short streams, but the Coze runner still omits a final signal at chunk boundaries. This remains a blocker: yielding the text does not guarantee QQ sends it.
• Isolated replay using the real runner and simulated QQ sending reproduced truncation: 15 rapid increments sent only the first 7 characters and left the stream open; 16 completed correctly. A 7-increment fallback case sent nothing. Live Coze-to-QQ reproduction remains unverified.
• A fix should guarantee a complete final snapshot on normal stream completion, even when its length is unchanged, without treating errors or cancellation as success. The added tests exercise a handwritten emulator rather than the runner; replace these with real-runner tests and adapter-throttling coverage. No product changes were made during this review.