Summary

Fixes #277 — the two most impactful Slack UX bugs:

• False fallback warning after real content: The Slack adapter posted ":warning: I didn't manage to produce a reply" even after real streamed text was delivered, because _postedThisTurn was not set after flushing the buffer on TurnCompleted.

• Invisible preamble during tool turns: When the model produced text alongside tool calls (e.g., "Let me search for that..."), the preamble sat invisible in the buffer until all tool iterations completed. Introduces a BufferFlush lifecycle signal emitted from HandleToolCallResponse that tells streaming adapters to flush their accumulated buffer before tool execution begins.

How BufferFlush works

1. LLM returns text + tool calls in a single response
2. HandleToolCallResponse emits TextOutput (for non-streaming path) + BufferFlush (for streaming path)
3. Slack adapter: if streaming deltas accumulated, BufferFlush flushes them to Slack immediately. If no streaming happened, the TextOutput posts directly via existing guard.
4. Buffer is cleared and ready for the next tool iteration's output

Files changed

Test plan

• Preamble_text_surfaces_before_tool_calls — verifies TextOutput + BufferFlush arrive before ToolCallOutput
• BufferFlush_roundtrips_through_dto — verifies DTO serialization roundtrip
• All 913 existing tests pass with zero regressions