[Bug]: reasoning parsers silently drop buffered partial delimiters at end of stream (no EOF flush hook)

[ighutake-debug]

Bug Description

All reasoning parsers can silently lose trailing text when a streamed response ends with a proper prefix of a delimiter token. The streaming pipeline only invokes parse_reasoning_streaming_incremental for decoded chunks and provides no end-of-stream flush, so any text a parser is holding in its internal buffer at stream end is never emitted.

Two concrete shapes:

1. BaseReasoningParser (used by deepseek_r1, qwen3, kimi, kimi_k25, kimi_thinking, glm45, step3, …): if the stream ends while the buffer is a partial token (e.g. </th), is_partial_token returns true forever and the fragment is dropped (crates/reasoning_parser/src/parsers/base.rs:90-93).
2. kimi_k2 (feat(reasoning): add unified kimi_k2 reasoning parser #1992): the trailing-partial hold-back that keeps split </think> / <|tool_calls_section_begin|> markers intact between chunks also holds a genuine stream-ending fragment (e.g. reasoning whose final character is <, or output truncated at </thi).

This also makes streaming disagree with one-shot behavior: detect_and_parse_reasoning on the same truncated input emits everything as reasoning.

Steps to Reproduce

let mut p = KimiK2Parser::new();
p.parse_reasoning_streaming_incremental("some reasoning</thi").unwrap();
// stream ends here (no more chunks) — "</thi" is never emitted

Expected Behavior

Held text is flushed when the stream ends: the pipeline calls a parser finalization hook at end-of-stream, and parsers emit any buffered remainder.

Actual Behavior

Buffered partial-delimiter text is silently lost.

Component

protocols (API types) / reasoning parser

Proposed Fix

• Add a finalization method to the ReasoningParser trait, e.g. fn flush(&mut self) -> Result<ParserResult, ParseError> with a default implementation returning the buffered remainder (or ParserResult::default() for parsers that hold nothing).
• Call it from the gRPC streaming response paths after the last decoded chunk.
• Update BaseReasoningParser and kimi_k2 to emit their held buffers from flush().

Surfaced during review of #1992 (codex P2 comment); kept out of that PR's scope per one-concern-per-PR.

Pre-submission Checklist

• I have searched existing issues and discussions
• I can reproduce this issue consistently
• I am using the latest version of SMG