fix: resilient DSML tool-call parsing and repair in long contexts

[jackygurui]

Summary

Long-context and long-generation scenarios expose several failure modes in DSML tool-call parsing. The model's attention degrades over thousands of tokens, producing malformed DSML that causes finish=error and breaks the agent loop. This PR adds a robust repair layer (try_repair_dsml) that handles all observed failure modes, plus targeted fixes for false positives and visibility improvements.

Problem

Three failure modes were observed in stress testing (256K context, q4-imatrix):

Mode	Description	Symptom
Unterminated	Model stops mid-DSML, missing closing tags	finish=error, "unterminated tool call"
Malformed closed	Outer tags balanced but inner tags broken	finish=error, "invalid tool call"
Hallucinated	<tool_calls>...</tool_calls> wrapping plain reasoning text	False tool call detection

Before this fix, all three modes would land in finish=error, silently aborting the agent loop. Additionally, DSML tags mentioned inside <thinking> text (e.g. the model explaining tool syntax in its reasoning) were being counted by the tag scanner in try_repair_dsml, causing false positive repairs.

Changes

1. try_repair_dsml() — single-pass DSML repair (2 commits)

   • Rewritten from 6-pass O(6n) to single-pass O(n) tag counting for all 6 tag types
   • Unterminated repair: appends missing closing tags in correct nesting order (parameter → invoke → tool_calls)
   • Hallucinated repair: when tool_calls tags exist but contain no <invoke>, strips the orphaned tags so the text is treated as plain content
   • Orphan end tag guard: returns early when closing tags outnumber opening tags (prevents size_t underflow)
   • Integrated into both the unterminated-tool-call path and the invalid-tool-call recovery path
   • Added 10 unit tests covering all 3 DSML styles, unterminated, malformed, hallucinated, and balanced cases

2. Ignore DSML inside <thinking> (1 commit)

   • try_repair_dsml now finds the last </thinking> boundary and scans for DSML tags only from there — deliberately matching the same strategy used by parse_generated_message_ex, which was introduced in commit 037ee39 ("Ignore tool calls emitted inside thinking") to fix the same class of false positives reported in #167
   • The hallucinated tag-stripping path also preserves the thinking section verbatim, only stripping tags from the post-thinking region

3. Debug visibility (1 commit)

   • Logs a stderr message when require_thinking_closed=true is triggered but the model never closes <thinking>

Results

Before this fix, the 38 repaired + 3 orphan + 2 invalid cases would all have been counted as invalid, with every occurrence causing a hard finish=error. After the fix:

Tool calls: 169 | Invalid: 2 | Repaired: 38 | Orphan: 3

Only 2 cases remain unrecoverable; the other 41 are gracefully handled (38 repaired, 3 hallucinated-tags stripped). 100% repair success rate on unterminated tool calls.