fix(reflect): don't surface raw harmony tool-call scaffolding as the answer (#2222)

[r266-tech]

Fixes #2222.

Problem

When the reflect LLM is a gpt-oss / harmony-format model served through litellm (e.g. vertex_ai/openai/gpt-oss-*-maas), the provider intermittently fails to parse a tool-call turn into structured tool_calls. The raw harmony scaffolding then lands in message.content (call_with_tools reads message.content and only builds tool_calls from message.tool_calls), and the reflect agent's no-tool-calls branch accepts that content verbatim as the final answer:

<|start|>assistant<|channel|>commentary to=functions.recall<|message|>{"query": "..."}<|call|>

_clean_answer_text only strips done(...) suffixes, so the harmony tokens pass through. With no tool executed, citations are empty too — the corrupt, unsynthesized output is indistinguishable from a real low-confidence answer and surfaces to end users (reported on v0.8.2).

Fix

In the if not result.tool_calls: branch, detect leaked tool-call scaffolding before accepting result.content as the answer. When detected, fall through to the existing forced final-synthesis path (build_final_prompt), which re-synthesizes from the gathered evidence instead of echoing the scaffolding.

Detection (_looks_like_unparsed_tool_call) is intentionally narrow to avoid rejecting valid answers:

• Harmony scaffolding is flagged only when a frame token (<|start|>/<|channel|>/<|message|>) co-occurs with call evidence (a <|call|> commit token or the to=functions/to=tool routing namespace). An answer that merely quotes one of these tokens while explaining the format has no such pairing and is left untouched.
• A done() payload emitted as bare JSON is flagged only when the object carries the tool's internal citation id fields (memory_ids/mental_model_ids/observation_ids), which never appear in a real free-text answer. A plain {"answer": ...} blob with no id fields is deliberately left alone, being indistinguishable from a valid JSON answer.

Known trade-off: an answer that embeds a full harmony tool-call example verbatim (frame token + call evidence together) would be re-synthesized rather than echoed. This is accepted as non-destructive (the fallback reuses the same gathered evidence; no data loss or error) and vanishingly rare for a memory-synthesis answer.

Tests

tests/test_reflect_agent.py:

• TestUnparsedToolCallDetection — positive cases (harmony call structures + leaked done-JSON with id fields) and negative cases (normal prose, answers explaining harmony tokens, bare {"answer": ...} JSON).
• test_unparsed_harmony_scaffolding_is_not_returned_as_answer / test_leaked_done_payload_with_ids_is_not_returned_as_answer — assert the corrupt content never reaches result.text and the agent routes to final synthesis.

This is the defense-in-depth (option B) from the issue; a deeper litellm-side harmony parser (option A) is out of scope here.