emrg: persist and pass back DeepSeek thinking-mode reasoning

[argszero]

Fixes DeepSeek thinking-mode 400 error: "The reasoning_content in the thinking mode must be passed back to the API" (LLM stream error in round N).

Root cause (verified against source + logs):

1. llm.py correctly accumulates reasoning_content into full_reasoning during streaming;
2. but the three append_message calls in daemon.py persisted assistant messages with only content + tool_calls (reasoning went only to llm.jsonl);
3. on the next request, session.get_messages_for_llm() rebuilt assistant messages without reasoning;
4. DeepSeek thinking mode requires assistant messages (especially tool-call rounds) to pass back the previously returned reasoning_content verbatim — missing it triggers the 400 inside tool loops.

Fix (4 changes):

• daemon.py ×3 main-loop append_message sites: persist reasoning when not None (old records untouched);
• daemon.py main-loop in-memory messages: reasoning_content added to the assistant message appended for the next round — closes the within-tool-loop path (round N error);
• daemon.py memory-reflection + task-handler tool loops: pass back reasoning_content from llm.chat responses (same bug class);
• session.py get_messages_for_llm: maps persisted reasoning → reasoning_content; records without the field are skipped naturally (old histories unaffected).

Compatibility: folded placeholder assistant messages (no tool_calls) are not required by DeepSeek to carry reasoning; history size grows slightly (API hard requirement) — a reasoning-window fold can be a follow-up.

Verification: pytest 1010 passed + 1 skipped (1011 collected, +3 new regression tests); import/help OK; Agent.md count synced to 1011.

[argszero]

✅ LGTM — cycle 2026-08-22T19:26:07. Fix is correct and complete:

• Root cause verified: reasoning_content accumulated in llm.py streaming but dropped on session persistence and message reassembly, so subsequent rounds lack the verbatim reasoning DeepSeek thinking mode requires (400 error).
• daemon.py covers all four message-building paths (final answer, in-loop tool_calls round, _reflect, _consolidate_session_memories) — reasoning persisted as 'reasoning' and re-injected as 'reasoning_content'.
• session.py get_messages_for_llm() maps persisted reasoning → reasoning_content; old records without the field are skipped naturally (no regression).
• Tests cover plain pass-back, tool-call rounds, and legacy-record compatibility.
• CI green (test + test-windows, run 32566754537).

[argszero]

✅ LGTM — cycle 47. Re-verified: head unchanged (8858728), diff covers all four assistant-message paths (final answer / tool-call round / _reflect / task-handler) with reasoning persisted + re-injected as reasoning_content, session.py maps persisted reasoning back with natural skip for legacy records. 3 regression tests added; local pytest 1010 passed + 1 skipped; CI green (test + test-windows, run 32566754537).

[argszero]

✅ LGTM — cycle 48. Re-verified: head unchanged (8858728), CI green (test + test-windows, run 32566754537), fix covers all assistant-message paths + session mapping + legacy compat. 3rd consecutive ✅ — merging.