emrg: fix double-accumulated reasoning in llm.jsonl (O(n²) blowup)

[argszero]

Fix double-accumulated reasoning in llm.jsonl (O(n²) blowup)

Problem

emrg/server/daemon.py appended every per-chunk cumulative reasoning snapshot to reasoning_parts and joined them at the end. But llm.py already accumulates reasoning internally and yields the full think text so far on every chunk (chunk["reasoning"] = "".join(reasoning_parts), llm.py:366).

Appending N cumulative snapshots and joining them re-accumulates the text N times — O(n²) growth:

• 47 real thinking tokens → 29,640 chars in the record (~630×)
• worst single llm.jsonl record: 9,968,917 chars (~10 MB)
• ~190 MB of disk across 3 rotations

Fix

Take the last snapshot — it is the complete think text:

full_reasoning=reasoning_parts[-1] ifreasoning_partselseNone

A comment in the code documents the llm.py accumulation contract so this cannot regress silently.

Verification

• uv run pytest tests/ -q → 1006 passed, 1 skipped
• from emrg.client.app import run_client + python -m emrg --help OK
• The llm.py accumulation contract itself is already covered by tests/test_llm.py (test_stream_accumulates_reasoning_content, test_stream_accumulates_openai_reasoning, test_stream_no_reasoning_means_none)

Ref: host rant 2026-08-23T10:15:06 (P0).

[argszero]

✅ LGTM — cycle 1287. The daemon-side double-accumulation is confirmed against the llm.py contract (chunk['reasoning'] is already the cumulative snapshot, llm.py:366); taking the last snapshot yields the complete think text. Single accumulation site in daemon.py — fix is complete.

[argszero]

✅ LGTM — cycle 1288. Fresh-eyes re-review of the final head (835ebb0, after master merge): the one-line fix correctly takes the last cumulative snapshot (chunk['reasoning'] is the full think text per llm.py:366, so [-1] is complete and O(n) instead of O(n²)). Empty reasoning_parts → None preserves the no-reasoning behavior; all three downstream full_reasoning sites (assistant-message persist + llm.jsonl logs) are unchanged. The conflict resolution also correctly kept #942's cache-pct debug line. CI test + test-windows pass on head; local pytest 1008 passed + 1 skipped.

[argszero]

✅ LGTM — cycle 1289. Third consecutive fresh-eyes review (head 835ebb0, CI test + test-windows green, MERGEABLE). The fix takes the last cumulative reasoning snapshot (chunk['reasoning'] = full think text per llm.py:366), eliminating the O(n²) double-accumulation that inflated a 47-token think to a 29,640-char llm.jsonl record; empty reasoning_parts → None preserves prior behavior. No ❌ since the first LGTM — merge condition (3 consecutive ✅ from different cycles) satisfied.