v3.44.2

simplicio-loop 3.44.2

• bench/llm_ab/run.py --tasks 4 --hard adds a hard Python set with hidden acceptance tests outside the arm repo: coupon logic with a half-up rounding trap, a two-bug fix, a two-file refactor, and a duration parser. --turbo-reasoning keeps the model's reasoning on for the turbo calls, so on and off can be compared. Turbo also sends a task's context files to the model. The checker accepts an absolute path.
• Hard-set result: with reasoning off, turbo passed 12/12 hidden-test tasks at 5.4 s and $0.0027 per run (mean of 3). The OpenCode arm passed 11/12 at 112.6 s and $0.0122. With reasoning on, turbo passed 7/8, and one call ran away to 131k reasoning tokens. Turbo keeps reasoning off.
• Every benchmark run of 3.44.0 to 3.44.2 is archived under bench/llm_ab/results/runs/ with a summary README.md, outside the release-to-release history.

PR #1374, issue #1373. Every benchmark run is archived in bench/llm_ab/results/runs/ (summary in its README.md).

Hard set, turbo reasoning off vs on (2026-09-29-hard/)

Four Python tasks with hidden tests (--tasks 4 --hard, loop at 5d11aca9).

task	normal, 3 runs	simplicio reasoning off, 3 runs	simplicio reasoning on, 2 valid runs
1 pricing (half-up rounding)	✓ ✗ ✓	✓ ✓ ✓	✓ ✓
2 inventory bug fix	✓ ✓ ✓	✓ ✓ ✓	✓ ✓
3 two-file refactor	✓ ✓ ✓	✓ ✓ ✓	✓ ✓
4 duration parser	✓ ✓ ✓	✓ ✓ ✓	✓ ✗
passed	11/12	12/12	7/8
wall, mean	112.6 s	5.4 s	18.1 s / 302.5 s
cost, mean	$0.0122	$0.0027	$0.0090 / $0.1621
reasoning tokens, mean	2,485	0	5,306 / 132,832

• normal off-2, task 1: half-up rounding was wrong (expected 1703 and 1712, got 1704 and 1713).
• Reasoning-on on-2, task 4: one call ran to 131,072 reasoning tokens in 294 s and returned no plan, so duration.py was never written.
• on-3-http402/ and on-3-rerun-http402/ are invalid runs. The key ran out of credits (HTTP 402), so every call failed. They are kept only as a record.

Decision: turbo keeps reasoning off.