chimera skillcard-bench

A/B reasoning with vs without injected TRS skill cards. Calls real models.

Options

--tasksstrdefault: 'hard'
Task suite: hard | big | demo. 'big' = 24 traps for a tighter paired CI.
--kintdefault: 1
How many cards to retrieve per task.
--min-overlapintdefault: 2
Relevance gate: inject a card only on >= N shared query terms (0=off).
--max-linesintdefault: 3
Render budget: max lines per injected card.
--use-storeboolean
Bench your own learned cards (skills.json) instead of the demo set.