Clawbolt Files

Named LLM endpoints, and ten evaluator bug fixes.

Named endpoints (#1545). A provider name used to answer four questions at once: which any-llm adapter to use, whether prompt-cache markers are worth stamping, how to ask for reasoning, and which price list applies. Behind a gateway only the first stays true. An llm_endpoints row separates them, and reasoning: none exists for endpoints that reject the parameter's presence rather than its value. Also fixes the evaluator sending both sides of a comparison to the same gateway, and gives it a per-side reasoning effort recorded on the run.

Evaluator fixes (#1546). A five-turn run could return a firm do_not_switch off one turn; truncation was charged to the candidate without checking the incumbent; the transcript window could starve the oldest sampled turn of history; a rolling deploy could mark a live run interrupted; polling wrote an audit row every two seconds; and "Show more runs" 422'd on the fifth click.

Migrations 044 through 046.