providers/openrouter: pin the serving backend, record which fleet served

[lec77]

Problem

OpenRouter routes a model id to whichever upstream fleet it likes, and it re-routes between sessions. Two runs of the same skill against the same model id can therefore execute on different backends, with different quantization, different context handling, and different prompt-cache behaviour.

That is not a small effect. In a measured session pair on one model id, the prompt-cache hit rate went 0% → 75%, mean call latency 1.8s → 3.1s, and per-task scores moved by up to 58 points on byte-identical inputs (permutation test p < 0.0005). A benchmark repeated a few weeks later can be comparing serving paths rather than skills, and nothing in the results says which.

Change

• SKVM_OPENROUTER_PROVIDER (comma-separated, e.g. deepinfra,nebius) sets provider.order with allow_fallbacks: false on both request builders. Unset keeps today's behaviour exactly — OpenRouter's own routing.
• Every response records the fleet that actually served it as LLMResponse.providerName, from OpenRouter's provider field. A run's serving path becomes diagnosable after the fact instead of being inferred from cache-hit rates.

Opt-in, so nothing changes for existing users until they set the variable.

Verification

bunx tsc --noEmit clean; bun test 1297 pass / 0 fail. New cases in test/providers/openrouter.test.ts cover the pinned request body, the unset case (no provider key at all), and providerName parsing.

Note: touches the same parseResponse return and LLMResponse shape as #106, so whichever lands second needs a one-line rebase.

🤖 Generated with Claude Code