Keep per-run model overrides out of comparison columns

When opts.model or opts.roleModels is supplied together with models (possible through /api/run, and through CLI --models ... --roles ...), forwarding the unchanged options lets OpenRouter's complete() override the provider's column-bound model. Consequently, columns can execute on the same override or on role-specific models while still being labeled as their requested comparison model, corrupting comparison results. Comparison mode should remove these overrides or explicitly force every call to the current column model.

Useful? React with 👍 / 👎.

The global engine singleton defined on line 20 caches resolved providers in a Map that is never cleared. If clients send requests with unique or arbitrary model names, this cache will grow indefinitely, causing a memory leak in the long-running server process. Additionally, sharing provider instances across concurrent requests means that stateful providers (like the mock provider with its call counter n) will share state, leading to non-deterministic behavior.\n\nTo fix this, instantiate the engine dynamically per-request inside the POST handler, and remove the global engine declaration on line 20.

There is a bug in the provider resolution logic when a provider is explicitly requested via PROVIDER but its corresponding API key is missing.\n\nFor example, if PROVIDER=openrouter is set but OPENROUTER_API_KEY is undefined, the outer condition want === "openrouter" is true, but the inner condition if (env.OPENROUTER_API_KEY) is false. This causes the function to silently fall through to the next provider checks, eventually returning the mock provider. This completely bypasses the helpful configuration error checks defined inside the provider modules (e.g., throw new Error("OPENROUTER_API_KEY not set")).\n\nRemoving the inner key checks allows the requested provider to be returned directly, so it can throw a clear and actionable error message.

  if (want === "openrouter" || (!want && env.OPENROUTER_API_KEY)) {\n    return openrouter(env, model);\n  }\n  if (want === "anthropic" || (!want && env.ANTHROPIC_API_KEY)) {\n    return anthropic(env, model);\n  }\n  if (want === "openai" || (!want && env.OPENAI_API_KEY)) {\n    return openai(env, model);\n  }

The comparison mode currently runs sequentially because of the await inside the for...of loop. This means if a user compares multiple models, they must wait for each model's pipeline to completely finish before the next one starts. For multi-call orchestrations or chains, this can be extremely slow.\n\nRunning the models in parallel using Promise.all would significantly improve performance (especially for the web UI). Note that if you run them in parallel, the onEvent console logs for the CLI might interleave, but this can be managed or accepted as a trade-off for the massive speedup.

  const tasks = models.map(async (model) => {\n    onEvent({ type: "model:start", model });\n    const provider = getProvider(env, { model });\n    const engine = createEngine(provider, onEvent);\n    let result = null, error = null;\n    try {\n      if (mode === "single")     result = await engine.runSingle(ids[0], inputs, opts);\n      else if (mode === "chain") result = await engine.runChain(ids, inputs, opts);\n      else                       result = { results: await engine.runAll(ids, inputs, opts) };\n    } catch (e) {\n      error = e.message;\n    }\n    onEvent({ type: "model:end", model });\n    return { model, provider: provider.name, mode, result, error };\n  });\n  const columns = await Promise.all(tasks);