Remove the InferenceGateway feature flag and enable the gateway everywhere

[mrubens]

Summary

The inference gateway has been running behind the InferenceGateway feature flag since it shipped. This removes the flag and makes the gateway the unconditional behavior for every deployment: gateway-covered provider API keys and ChatGPT subscription auth always stay on the control plane, and task sandboxes route inference through /api/inference with their run-scoped token.

Changes

• feature-flags: removed the InferenceGateway enum entry, its config block, the isInferenceGatewayEnabledForDeployment fail-closed evaluator helper, and the server export.
• db / sdk (dequeue): the inferenceGateway boolean option is gone from the resolver's public API. resolveEffectiveModelRuntimeEnv is now the control-plane entry point (routing, titles, summaries — no run token, so raw provider keys), and a new resolveSandboxModelRuntimeEnv is the sandbox entry point (gateway-covered keys stay on the control plane, advertised via R_INFERENCE_GATEWAY_KEYS; ChatGPT subscriptions route through the gateway). fetchResolvedRuntimeEnvVars calls the sandbox variant; the shared implementation keeps the distinction as a private parameter.
• compute-providers / controller: removed the inferenceGatewayEnabled option from BuildWorkerEnvOptions and all five worker-env builders; buildBaseWorkerEnv now always withholds gateway-covered provider keys from the worker daemon env. Deleted the controller's inference-gateway-flag.ts helper and its use in the five spawn-worker paths.
• api: the inference gateway handler no longer checks the flag before serving; run-token auth remains the gate.
• web: dropped the flag from mockFeatureFlags.
• tests: removed flag mocks and the flag-disabled rejection test; flipped worker-env expectations so gateway-covered keys (e.g. OPENROUTER_API_KEY, AI_GATEWAY_API_KEY) are asserted withheld while custom provider keys still flow through.

Notes

• A Redis outage can no longer flip inference back to direct provider keys; the old fail-closed path is gone entirely, which is strictly safer for key containment.
• Deployments that still have inference_gateway in their deployment metadata will see it surface as a legacy metadata control in the admin UI; it is ignored and can be deleted.

Validation

• pnpm check-types, pnpm lint, pnpm knip all pass.
• Targeted vitest runs pass for feature-flags, compute-providers worker-env, sdk dequeue-helpers, db model-runtime-config, controller spawn workers, api inference gateway, and the worker gateway tests.

[roomote-roomote]

No new code issues found. See task

_{Reviewed d837861}