feat: isolate default and stateful execution profiles

[danny-avila]

Summary

• add a stable CODEAPI_EXECUTION_PROFILE=default|stateful deployment identity
• preserve the existing default BullMQ queues while isolating stateful jobs onto dedicated queues
• add X-CodeAPI-Expected-Profile fail-closed routing assertions and advertise the actual profile on every response
• reject contradictory profile/backend/session-mode configurations at startup
• expose low-cardinality profile metrics and document the contract in OpenAPI, Helm, and operator docs

Why

This establishes the Code API boundary needed for #38 and AI-1650: LibreChat/Agents can route each executing agent to either the AWS-free stateless deployment or the stateful Lambda MicroVM deployment without letting the two worker pools consume each other jobs.

There is deliberately no silent fallback and no workspace/file migration between profiles.

Downstream

• feat: route code execution per agent profile danny-avila/agents#416
• feat: route code execution per agent profile danny-avila/LibreChat#14848

Compatibility

• HTTP/stateless remains the default and keeps python-queue / other-queue.
• Existing affinity/strict deployments infer stateful when the new env var is omitted during rollout.
• API-only pods validate profile/session semantics without requiring worker-only AWS configuration.
• Older clients may omit the expected-profile header; trusted new callers always send it.

Validation

• 542 pass, 0 fail across the complete service test suite
• bun run build
• OpenAPI YAML parse
• helm lint helm/codeapi -f helm/codeapi/values-local.yaml
• exact-head CI: 5/5 green
• git diff --check

[danny-avila]

@codex review