feat(auth,prompts,inference): multi-tenancy MVP for MaaS deployments

[franciscojavierarceo]

Summary

Implements multi-tenancy support (Phase 1 + Phase 2) for MaaS, llm-d, and vLLM deployments.

Phase 1: Identity & fairness

• attribute_headers on UpstreamHeaderAuthConfig — maps multiple HTTP headers to attribute categories (e.g., X-MaaS-Group -> teams, X-MaaS-Subscription -> namespaces). Values merge with the existing attributes_header field. Enables MaaS Authorino integration where identity is spread across multiple upstream headers.
• fairness_header_attribute on vLLM config — injects x-gateway-inference-fairness-id on outgoing API calls from the authenticated user's attributes. Used by llm-d EPP Flow Control for per-tenant fair scheduling. Implemented as a _get_extra_request_headers() hook on OpenAIMixin so the pattern is reusable by other providers.

Phase 2: Data isolation (KVStore -> AuthorizedSqlStore)

Migrated all remaining stateful resources from plain KVStore (no access control) to AuthorizedSqlStore (row-level ABAC via owner_principal + access_attributes):

Resource	Before	After	Breaking change
Prompts	KVStore	AuthorizedSqlStore	Yes - recreate prompts
Connectors	KVStore	AuthorizedSqlStore	Yes - re-register connectors
Batches	KVStore (provider config)	AuthorizedSqlStore	Yes - in-progress batches lost

Not migrated (with rationale)

Resource	Reason
Vector store metadata	Routing table already enforces ABAC at the API boundary; KV metadata is defense-in-depth only
Distribution registry	Admin-only infrastructure, not user-facing data
Quota tracking	Per-client rate limiting, should NOT be tenant-scoped
Agent state persistence	Dead code (persistence_store initialized but never read/written)

Design decisions

1. _get_extra_request_headers() hook — added to OpenAIMixin rather than vLLM-specific injection. Any OpenAI-compatible provider can override it.
2. set_default_version crash safety — new default set before clearing old defaults, so a crash leaves two defaults (recoverable) rather than zero (data loss).
3. Distribution template updated — template.py default stores now use SqlStoreReference for prompts and connectors, ensuring all codegen'd distribution configs are consistent.

Files changed (27 files)

Area	Files
Auth	core/datatypes.py, core/server/auth_providers.py
Prompts	core/prompts/prompts.py, core/storage/datatypes.py, core/stack.py
Connectors	core/connectors/connectors.py
Batches	providers/inline/batches/reference/{__init__,batches,config}.py
Inference	providers/remote/inference/vllm/{config,vllm}.py, providers/utils/inference/openai_mixin.py
Templates	distributions/template.py
Distro configs	9 YAML files across ci-tests, starter, nvidia, oci, open-benchmark, watsonx
Docs	docs/docs/providers/{batches/inline_reference,inference/remote_vllm}.mdx
Tests	9 test files

Test plan

• 389 unit tests pass across all changed modules
• All 40 pre-commit hooks pass
• Integration tests (replay): uv run --no-sync ./scripts/integration-tests.sh --stack-config server:ci-tests --setup gpt --suite responses
• Verify attribute_headers with MaaS Authorino headers
• Verify fairness_header_attribute sends header to llm-d EPP

Generated with Claude Code

[mergify]

This pull request has merge conflicts that must be resolved before it can be merged. @franciscojavierarceo please rebase it. https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork