feat: lazy + configurable LLM fallback chain

[felixschmetz]

Summary

• Let the backend boot when no chain-provider API key is configured: _build_llm_chain() returns an UnavailableLLM null-object instead of raising. Instant search works immediately; classic/agentic surface LLMUnavailableError on first use and return HTTP 503 via the existing exception handler.
• Expose LLM_FALLBACK_CHAIN as an env var (format: provider:model,provider:model) so private-cloud deployers can override the chain without forking. Unset keeps the current [together:zai-glm-5, anthropic:claude-sonnet-4.6] default — no change for Airweave self-serve.
• Services keep required, non-null llm: LLMProtocol signatures. Null-object carries the "unavailable" state so no Optional threading or None-checks inside search services.

Context

A deployer was blocked on v0.9.65 because only MISTRAL_API_KEY was set, and the hardcoded chain required a Together or Anthropic key. Their actual plan is to use Instant search (no LLM needed); the primary ask was "don't break deployment." A follow-up ask from another deployer: private-cloud customers don't always have the same providers as self-serve, so the chain should be configurable per deployment rather than via a fork.

Test plan

• New unit tests (17 total): adapters/llm/tests/test_unavailable.py, domains/search/tests/test_config.py, core/container/tests/test_llm_chain_wiring.py
• Existing unit suite still green (1111 tests), LLM adapter tests green (65), billing listener tests green (23)
• ruff check clean on changed files; ruff format --check clean; lint-imports green; no new mypy errors
• Smoke script: factory returns UnavailableLLM with no keys set and logs the expected INFO message; parser accepts mistral:mistral-large and rejects unknown providers/models with the accepted-value list
• Manual docker compose up with only MISTRAL_API_KEY — confirm /search/instant works, /search/classic and /search/agentic return 503
• Manual override: LLM_FALLBACK_CHAIN=mistral:mistral-large with MISTRAL_API_KEY — confirm all three search tiers work
• Manual invalid chain: LLM_FALLBACK_CHAIN=bogus:foo — confirm startup fails with an actionable parser error

---

Summary by cubic

Make the LLM fallback chain lazy and configurable so the app boots without LLM keys; instant search works, and classic/agentic return HTTP 503 until a key is set or a chain is configured. Also removed unused weaviate and weaviate-client packages to unblock Poetry 2.x Docker builds.

• New Features

  • Injects a null-object UnavailableLLM when no providers resolve, so boot doesn’t fail; classic/agentic raise LLMUnavailableError mapped to 503.
  • Adds LLM_FALLBACK_CHAIN env var (provider:model,provider:model); parser validates provider names, model names, and valid provider–model pairs at startup.
  • Docs: added and refined “Configuring the LLM provider chain” with API key list, override format, and startup validation notes.

• Migration

  • Self-serve: no change.
  • Private-cloud: set a provider API key or set LLM_FALLBACK_CHAIN (e.g., mistral:mistral-large) to enable classic/agentic; invalid values fail fast with accepted providers/models and valid combinations listed.

^{Written for commit ada1033. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 9 files