refactor: table-driven LLM provider registry

[amal66]

Summary

Replaces the per-provider if/else chains in the LLM dispatch layer and the per-provider switch in the API-key lookup with a table-driven provider registry. Adding a provider (including OpenAI-compatible endpoints) becomes a single registerProvider() / registerApiKeyProvider() call — no edits to index.ts, models.ts, or userApiKeys.ts at every call site.

Changes

• backend/src/lib/llm/registry.ts (new): LLMProviderAdapter contract (id, matchesModel, stream, complete, tiered models) plus registerProvider, getRegisteredProvider, findProviderForModel, registeredProviderIds, allRegisteredModels, and a test-only reset hook.
• backend/src/core/apiKeyProviders.ts (new): table of provider → env-var names (claude, gemini, openai, openrouter, courtlistener) with registerApiKeyProvider, getRegisteredProviders, envApiKey, hasEnvApiKey, isApiKeyProvider, normalizeApiKeyProvider.
• backend/src/lib/llm/index.ts: built-in providers (claude/gemini/openai) are registered on module load; streamChatWithTools() / completeText() dispatch through the registry instead of if (provider === "claude") ... chains.
• backend/src/lib/llm/models.ts: providerForModel() consults the registry first (prefix heuristics kept as fallback for test contexts); resolveModel() also accepts any model declared by a registered provider.
• backend/src/lib/userApiKeys.ts: env-var switch and the hand-maintained PROVIDERS array replaced by the table; getUserApiKeyStatus() / getUserApiKeys() build their result dynamically from the registered provider list.
• backend/src/lib/llm/types.ts: Provider becomes an open string id; UserApiKeys gains an index signature so third-party providers can carry keys without type edits; CompleteTextParams extracted.
• backend/tsconfig.json: exclude __tests__/*.test.ts from the production build.
• Tests (new): backend/src/lib/llm/__tests__/registry.test.ts (registration, first-match-wins routing, insertion order, model-set union) and __tests__/models.test.ts (provider inference + model resolution).

Behavior is preserved: same three built-in providers, same env-var precedence (ANTHROPIC_API_KEY || CLAUDE_API_KEY, etc.), same model routing, same API-key status/lookup results.

Why

Today every new provider requires touching each dispatch site (streamChatWithTools, completeText), the model tables, and the API-key switch. With the registry, local LLMs and other OpenAI-compatible endpoints can be added from a single setup file — the extensibility path Will endorsed for OpenAI-compatible endpoints — without touching call sites.

Testing

• npm run build --prefix backend — passes (tsc, no errors).
• npx vitest run src/lib/llm/__tests__ — 2 files, 20 tests, all passing.
• Upstream has no test runner yet, so vitest was installed temporarily and uncommitted (npm install --no-save vitest); the committed tests are ready to run atop a test-harness PR once one lands.

Provenance

All changes are mechanical ports of code in amal66/mike@origin/main (commit b3166dd); exceptions:

• Fork-only features intertwined with the ported code were omitted, not rewritten: the air-gapped-mode branches in resolveModel/registerBuiltinProviders, the Ollama/demo provider setup, and the retry/circuit-breaker wrapper around dispatch (upstream calls adapters directly, as before). Doc comments referencing those omitted features were trimmed accordingly.
• userApiKeys.ts keeps upstream's existing encryption and logging; only the fork's table-driven lookup delta is ported.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CEguyEgXa9JjCciXCcVemC