refactor: type-safe embedding model system, reindexing safety fixes, and tool formatting extraction

[lucreiss]

Summary

Restructures the embedding model/provider type system to be data-driven and type-safe,
fixes several bugs in the reindexing and compatibility check flow, extracts tool formatting
logic into maintainable utilities, and adds support for Google's task-specific embeddings
and newer embedding models.

Changes

• Config restructuring: Extract EMBEDDING_MODELS constant as single source of truth
  for all model metadata; derive EmbeddingProvider, EmbeddingModelName, and related
  types from it at compile time
• Embedding provider improvements: Split embed() into embedQuery()/embedDocument()
  to support Google's task-type optimization (CODE_RETRIEVAL_QUERY vs RETRIEVAL_DOCUMENT);
  add gemini-embedding-001 with Matryoshka truncation; replace deprecated
  text-embedding-004 with text-embedding-005; add Google batch request support
  (up to 20 texts per call)
• Reindexing safety fixes: checkCompatibility() no longer silently returns
  compatible: true on uninitialized state (throws instead); clearIndex() now deletes
  stale index metadata so force-rebuild works correctly with a new provider; index()
  blocks before writing if the index is incompatible; search() and findSimilar() now
  call ensureInitialized() before the compatibility check (previously ran on null state);
  findSimilar() gains a compatibility check it previously lacked; add provider mismatch
  detection via new IncompatibilityCode enum
• Tool formatting extraction: Move all formatting functions from src/tools/index.ts
  to src/tools/utils.ts for maintainability and organization
• Indexer type exports: Export SearchResult, HealthCheckResult, StatusResult
  interfaces from the indexer module
• Test coverage: Comprehensive tests for config parsing, model validation, and all
  extracted formatting utilities

Testing

• Unit tests added/updated
• Manual testing performed
• Build passes (npm run build)
• Tests pass (npm run test:run)
• Lint passes (npm run lint)

Related Issues

N/A

[Helweg]

Great restructuring — the data-driven model system and embedQuery/embedDocument split for Google task types are clean improvements.

A few things:

1. Bug in createEmbeddingProvider default branch (src/embeddings/provider.ts): The error message interpolates the whole configuredProviderInfo object — throw new Error(\Unsupported embedding provider: ${configuredProviderInfo}`)— which will produce[object Object]. Should be configuredProviderInfo.provider`.

2. Score label change in codebase_search: formatSearchResults now labels all scores as (similarity: X%). Previously codebase_search used (score: 0.85) while only find_similar used the percentage format. For hybrid search results (semantic + BM25 fusion), "similarity" is misleading since the fused score is a relevance rank, not a cosine similarity. Consider keeping the old format for codebase_search.

3. Provider mismatch check: The new PROVIDER_MISMATCH incompatibility will force a full rebuild when switching between github-copilot and openai with the same model (text-embedding-3-small), even though Copilot proxies to OpenAI and produces identical embeddings. The dimension and model checks already catch genuinely incompatible scenarios — is the provider-only block intentional, or should it be a warning?

4. Missing trailing newlines: src/config/constants.ts, src/config/index.ts, and src/tools/index.ts are all missing a final newline.

Otherwise LGTM.

[lucreiss]

Thanks for the thorough review! I addressed all four points:

1. createEmbeddingProvider default branch — You're right about the [object Object] bug.
I went a step further and replaced the default with an exhaustive never check. Since
ConfiguredProviderInfo is a discriminated union over the four known providers, the default
was already unreachable — removing it and adding never turns a missing case into a
compile-time error rather than a runtime one.
2. Score label in codebase_search — Good catch, I missed this when extracting
formatSearchResults into utils. codebase_search now uses (score: 0.85) format again
(appropriate for fused hybrid results that blend cosine similarity + BM25), while
find_similar keeps the (similarity: 92.0%) percentage format (pure cosine similarity).
I added a ScoreFormat parameter to formatSearchResults and two new tests covering both
formats.
3. Provider mismatch — I had this as intentional as a defensive measure — model names
like text-embedding-3-small are generic enough that a different provider could theoretically
reuse the name with different embeddings. That said, you're right that forcing a full rebuild
for github-copilot ↔ openai with matching model + dimensions is unnecessarily disruptive.
I changed it to return compatible: true with a logged warning instead. The dimension and
model checks remain as the real safety net.
4. Trailing newlines — Fixed in all three files.

All checks pass: build, typecheck, lint, 321/321 tests (2 new).

[Helweg]

The never exhaustive check is a nice touch — strictly better than the runtime default. And the ScoreFormat parameter is a clean way to handle the two contexts.

LGTM, ship it.