Add optional Ollama-backed semantic sidecar reranking (lexical-first, additive)

[HiveForensicsAI]

Motivation

• Provide an optional local semantic reranking layer backed by Ollama embeddings while preserving Knolo's deterministic lexical-first retrieval model.
• Keep semantic logic additive, explainable, and local-first without introducing a vector DB or changing the .knolo pack format.
• Use a sidecar index to enable fast, portable reranking over only lexical top-N candidates and allow gating by lexical confidence.

Description

• Added a semantic core under packages/core/src/semantic/ with types.ts, cosine.ts, sidecar.ts, provider.ts, and rerank.ts implementing provider interfaces, normalization/cosine utilities, sidecar serialization/validation, and lexical-topN reranking/blending.
• Extended query() API and validation to accept semantic options (sidecar/provider/topN/minLexConfidence/blend/etc.) and attached per-hit evidence fields (retrieval, lexicalScore, semanticScore, blendedScore, modelId) while preserving the default lexical-first behavior.
• Implemented an Ollama adapter in a new package @knolo/semantic-ollama that implements EmbeddingProvider with batching, timeouts, and clear errors, and added CLI sidecar commands in the knolo binary: semantic:index, semantic:inspect, and semantic:validate.
• Added tests and README docs: sidecar rerank + validation tests, cosine helper tests, CLI semantic validation tests, and README usage/troubleshooting for the sidecar workflow.

Testing

• Built core and adapter packages with npm run build --workspace @knolo/core and npm run build --workspace @knolo/semantic-ollama and both builds succeeded.
• Ran core unit tests with npm run test --workspace @knolo/core and all tests passed (including testSemanticSidecarRerankAndValidation and testCosineHelpers).
• Ran CLI tests with npm run test --workspace @knolo/cli and all CLI tests passed (including semantic:validate success/failure cases).
• Verified knolo semantic:index / semantic:inspect / semantic:validate flows via unit tests that exercise sidecar creation/inspection/validation.

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b8ac727f6d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Wire semantic.sidecarPath into rerank eligibility

query() accepts semantic.sidecarPath, but semantic rerank is gated by if (!pack.semantic && !opts.sidecar) return false and the function never resolves sidecarPath into opts.sidecar. In practice, callers that pass only sidecarPath (with semantic.enabled, queryEmbedding, and even force) silently fall back to lexical-only retrieval, so the new option does not actually enable sidecar reranking.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Preserve pre-rerank score in evidence.lexicalScore

When semantic reranking runs, prelim is replaced with blended/semantic scores, but evidence.lexicalScore is still assigned from r.score here. That means hybrid results report reranked scores as "lexical" scores, which corrupts explainability/telemetry and any downstream logic that uses lexical confidence when semantic rerank is enabled.

Useful? React with 👍 / 👎.