feat(library): add local semantic knowledge library

[keli-wen]

Summary

• Add the focused quantmind.library public API with only LocalKnowledgeLibrary, SemanticQuery, and SemanticHit.
• Persist canonical knowledge in SQLite using knowledge_items for typed aggregate roots and knowledge_nodes for normalized tree nodes; keep semantic_records and NumPy exact-cosine matrices rebuildable and derived.
• Index flat items, tree roots, and non-root tree nodes at the feat(library): local semantic knowledge library MVP #111 grain, with deterministic per-target embedding invalidation, transactional put/delete, typed get, combined metadata filters, financial-time filters, and explicit not-found/stale/corrupt behavior.
• Add optional BaseKnowledge.available_at, keep provider/storage seams private, and enforce the library-to-knowledge import boundary without introducing a generic RAG/backend API.
• Keep LocalKnowledgeLibrary as a 248-line public facade and group concrete implementation under quantmind.library._internal: retrieval-target construction, index-only embeddings, the concrete SQLite store, and deterministic exact-cosine ranking.
• Remove the standalone private ports module; the fakeable embedding protocol now lives beside its OpenAI implementation and remains private.
• Add an executable AI-infrastructure scenario backed by a real arXiv paper tree, Microsoft investment announcement, and NVIDIA earnings release.
• Precompile the auditable canonical JSON into a 96 KiB SQLite bundle containing three items, four normalized tree nodes, and six real text-embedding-3-small 1536-dimensional targets; the user example only performs open → search → get and embeds the query once.
• Move the maintainer build tool to scripts/examples/build_ai_infrastructure_bundle.py and add examples/library/README.md so agents can discover the run path, artifact provenance, and rebuild procedure locally.
• Document why SQLite is the canonical default and where a future vector database such as Chroma could replace only the private derived-index layer.

Related Issue

Closes #111.

Example Sources

• Compute Trends Across Three Eras of Machine Learning
• Microsoft's FY2025 AI-datacenter investment announcement
• NVIDIA's FY2026 Q1 results

Verification

• pytest -q tests/library --no-cov — 23 passed.
• bash scripts/verify.sh — 317 passed with 86.47% total coverage; Ruff, basedpyright, import-linter, and pytest all passed.
• SQLite PRAGMA integrity_check returned ok; the compiled bundle contains exactly 3 canonical items, 4 canonical nodes, and 6 text-embedding-3-small × 1536 semantic records.
• The relocated maintainer build and user-facing example were both run successfully against the configured OpenAI API; the live query returned all six targets across News, Paper, and Earnings.
• Commit and pre-push hooks passed without bypassing; the pre-push hook reran scripts/verify.sh successfully.
• The committed bundle was checked for API-key-like strings, and the offline test uses a private deterministic query provider without network access.
• Unsupported knowledge types are tested to fail before any embedding-provider call.

Checklist

• The title uses English Conventional Commit format.
• The related issue is linked.
• bash scripts/verify.sh passes.
• Public behavior has focused unittest.TestCase / unittest.IsolatedAsyncioTestCase coverage, one focused example, and documentation.
• Canonical tree nodes are stored separately from aggregate-root payloads.
• The user-facing example does not parse source JSON, ingest documents, or build document embeddings.
• Private implementation files describe concrete responsibilities and do not imply alternative library backends.
• The change is scoped to feat(library): local semantic knowledge library MVP #111 and contains no paper-flow integration, answer generation, knowledge graph, PageIndex, public backend interface, or generic RAG abstraction.