fix: resolve issues #40, #41 — fact_id consistency + enterprise_context description

fix: reindex flush + stub detection for #37

fix: resolve issues #36, #37, #38 — search_facts reliability, 32/32 tests pass

fix: resolve issues #29-#34 — 6 bugs, 21/21 regression tests pass

fix: #28 singleton cache for SentenceTransformer interceptor — 1 load, not 3

fix: close codex audit findings — BLOCKER + HIGH + MEDIUM

fix: #21 session restore + #22 causal context — verified by 26/26 tests

fix: #24 don't set DEFAULT_MODEL for ollama/openai (incompatible name format)

fix: resolve issues #19-24 — search_facts, Windows paths, causal, fingerprint, single model

fix: tool filtering via _tool_manager._tools dict (MCP SDK v2)

fix: tool filtering for MCP SDK v2+ (_tool_manager)

fix: Float32SafeEmbedder for all providers, class-level embedding patch, profiles

feat: multilingual default embedding, LM Studio/OpenAI-compat support

fix: default tool set (12 tools instead of 50)

fix: healthcheck for streamable-http, tool filtering, embedding strict mode