semantic: use negotiated float16 transport

[justrach]

Summary

• request compact base64 little-endian float16 vectors from the hosted CodeDB embedding endpoint
• preserve JSON-float requests for custom OpenAI-compatible endpoints
• decode either transport with strict length, encoding, index, model, and finite-number validation
• retry the hosted request once with a freshly signed legacy JSON body when an older edge rejects negotiation
• share the validated vector parser with advisory cosine scoring

Validation

• zig build test
• python3 scripts/e2e_mcp_test.py --binary zig-out/bin/codedb --project "$PWD" (76/76)
• live production hybrid context query succeeded against embeddings.wiki.codes
• origin parity: 25x512 minimum self-cosine 0.9999967; compressed response size reduced 58.8%

Private origin/edge support is already deployed and remains backward-compatible with released clients.