Claude/qwen claude reverse eng v hu hv - #76

…1-indexed

Release assets: shard-01..shard-11 (1-indexed, from indexing pipeline)
Manifest/local:  shard-00..shard-10 (0-indexed, matching bgz7_path())

Fix: download URL uses shard+1, local storage stays shard.
Without this fix, hydrate --download tries to fetch shard-00 which
doesn't exist in the release, causing 404 on every model.

https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK

Endpoints:
  POST /v1/chat/completions — all 7 models (qwen35-opus46/45, 9b, reader-lm, bge-m3, llama4-scout, openchat-3.5)
  POST /v1/embeddings — Base17 fingerprints as 17-dim OpenAI embedding format
  GET  /v1/models — 7 models with descriptions and bgz7 sizes
  GET  /health — ok

Fixes:
  - Model validation: unknown models → 404 with available list
  - Model descriptions include bgz7 size and purpose

Latency: 7ms per request. Tested with 10-test curl suite.

https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK

bge-m3: XLM-RoBERTa forward pass (24 layers, GELU, mean pool, L2 norm)
  + SentencePiece tokenizer stub + embed_to_base17() projection
reader-lm: Qwen2 forward pass (28 layers, GQA 12:2, RoPE, SwiGLU)
  + BPE tokenizer stub + generate() + html_to_markdown()

Both fall back to deterministic hash when no weights loaded.
Key finding: bgz7 dims have σ=0.5, range=[-2,+2], L1 μ=2.9 σ=1.8
→ SimilarityTable σ-buckets needed for discrimination (not absolute threshold)

https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK

bgz7_to_batch(): Base17 fingerprints → Arrow FixedSizeList[i16,17]
compute_heel(): column-wise bundle of all rows → HEEL vector
hydrate_bgz7(): read bgz7 file → RecordBatch (ready for Dataset::write)

4 tests passing. Schema: tensor_name, row_idx, base17[17], palette_s/p/o.
HEEL test confirms non-zero for asymmetric data.

https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK