Qualify E5-Mistral-7B, and support multiple reference sets

[bsbodden]

Extends the gate past a single hardcoded reference, qualifies one more model, and reports a real divergence the gate caught.

Multiple references

The gate pinned reference-qwen3-embedding-0.6b-q8_0.json, so it could only test one model. References are now listed in references.txt and selected by artifact SHA-256, so a model can only be compared against vectors generated from its own exact bytes. Pointing it at an unknown artifact exits 2 rather than silently comparing against the wrong model.

Qualified

E5-Mistral-7B-Instruct Q4_K_M reproduces llama.cpp 6ea215d17: min cosine 0.9995117, mean 0.9996140, max norm deviation 5.54e-09.

Not qualified, and why

The three Qwen3-Embedding Q4_K_M variants fail. The cause is narrower than "Q4_K is lossy" — holding the model fixed and changing only quantization:

model	arch	quant	min cosine
Qwen3-Embedding-0.6B	qwen3	Q8_0	0.99950	pass
Qwen3-Embedding-0.6B	qwen3	Q4_K_M	0.99700	fail
Qwen3-Embedding-4B	qwen3	Q4_K_M	0.99849	fail
Qwen3-Embedding-8B	qwen3	Q4_K_M	0.99797	fail
E5-Mistral-7B	llama	Q4_K_M	0.99951	pass

Q4_K passes on llama and fails on every qwen3, so neither quantization nor architecture explains it alone.

Ruled out:

• Tensor type mix. Qwen3-Embedding-4B is {F32:145, Q6_K:37, Q4_K:216}; E5-Mistral is {F32:65, Q6_K:33, Q4_K:193} — same three types.
• mappedKQuantLongOffsets at true/false/default gave results identical to 7 decimal places.

Remaining difference: qwen3 carries per-layer attn_q_norm/attn_k_norm weights that llama does not — those extra ~80 F32 tensors. Tracked for investigation; it may also affect generative qwen3 Q4_K_M artifacts, which the RAG policy does not test numerically.

Their references are deliberately not committed, since the gate would report a failure for artifacts nobody is publishing.

Verification

:models-bench:check and complianceCheck pass; new tests cover the reference index, digest-based selection, rejection of an unknown artifact, and that every committed reference covers the same probe set.

🤖 Generated with Claude Code