fixtures: add the Qwen3-Embedding embed probe corpus so owned-cuda can certify

[Qiiks]

Owned-CUDA has no declared or inherited certification path: ensure_model_certified requires a measured probe row keyed on the machine-profile hash, and the only writer of those rows is probe.start. The embed probe resolves reference fixtures by (family, reference_model), but the shipped embed fixtures only cover minilm and gte-modernbert — so a Qwen3-Embedding model on the owned-CUDA lane dead-ends at:

"gate": "reference_fixture",
"blocking_reason": "reference_fixture_missing",
"reference": { "family": "qwen3-0.6b", "model": "<model_id>" }

with no supported way to certify. This PR adds the missing fixture set.

What's added

• crates/synapse-module/src/fixtures/probe_corpus_qwen3_embedding_fp32.json — 64 items (same corpus texts as the shipped MiniLM fixture so cross-lane comparisons stay like-for-like), reference vectors for Qwen3-Embedding-0.6B, 1024 dims, pooling: "last", normalize: true.
• probe_fixtures() loads the new set and computes its reference dims, matching the existing pattern.
• probe_reference_key resolves the qwen3-embedding model-id family/reference names when the engine build flags carry no family (same fallback shape as minilm / gte-modernbert-base).
• A registry test pinning family / reference_model / dims for all three embed sets.

Reference provenance (the part that matters)

The reference vectors were generated with candle-transformers 0.10.2 on CPU, f32 weights, last-token pooling + L2 normalization, from Qwen/Qwen3-Embedding-0.6B (safetensors revision 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3). The generation path shares no code with the CUDA kernels under test, which is the property the probe needs from a reference. The generation_command field records the invocation; the generator is a small candle-based tool (happy to land it under bench/lanes/ in a follow-up if you want it in-tree rather than as a documented command).

Tokenization follows the module's owned-Qwen3 policy: add_special_tokens=true, pop a trailing terminal, truncate to 2048, push the terminal — i.e. exactly what apply_owned_tokenizer_policy + tokenize_batch produce for this model, so the reference and the probe compare the same byte stream.

Verification

• Live probe vs the owned-CUDA lane (RTX 4050, driver API 13030, CC 8.9): mean cosine 0.99999771, min cosine 0.99999539, worst-decile rank overlap 1.0000 — thresholds are 0.999 / 0.9. This is the first time an owned-CUDA Qwen3 embed model can reach the probe's quality gate at all.
• cargo test -p synapse-module --lib — 423 passed, 0 failed (the new registry test included; one unrelated worker_host test flaked once on this box and passes on re-run).
• cargo fmt --check and cargo clippy -p synapse-module --lib clean.

Scope

Fixture + matching only. I did not touch the certification flow itself; with this set present, a normal probe.start on an owned-CUDA qwen3-embedding model can produce a measured row and serve.

One adjacent gap I noticed while building this, filed separately if you want it: the CUDA capability floor is env-only (SYNAPSE_CUDA_DRIVER_API / SYNAPSE_CUDA_COMPUTE_CAPABILITY), and the driver value is a raw driver-API integer (13030 here), not the marketing version string — nothing probes it automatically, so a correct machine still refuses until the operator sets both. Happy to follow up with a probe that reads cuDriverGetVersion / cuDeviceGetAttribute at startup.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Summary by cubic

Adds the Qwen3-Embedding probe fixture corpus so owned-CUDA Qwen3 embed models can certify instead of dead-ending at reference_fixture_missing.

Changes

• Adds a 64-item fixture with reference vectors from candle-transformers CPU f32 (last-token pool, L2), independent of the CUDA kernels under test.
• Loads the fixture in probe_fixtures() and resolves qwen3-embedding model-ids to the new family/reference names, mirroring the existing minilm/gte-modernbert fallbacks.
• Adds a registry test pinning family/reference/dims for all three embed fixture sets.

Verification

• Live owned-CUDA probe scores 0.99999771 mean cosine (threshold 0.999) and worst-decile rank overlap 1.0.
• cargo test -p synapse-module --lib passes (423 tests, 0 failures) and cargo fmt/clippy are clean.

^{Written for commit 6f7cdca. Summary will update on new commits.}

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[cubic-dev-ai]

No issues found across 2 files

Not reviewed (too large): crates/synapse-module/src/fixtures/probe_corpus_qwen3_embedding_fp32.json (~65,932 lines) - if these are generated or fixture files, add them to ignored paths to exclude them from future reviews.

<sub>Tip: cubic can generate docs of your entire codebase and keep them up to date. Try it here.

Re-trigger cubic</sub>