fix(cuda): advertise enforced embedding batch limits for owned-cuda

[Qiiks]

Summary

recommended_batch_for_engine returns None for owned-cuda, so models.list omits recommended_batch for CUDA models. Consumers that require the field reject the envelope (AFT's batch pager refuses discovery — reproduced end-to-end here: AFT failed model discovery with no usable recommended_batch rows/token_budget, and a normal semantic search made zero embedding calls as a result).

This is the same omission PR #19's review surfaced; it was kept out of #19 deliberately. Alfonso's 4a03d6bf landing covers owned-metal + ane; the owned-cuda arm fell through.

Change

• One-line producer fix: owned-cuda joins the owned-metal arm — rows=MAX_ENGINE_BATCH_ITEMS (8), token_budget=DEFAULT_ENGINE_BATCH_TOKEN_BUDGET (3072). These are the module-enforced limits, matching what the worker actually admits; no invented values.
• Regression: extends recommended_batch_policy_uses_engine_constants_and_omits_unknown_advice to assert the CUDA arm serializes with those exact constants (fails pre-fix, passes post-fix).

Verification

• cargo test -p synapse-module --lib recommended_batch — 1 passed (fails on master: no CUDA arm).
• cargo clippy -p synapse-module --lib -- -D warnings clean; cargo fmt --all -- --check clean.
• Live end-to-end: AFT on native Synapse indexed and retrieved (refund_cancelled_purchase, semantic_score 0.5837) only after the deployed module carried this arm; pre-fix it refused discovery.
• No GPU benchmark; rebased on current master (6015c48); Cargo.lock untouched.

Real consumer proof already posted in the #19 thread (issuecomment-5715898282); this PR is the upstreamable form.

---

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

---

Summary by cubic

Fixes CUDA models missing recommended_batch in models.list because recommended_batch_for_engine returned None for owned-cuda. The engine now returns the same enforced batch limits as owned-metal (8 rows, 3072 token budget), so consumers that require the field no longer reject the envelope.

• Adds a regression test asserting the CUDA arm serializes those exact constants.

^{Written for commit cefe807. 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.

[synapse-alfonso]

Verified and landing. Confirmed at source that the advertised values are the enforced ones: the batch planner clamps every engine to MAX_ENGINE_BATCH_ITEMS (8) and DEFAULT_ENGINE_BATCH_TOKEN_BUDGET (3072) with no engine branch, so a CUDA row gets exactly what it advertises. Mutation-checked: with the arm reverted, the extended test fails at the CUDA assertion; with it, passes. Clippy clean, CI green, Cargo.lock untouched.

Thanks for tracing the consumer failure end to end — a producer omission that reads as "no error" on our side and "no embedding calls" on AFT's is exactly the kind of silent gap worth a regression test.