Previously, ModelProviderRequest::get_cache_key() serialized the request twice: once via serde_json::to_value to produce a serde_json::Value, then again via .to_string() to get a JSON string for hashing. Between the two steps it also had to allocate the intermediate Value, find and remove the inference_id key from it, and then allocate the final String buffer.

This PR replaces that with a single streaming serialization pass directly into the blake3::Hasher (which implements std::io::Write), using serde_json::to_writer. To exclude inference_id from the hash without an intermediate allocation, it is now marked #[serde(skip)] on ModelInferenceRequest.

What changed

• tensorzero-inference-types

  • Added #[serde(skip)] to the inference_id field on ModelInferenceRequest. The field is only serialized for cache key purposes and is intentionally excluded from the hash, so skipping it during serialization is correct and safe.

• tensorzero-core/src/cache.rs

  • Replaced the three-step to_value → remove → to_string sequence with:

    serde_json::to_writer(&mut hasher, request)

• Added a Criterion benchmark: benches/cache_key.rs for regression tracking.

Benchmark results (release mode, same machine)

✅ All 88 cache unit tests pass.