[mateusands]

Two defects found by an external review of c726aa4. I reproduced both before fixing them, and the plan review found a third.

Defect	Cause	Fix
Different requests could share a cache key (false hit)	The canonical copy was a plain {}: an own __proto__ key replaced the prototype and vanished, and toJSON was ignored, so every Date hashed as {}.	Null-prototype copy; toJSON followed like JSON.stringify (only when it is a function); other non-plain objects (Map, Set, RegExp, class instances) are unkeyable and bypass the cache.
A BigInt or a cycle in params made wrap() reject instead of calling the model	deriveKey threw and no caller caught it, breaking fail-open.	deriveKey turns any failure into a bypass.
An OpenAI stream replayed from the cache lost logprobs, system_fingerprint and service_tier	The assembler stored logprobs: null and kept only id / created / model.	Accumulate like ChatCompletionStream (first logprobs copied, later content appended; last value wins for the top-level fields) and replay them.

KEY_VERSION stays 1. Plain JSON, which is what the SDK wrappers send, hashes exactly as before. A test pins the hash computed on c726aa4 for a typical request with system prompt, prior turns, tools and nested params. Only inputs that used to collide get new hashes; their old rows simply miss.

Testing

• 249 tests from a clean install. Lint, typecheck, build and the consumer check pass. The integration suite is also green on pgvector 0.7.4 / PG17 and 0.8.0 / PG13.
• Every new test failed before its fix.
• Guards checked by breaking them on purpose:
  • going back to a plain {} copy fails the __proto__ test;
  • not appending later logprobs fails the replay test.
• The replay test compares against the SDK's own ChatCompletionStream.finalChatCompletion().

Code review follow-up

An independent code review found six more points, fixed in 8831524 and 65e89c8:

• the catch now covers only serialization: a malformed input throws again, and a bypass reports why to onError with a new key stage;
• plain objects from another realm (vm, jsdom) are hashed again;
• toJSON is applied once per value, as JSON.stringify does;
• a null service_tier survives replay;
• comments and the changelog are corrected.

Each fix has a test, and the review re-checked both commits. The golden hash still holds. CI green on 65e89c8.