Stop recording a client hang-up as a provider error

[anilguleroglu]

Why

Follow-up to #218. A streamed call whose client disconnects mid-stream lands in the same catch block as a genuine provider failure, so every user who pressed stop was written to the usage log as status: 'error' with usage: {}. Two consequences followed: the model error rate — and alerting built on it — tracked user behavior instead of provider health, and the output tokens the provider had already generated (and billed us for) were recorded as zero.

What changed

Only cancel() aborts the stream's AbortSignal, and only a closed response socket reaches cancel() — so abortController.signal.aborted is an exact test for "the client walked away." That case is split out from the provider-error path:

• A third usage status: cancelled. Both the SQLite and MongoDB aggregations match success/error by exact value, so this is additive — a cancelled call still counts toward totalCalls and the token sums, without landing in either the success or error bucket. No CHECK constraint on the SQLite column and StatusBadge already recognized the value, so no migration.
• The generated output is recorded, not discarded. Provider usage rides on the terminal chunk, which is exactly the chunk a cancelling client didn't stay for. When it never arrives, output is estimated from the streamed text with the same chars/4 rule the quota pre-flight already uses, and the entry is marked output_tokens_estimated so a measured count and a derived one stay distinguishable in cost reporting.
• Stop writing to an already-cancelled controller. The old error path enqueued an error frame and [DONE] into a stream with no reader left, throwing a second exception.
• The output guardrail now runs on cancelled streams too. Those tokens reached the caller before the disconnect; only auditing completions that ran to the end meant hanging up was a way to skip the audit. Both paths go through one auditStreamedOutput, tagged chat.completions:stream:cancelled when the caller left early so the two stay separable in the evaluation log.

UI / reporting impact

Model Hub's Logs and Overview tabs can now show a neutral Cancelled badge. The model error rate will drop for any tenant with meaningful stream-cancellation traffic, since those calls are no longer miscounted as errors — this is the fix, not a regression, but worth knowing before someone looks at a dashboard and asks why the error rate moved.

Verification

• TypeScript: clean
• next lint: clean
• Full suite: 3108 passed, 4 skipped (215 files) — includes 3 new mutation-tested cases (each fails when its guarding condition is flipped to false)
• Docs build: clean
• Production build: successful

🤖 Generated with Claude Code

https://claude.ai/code/session_016vhAPUV6foqg5ap3x6j2K5