fix(openai): require terminal stream events before finishing

[ethanndickson]

Context

#232 gated the Chat Completions text Stream and the Responses API text Stream on fantasy.NewIncompleteStreamError(). This PR closes two remaining gaps that those guards don't cover, plus a related misclassification.

What this PR fixes

1. Chat Completions object stream — silent truncation. streamObjectWithJSONMode has no terminal-event check today. A truncated stream where any prior chunk parsed cleanly emits Finish with the partial lastParsedObject — the caller receives a JSON object missing fields, reported as success. Fix: track sawFinishReason; otherwise yield fantasy.NewIncompleteStreamError() (retryable *fantasy.ProviderError with Cause: io.ErrUnexpectedEOF) with ctx.Err() precedence.

2. Responses API object stream — misclassified truncation + silent-finish risk. Post-loop wraps io.EOF from stream.Err() via toProviderErr, which produces a non-retryable error (IsRetryable() matches io.ErrUnexpectedEOF, not bare io.EOF), so transient truncation surfaces as a fatal provider error and retries never engage. And with no terminal-event check beyond the error event, an SDK that returns nil on a clean mid-message close flows straight to Finish with a partial parsed object. Fix: track sawTerminalEvent (set on response.completed / response.incomplete), treat io.EOF from stream.Err() as a fall-through to that check, and emit fantasy.NewIncompleteStreamError() when neither arrived.

3. response.failed misclassified as transport truncation on both Responses streams. Today the event falls through, fails the terminal-event check, and gets retried as if it were a network truncation. Fix: handle response.failed explicitly, yielding a non-retryable *fantasy.Error{Title: "response failed"} with the provider's error message and code. The pre-existing top-level error event is normalized through the same helper as *fantasy.Error{Title: "response error"}, so only synthetic stream truncation is wrapped with io.ErrUnexpectedEOF.

Two smaller refinements ride along: the existing Chat Completions text Stream guard now respects ctx.Err() precedence, and the Responses API text Stream's existing FinishReasonUnknown-sentinel guard is renamed to an explicit sawTerminalEvent for clarity (semantics unchanged).

Tests

Three new table-driven tests cover happy path + each missing terminal signal + (where applicable) response.failed / error events:

• TestChatCompletionsStreamObject_RequiresFinishReasonBeforeFinish
• TestResponsesStream_RequiresTerminalEventBeforeFinish
• TestResponsesStreamObject_RequiresTerminalEventBeforeFinish

The object-stream truncation subtests would emit Finish on main today.

Prior art

OpenAI's own openai/codex (Apache-2.0) implements the same gate for the Responses API: codex-rs/codex-api/src/sse/responses.rs handles response.completed, response.incomplete, and response.failed as distinct terminal events, and emits ApiError::Stream("stream closed before response.completed") on early EOF. Regression test stream_no_completed.rs::retries_on_early_close asserts the retry path fires; an inline error_when_error_event test covers response.failed end-to-end. vercel/ai (Apache-2.0) makes the same isResponseCompletedChunk / isResponseFailedChunk distinction in its OpenAI Responses adapter.

Follow-ups

Similar terminal-event gaps exist in a few of the other providers in this repo; those will land in separate PRs so each can be reviewed in isolation.