Make the SSE reader emit exactly one terminal outcome; add ModelUsage parity

[HartBrook]

Second of the two P0 follow-up PRs. Independent of #13 — this touches only packages/llmpane-react, so the two can merge in either order.

The SSE reader had no terminal state

readSSEStream fired callbacks opportunistically as it walked the buffer, which produced four distinct defects:

Scenario	Old behaviour
Chunk with both error and done	onComplete and onError — the same failed run reported as success and failure
Body closes after a done chunk	onComplete fires a second time
Stream aborted	No callback at all — optimistic messages pending forever
Body closes with no terminal event	Reported as success, though the outcome is unknown

Every exit path now settles through a single guard that delivers exactly one of onComplete, onError, or the new onAbort, and reading stops as soon as a terminal chunk arrives. A body ending without a terminal event is an error. The terminal chunk is still delivered to onChunk first, since consumers read messageId and conversationId off it.

useChat had matching problems

It classified error chunks inside onChunk while also receiving onError from the reader, so consumers saw every failure twice. That duplicate is removed.

Failed and cancelled runs also left both optimistic messages pending forever. They now settle deliberately:

• User message is kept and finalized — the server persists it before streaming begins.
• Partial assistant message is dropped — the server discards partial output, so keeping it on screen would make it silently vanish on the next reload.

Two subtleties the tests forced out:

• The settle helper is keyed off the temp IDs rather than streamingMessageRef, because stopStream() clears that ref before the abort reaches the handler. Keying off the ref left the assistant bubble pending — caught by the cancellation test.
• It is declared before the request, so an abort landing during fetch (rather than mid-stream) settles too.

Protocol parity

Python's StreamChunk has carried a usage: ModelUsage field that the TypeScript contract never exposed. ModelUsage is now part of the React types and exported from the package entry point.

total_tokens is a computed property on the Python model and does not cross the wire, so it ships as a totalTokens(usage) helper rather than a phantom field.

Verification

• 151 React tests pass (was 130); lint, typecheck, and build clean; verified ModelUsage and totalTokens actually appear in the built dist/index.d.ts.
• New coverage asserts callback counts and ordering — not merely "was it called" — across success, terminal error, structured errorInfo, early EOF, empty body, non-ok response, malformed events, [DONE] marker, a terminal split across a trailing partial line, and cancellation.

One existing test changed, deliberately. "calls onComplete when stream ends" fed a stream with no terminal done chunk and asserted onComplete — it encoded the bug. It is replaced by two tests: one for the real success path, one asserting an early close is an error.

Compatibility

onAbort is additive and optional. The behavioural changes are the point of the PR, and worth noting for anyone upgrading:

• If you handled error chunks yourself inside onChunk, you were previously getting onError as well; that double-call is gone.
• A stream that closes without a terminal event now reports an error instead of success.
• Cancelled runs now invoke onAbort instead of nothing.

Checked against a production consumer: it passes neither onError nor onComplete, reading only the returned error state and calling stopStream(), so this is a strict improvement there with no API break — cancelled runs stop leaving permanently pending bubbles in its message list.