fix(simple-engine): preserve streaming finish reasons

[Thump604]

Natural generator exhaustion currently emits a final chunk with no finish reason. Strict OpenAI clients treat that as a truncated stream.

This stamps natural exhaustion as stop. It also fixes the engine-enforced token-limit fallback so a backend chunk with no reason is reported as length, while preserving explicit backend reasons.

Tests cover natural and empty exhaustion, token limits, exceptions, and request cleanup.

This adopts and extends #629 because the required regression patch could not be added to its fork-owned branch. Miguel Vilhena is credited as co-author.

Verification: 2,301 passed, 24 skipped.

Refs #628

[janhilgard]

Approving. Reproduced both halves against main and confirmed the fix, driving _stream_generate_impl with a backend that exhausts without ever emitting a finished chunk — the shape #628 describes.

main   natural exhaustion  ->  finish_reason=None   finished=True
main   token limit         ->  finish_reason=None
#681   natural exhaustion  ->  finish_reason='stop'
#681   token limit         ->  finish_reason='length'

Both were None on main, so a strict client sees every stream as truncated, not just the natural-stop case. The issue title says "length works", which is true at the API level but not here — the length case was equally unset at this layer.

The root cause is worth spelling out because it is easy to reintroduce: getattr(chunk, "finish_reason", "stop") supplies its default only when the attribute is missing, never when it is present and None. Backends set it to None routinely, so the default was dead code. Splitting the lookup from the fallback is the fix, and deriving length from completion_tokens >= max_tokens rather than assuming stop is the right way round.

Crediting Miguel Vilhena as co-author while adopting #629 is the right handling of a fork-owned branch.

Merge note

This conflicts with #679, which is approved and touches the same block. The conflict is textual only — #679 re-indents that region into a _generation_worker_in_use() context and rewrites the tail's terminal chunk — but the semantic resolution needs stating, or a rebase will silently drop one intent:

• #679 yields finish_reason="abort" from that tail when a stream is cut short by stop().
• #681 yields "stop" from the same tail on natural exhaustion.

Combined, it should be "abort" if aborted_by_stop else "stop". Picking either branch wholesale loses the other. Merges cleanly against #690 with no conflict.

CI green on all nine jobs.

[waybarrios]

I carefully reviewed and it is all set. Dont see any major flaw or issue here.