fix(sglang): stream tool_call deltas incrementally instead of bufferi…

[guojn1]

…ng until finish

Client-side TTFT of pure tool-call streaming requests was effectively equal to the whole generation time: the post-processor only attached accumulated tool_calls to the finish chunk, and content-less deltas were suppressed, so no bytes reached the client during generation (Nexus TTFT/total median ~99.6%/97.4% vs ~17.2% for native SGLang).

The post-processor now emits OpenAI tool_call deltas as the parser produces them: the id+name delta goes out with the first argument fragment and argument fragments stream as they arrive (the un-emitted tail of the accumulated arguments is the source of truth, so event granularity cannot produce duplicates).

Correctness guards from the buffered design are retained:

• names without any argument fragment are never emitted, so a misidentified prompt word cannot leak as a dangling call;
• unknown tool names (not in the request's tool list) are suppressed mid-stream and purged at finish, exactly as before;
• the finish-time full-text re-parse remains authoritative, now in merge mode: argument suffixes for partially streamed calls are patched (with streamed indices kept stable), calls the streaming parser missed are appended under fresh indices, and when nothing was streamed the re-parse rebuilds state exactly as before;
• malformed (non-JSON) accumulated arguments of already-streamed calls can no longer be retracted, so they now trigger the finish-time re-parse for authoritative recovery instead of being silently dropped; finish_reason is still rewritten to tool_calls when a call was emitted.

Tests: existing assertions now reassemble the delta stream per the client contract (semantic regression), plus new cases locking the TTFT contract (deltas arrive before the finish chunk, first entry carries id+name, distinct indices/ids for parallel calls) and the new malformed/unknown-name streaming contract. Not executed locally (no sglang env); run on gpu_1 CI.

Overview:

Details:

Where should the reviewer start?

Related Issues

⚠️ This section is required. Choose one path below and delete the other.

🔗 This PR is linked to an issue:

• Closes #XXXX

🚫 This PR is NOT linked to an issue:

• Confirmed — no related issue