fix(llm): treat empty tool call identity in stream deltas as absent - #40969

Problem

Streaming tool calls through an OpenAI-compatible endpoint fails with OpenAI Chat tool call delta is missing id or name on providers that send id as an empty string (instead of omitting it) on continuation deltas. Alibaba DashScope (*.maas.aliyuncs.com/compatible-mode/v1, also serving qwen3-max / deepseek models) does this. The session errors out mid-turn and the pending tool call is never resolved.

Actual chunks captured from DashScope (qwen3.8-max):

data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_a1ad8367bde7444f82142e8e","type":"function","index":0,"function":{"name":"bash","arguments":""}}]}}]}
data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"","type":"function","index":0,"function":{"arguments":"{\"command\": "}}]}}]}

The first delta carries the real identity. Continuation deltas carry id: "" with no function.name. In ToolStream.appendOrStart, delta.id ?? current?.id keeps the empty string (it is not nullish), so the fallback to the accumulated tool never happens and the truthiness guard reports the identity as missing.

Fix

Fall back on empty-string identity the same as on absent identity (|| instead of ??). A genuinely missing identity on the first delta for an index still errors as before.

Tests

Added a ToolStream regression test replicating the DashScope shape: identity on the first delta, id: "" on the continuation delta. bun test in packages/llm: 299 pass, 0 fail.