fix(streaming): handle an initial UTF-8 byte-order mark

[sylvesterkaczmarek]

Fixes #1982.

Track the first decoded SSE line and remove at most its single leading U+FEFF before parsing fields. The shared decoder covers synchronous/asynchronous and stable/beta clients. Subsequent U+FEFF characters are preserved. This adds a first-line normalization block without rewriting event framing, resource methods or public signatures.

Validation

The 48 fixture-based cases cover prefixed/unprefixed streams, event/data as the first field, one-byte/two-byte/larger transport chunks and preservation of later U+FEFF characters. Against unchanged main: 24 failed, 24 passed.

Final validation ran ./scripts/test for tests/test_streaming.py and both tests/lib/streaming/test_messages.py and test_beta_messages.py, followed by the script's MCP v2 checks, on Python 3.14.7. Repository formatting, ./scripts/lint and git diff --check passed. Earlier related runs also exercised Python 3.13.15 with Pydantic v1/v2; the Python 3.10 floor was not validated locally.

Tested on macOS arm64 with mocked HTTP and recorded fixtures. The complete generated API suite and live services were not run. No dependency, lockfile, workflow or release-metadata changes.

Reference: https://html.spec.whatwg.org/multipage/server-sent-events.html#event-stream-interpretation. The Go SDK has a separate initial-BOM PR at anthropics/anthropic-sdk-go#455.

[sylvesterkaczmarek]

The regression and related tests and the repository lint checks pass locally, as detailed in the PR. Could a maintainer approve the pending Actions run?