Files
funkpopo faf8d558ee perf(proxy): frame SSE events in place across all stream transforms
Only the logged passthrough stream got batched SSE framing; the 13 other
call sites still went through take_sse_block, which allocates a String per
event and shifts the rest of the buffer once per event. The whole-body
aggregators (Responses/Chat SSE -> JSON, Anthropic SSE -> message) copied
the full body first, so draining was O(events x body).

- Add SseBuffer for streamed bodies: events are slices of the buffer,
  consumed text is dropped once per chunk, and a scan that found no
  delimiter resumes near the end instead of rescanning a long partial
  event. It owns the split-UTF-8 remainder, and take_tail() keeps the
  existing lossy flush at EOF.
- Add SseBlocks, a borrowing iterator with remainder() for complete
  bodies, so aggregators no longer copy or drain the body.
- Port every take_sse_block / find_sse_block caller and remove both.
  Event boundaries, the trailing-event fallback and clear() semantics
  are unchanged.

Release microbenchmark (old main / previous PR commit / this commit):
  4096 events in one chunk      668ms / 15.9ms / 0.15ms
  64 events x 512 chunks       91.7ms /  2.7ms / 0.97ms
  2 MiB event, 16 KiB chunks    111ms / 10.8ms / 1.6ms
  whole body, 20k events        15.3s /  1.26s / 0.49ms

cargo clippy --lib -D warnings is clean; cargo test --lib: 4305 passed.
2026-11-03 11:42:08 +08:00
..
…
2026-11-01 22:30:02 +08:00