AFTER / HEAD  MCP SSE timing capture
Commit: beef8702e18bc94522f1e093af52759e91bc7e3d

Linux x64; Node v26.13.1; AMD Ryzen 9 9950X; 8 logical CPUs exposed; same VM, sequential fresh processes; synthetic
data only
Five measured runs; same driver; all returned payloads exactly verified.

responsePayloads -> Stream.runCollect (milliseconds)
MiB  Delivery       Median       Minimum       Maximum
---  ------------  -----------  -----------  -----------
  1  one chunk           28.03        22.85        53.07
  1  256 bytes           17.90        11.43        25.88
  2  one chunk            4.05         3.38         5.94
  2  256 bytes           18.17        15.28        37.26
  4  one chunk            7.89         5.25        13.23
  4  256 bytes           30.06        24.51        98.94

4 MiB / 256-byte individual runs (ms):
  98.94, 34.30, 26.46, 30.06, 24.51

Real acp-mcp-call CLI, loopback HTTP + controlled 256-byte delivery:
  Median 410.70 ms; range 328.43-458.49 ms
  Runs: 353.24, 410.70, 423.09, 458.49, 328.43

Result: exactly 4,194,304 text characters; unchanged tool result.
4 MiB parser budget: 1,500 ms.  PASS

CLI timing includes process startup, MCP handshake, HTTP, parsing and stdout.
Fetch rechunker buffers HTTP bytes, then yields 256-byte chunks immediately.
This is controlled parser / CLI evidence, not an interactive ACP-provider thread.
