MCP SSE performance evidence

Linux x64, Node v26.13.1, same VM, sequential fresh processes.
Base 0597532f16d54189c4a696a249cc66c6d44c7ab5
Head beef8702e18bc94522f1e093af52759e91bc7e3d

Install each checkout with vp i first.
Parser reproduction (repeat unchanged per checkout):
  node measure-sse.mjs CHECKOUT OUTPUT_DIRECTORY ROUND
CLI reproduction (repeat unchanged per checkout):
  node measure-cli.mjs CHECKOUT OUTPUT_DIRECTORY ROUND
Keep rechunk-fetch.mjs beside measure-cli.mjs.
run-comparison.py and run-cli-comparison.py alternate base/head order.
Their checkout paths are examples under the caller's home/wt directory.

Parser medians use rounds 2-6; round 1 was a pilot and is excluded because
it overlapped an initial exploratory run. The five reported rounds were
sequential. CLI medians use all five rounds. results.json contains every
reported raw sample and exact-payload validation. Captures are rendered
measurement output, not screenshots of an interactive thread.

The parser workload matches the PR's supplied harness: 1, 2, and 4 MiB
JSON-RPC text results in SSE, each delivered in one chunk and 256-byte
chunks, timed from Response construction through Stream.runCollect.
Exact text and envelope are verified outside the timed region.

The CLI workload invokes the real apps/server/src/bin.ts acp-mcp-call
command against a synthetic loopback HTTP MCP server, with initialize,
notifications/initialized, and tools/call. Since TCP coalesces writes,
a fetch wrapper buffers the HTTP response and delivers immediately
available 256-byte chunks to the unchanged production parser. CLI time
includes startup, handshake, HTTP, parsing, and stdout. This is not an
interactive ACP-provider thread and does not depend on real model behavior.

No tests were used as evidence. No production data or credentials used.
