[rn7s2]

Every lane of an evo-swarm on Windows died about a second after it started listening, and the coordinator reported each one as:

[lane 1] init-failed — ERROR 12029: Cannot connect

12029 is ERROR_WINHTTP_CANNOT_CONNECT, so this read as a networking fault — but the lane's ready.json and its own lane.log both showed a listener that had come up and then gone. The lane's process was gone before the coordinator's first request could be answered.

Cause

A lane is launched with --watch-stdin and stdin a pipe the coordinator holds, so a lane can never outlive the swarm: end of file on that pipe is its shutdown. watch-stdin-eof parks a thread on the stream make-stdin-stream returns and reads it with read-byte.

On Windows make-stdin-stream returned a character stream whenever stdin was not a console — which is exactly a lane's case, a pipe. read-byte on a character stream signals a type error rather than returning a byte, and the handler around that read treated any error as end of file, so every lane shut itself down at once.

Measured, with the lane's stdin a pipe held open by the parent:

	with --watch-stdin	without
old code	exits within ~1s	stays up
fixed	stays up	stays up

Changes

• make-stdin-stream's Windows pipe branch is now a byte stream (:element-type '(unsigned-byte 8)), the only kind its one reader can read. The console branch is untouched: on a console it still returns the console input handle, which drain-console-input reads as key events. The character branch had no remaining caller — read-available-input's Windows character path only ever saw a console handle — so untranslated-newlines and that branch are gone, and read-available-input reads bytes everywhere but the console.
• watch-stdin-eof calls its thunk only on a genuine end of file. A read that fails for any other reason leaves the thread with nothing to watch and returns; this signal is a lane's whole life, and a stream that cannot be read is not a parent that went away.

Verification (real Windows lanes, not the harness)

• With --watch-stdin, the lane stays up and answers.
• Both lanes of a two-lane swarm reach idle, restarts: 0, with a ready model.
• A delegated task runs in the lane with a real model call.
• The contract the flag exists for still holds: closing the coordinator's end of the pipe shuts the lane down cleanly (exit 0, ready file deleted).
• test-stdin-stream-for-watching fails on the old element type and passes on the new one.

Co-authored-by: EvoAgent evo@ruiqilei.com

[rn7s2]

The build-and-test (ecl) failure here is pre-existing on main, not from this PR.

Unmodified main at 449363e fails with the identical ECL internal error:

Internal or unrecoverable error in:
Unable to mprotect environment.
  [22: Invalid argument]
Aborted (core dumped)
make: *** [Makefile:74: test] Error 134

Reproduced by dispatching CI on main itself (run 39290253272): build-and-test (sbcl) and build-and-test-windows pass, build-and-test (ecl) fails the same way. No test assertion fails in either run — ECL dies inside its own compiler/GC (cl_read_line → mprotect), after the OAuth unit tests, which is a known ECL 24.5.10 failure on newer kernels that restrict PROT_EXEC.

build-and-test (sbcl) and build-and-test-windows both pass on this PR, which are the configurations this change actually affects (the Windows make-stdin-stream branch is SBCL-only — ECL is not supported on Windows).

Co-authored-by: EvoAgent evo@ruiqilei.com