Read fragmented HTTP status lines in the job tests

[O6lvl4]

The oversized-request check in job_server_checks.py assumes one recv(4096) returns the HTTP status. In the Ubuntu CI run for #46, it received only HTTP/1.1 and rejected the response before the status code arrived.

Read through the status-line CRLF across TCP fragments, bounded by one 10-second monotonic deadline and 4096 bytes. EOF and timeout remain failures. The request still sends headers only, so the check continues to prove that an oversized body is refused before being read. No server or application code changes are included, and this branch starts from develop independently of #46.

Add 10 deterministic harness tests and run them on both CI platforms before the build. They cover fragmentation within the status and CRLF, early return without the response body or EOF, unexpected status, premature EOF, socket timeout, a shared deadline, and the size bound.

Verification:

• All 10 new harness tests and all 17 existing evaluator unit tests pass.
• Replacing the reader with the original single recv fails the fragmented-response regression; the new reader passes. A separate localhost TCP test uses a handshake to force the same fragment boundary and also proves old-fail/new-pass without timing sleeps.
• almide check src/main.almd, Python compilation, git diff --check, and the pinned code-quality gate (90/A) pass.
• The full local job suite and an isolated real-server check stop at startup because this executor has no usable Landlock support (ENOSYS), with os_sandbox = "required" unchanged. Those attempted checks reused the existing 0.6.17 binary from Escape HTTP error responses as valid JSON #46, whose job service is unchanged. Full platform integration remains for CI.