test(runners): two tests no longer race a fast or a loaded runner

[sleep3r]

Why

The release PR's CI (#45) failed once on a drain lets a held hidden-test run finish… (tests/runner-agent-lifecycle.test.mts): ENOENT … attempts/<id>/inputs/test.csv. The same test passed in #46's CI, on macOS and in a Linux node:22 run. A flake hunt on a Linux VM (the feature's 46 suites, six at once, three rounds) then caught a second one: an upload refused in the middle of its body is answered… (tests/remote-runners-doors.test.mts) took 1.6–1.7 s against a 1.5 s bound in three of six runs.

What

• The drain test: the stand-in run ended at once, so on a fast runner the agent could settle it and remove its tree between the test's wait for the run and its read of the run's /data. The run now waits for that read (seam.next = 'wait', then seam.release()), as the sync test's run already does.
• The upload test: what it means is that home does not drain a body it refuses before reading. It now counts the chunks sent before the answer (24 of 1024 locally; the sockets' buffers hold a few megabytes, a drained body all 64) and keeps a 10 s bound against an answer that waits for a timeout.

Nothing outside the tests changes.