hertz -> doyle. RELAUNCH UNDERWAY at 09:02Z, inside your 09:10Z bound. Window stays open.

NEW FROZEN SHA a17dbfbd (a17dbfbd308a4c31db57fb7c4ce2af4f86c2b7d0), "test: a node root is a 302 to its docs, not a page". One file, test-only, no product edit. The assertion now checks BOTH status and Location — FOUND plus location "/local/docs/" — so it cannot pass on a bare redirect to somewhere else, which the status-only form would have. 7b098cc2 is superseded; its 20/21 stays on the record as red-then-superseded, not a pass, as you ruled.

RUNNING NOW, all three, in order: producer 1 (21 cases at the new sha, first attempt at a new sha), then producers 2 and 3 (first attempts, never started). Deadlines 480/420/300s.

INSTRUMENT DEFECT FIXED AND CONTROLLED, not just edited: lane_run.py now reconfigures stdout and stderr to UTF-8 with replacement at import, with the reason written at the fix — a cp1252 stdout raises on nextest's box characters and kills the driver AFTER the producer has written its evidence, losing every later producer in the sequence. That is the exact shape that turned one red into two never-started producers. Positive control run before relaunching: printed U+2500 box-drawing and U+2713 through the fixed driver, exit 0. I did not want to relaunch on the assumption that a one-line encoding change worked.

For the register, the mechanism in one line: a driver that prints a producer's output can destroy the REST of a run while leaving the finished producer's evidence intact, so the surviving evidence looks complete and the missing producers look like they were never queued. The tell is that evidence files exist for producer N and no files at all exist for N+1 — not a failure, not a skip, an absence.