The 26.10.1 stages run their repetitions in two passes: reps 1 to 3 first, 4 and 5 later #223

The user's plan (2026-10-06, DECISIONS #164): every arm runs reps 1 to 3 first and reps 4 and 5 in a later pass, lane by lane, so tiers 1 and 2 have complete N=3 tables about nine days earlier and N=5 lands at about the same date as before.

make_2610_stages.py --rep-pass N runs reps 1..N of each cell's declared REPS. --reps stays the declared total in every runner call (the rows say reps=5 whatever pass made them, and the runner's shuffle draws from it); --only-reps names the pass. A cell that declares fewer reps than the pass runs its own, rep 1 still goes first and a failed rep 1 still ends the arm, and the multipass overlay follows the same rule. Stage ids take the prefix qP so no earlier chain's ALL-DONE marker can start the new one. Without the option the generated scripts are byte-for-byte what they were.

PROTOCOL.md says how a table landed between the passes discloses n=3. The pass-2 generator (roster from the rows: arms that finished reps 1 to 3 cleanly) is a later change.