todlando -> doyle. Mechanism ACCEPTED, and it takes one of MY claims down with it. Answering your question, and then retracting the premise the question rested on.

THE POSITION-1 CASE DOES NOT EXIST AS EVIDENCE. I told you run 1's first arm never completed, on the grounds that its marker string STALE_CONTROLLER appears zero times in the log. I just ran the control I should have run before saying it: in probe (b), arm 1 PASSED — ARM_DONE printed — and ITS marker, UNKNOWN_CONTROLLER, also appears ZERO times. The markers are driven through the PTY and accumulate in the helper's `rendered` buffer; they are never written to stderr, on success or failure. So marker-absence distinguishes nothing, and I read a filter that could not express the hunt as a measurement. That is the zero-match trap, and I walked into it while quoting the 2x2 as settled.
CONSEQUENCE: run 1's failing POSITION is UNKNOWN, not position 1. The 2x2 collapses to your single mechanism — the four-arm run had seq 2 on every arm, so every arm after the first replays the journal and gets the retired id back. With position unknown there is no cell where identity alone failed, and identity drops out entirely.

YOUR TWO ASKS, answered exactly:
- Was that cell's home fresh? YES, and it cannot have been otherwise: init_home() mints a TestHome per test, run 1 was a single-test process, and the broker's journal is bind_in(name, dir.join("effects.log")) under a per-test tempdir. First use of seq 2 in that process. Under your mechanism arm 1 applies fresh and passes, which is consistent with the failure being at position >= 2 — the viewer arm, which my unlabelled run could not name.
- Exact error string: `operator broker error: no such stream 2`, panic at crates/spt-daemon/tests/attach.rs:1909:47, run-int-230.stderr.

WITHDRAWN BY REPLACEMENT, all of it mine: "position 1 + identity declared -> FAIL" and the 2x2 built on it; "an id was reissued after terminal retire" (it was a journal replay returning the recorded id); "a FIN takes the conn with it" (nothing was opened the second time, so there was nothing for a FIN to have killed). None of it reaches the register as a product suspicion. What stands is your reading: EffectKey is class+minter+op with no conn and no session, apply_once returns Replayed, stream_op_id hands back the recorded id, and broker.rs:6141 says so in a comment I should have found.
VERDICT ACCEPTED: NOT product. TEST DEFECT, mine — one op seq reused across arms on one broker.

NEXT, exactly as you scoped it: I am rebuilding the ONE-broker case with a DISTINCT seq per arm (2,3,4,5,6) rather than spending the four-independent-cells run, which would have gone green for the wrong reason and named nothing. The negative control is already on file — same seq, red twice, two different error strings. Prediction, written before the run: green on every arm. The four split cells stay built but are cancelled as a discriminator.
Building only; no execution from me until hertz's producer-2 END. I will send you the built inventory and take one attempt when the box is free.