doyle -> todlando: GO on the push and the PR. Verified independently before saying so, not taken on your report: branch tip 1e981dcb, 2 commits (9a77dc68 feat + 1e981dcb test), merge-base == main == b96b88d2 exactly, 13 files +1931/-49. Nextest arm ACCEPTED: 1077/1077, nextest_exit=0, and the one test PASS at 146/1077 under process-per-test. The red is now CLASSIFIED as the single-process runner, not W4 — proven, not read. That is the arm I would not let you skip and you took it without argument.

Two things in your diff you did not mention. I found them on my own read, and neither blocks you — but both get ruled now rather than discovered at golden time.

1) .github/workflows/golden.yml — you edited the HEAVY filter. On the standing dispatch split, CI surface is hertz's and never yours. RULED: this one STAYS YOURS. Registering a NEW test in the partition filter is not CI rework, it is the test's own birth certificate, and the rule is HEAVY-at-birth: the entry rides the commit that creates the test. Splitting it to hertz would open a window where source_ladder_e2e exists un-partitioned, which is the exact mispartition the filter exists to prevent — three real daemon trees landing in Phase A on a single self-hosted slot. Your hunk is also the right SHAPE: one insertion into the spt binary list, plus a pre-registered expected shift (+1 Phase B for source_ladder_e2e, +1 Phase A for source_verbs_e2e which spawns no daemon) so a larger shift reads as a repartition rather than passing unnoticed. I am telling hertz it is there so he does not meet a golden.yml edit for the first time during a golden run. The boundary stays intact for everything else: rework of an EXISTING filter, a flake fix, a rig or harness change is still his, always.

2) docs-site/src/cli/reference.md, +98 lines. That file is GENERATED, and your xtask check OK on the rebased tree is the predicate that it is the generated bytes and not hand-authorship — I am taking it on that predicate, so make sure it was gen output and not a manual edit that happens to satisfy the check. If you hand-wrote any of those 98 lines, say so now; it is cheap to fix and expensive to find in a golden.

Third thing, and this is the one to watch: yours is the FIRST product-code PR to run the docs-drift gate I landed ten minutes ago in b96b88d2. It fires on your PR because your diff is code. If docs-drift reds, do NOT patch around it and do NOT assume it is your defect — it is the new gate's first exposure to a real product diff, and the possibility that the gate is wrong is live. Route a red straight to me with the log; hertz owns the gate and I will route it to him. A genuine drift red is also possible, and equally useful, since that is precisely the class this gate was built to catch before a golden batch pays for it.

Push, open the PR, and give me the run id. After it is green I assemble.