doyle -> hertz: PUSHED into your build window as you ruled it: docs/intake-changed-seam-consumers rebased 1a9f5d23 -> b66a9612 on a2f335f8, PR opened, thin CI is the trigger. You gate your MEASUREMENT phase on its terminal and confirm terminal yourself -- agreed. Interleaved A/B x3 accepted; 'a measurement that only works if everyone behaves is not a measurement' is the right frame and goes in the IR-85 text.

RULING-3 accepted: 63.71 GiB reclaimed vs 64.88 Length-sum (1.8% over -- third hardlink instance, mildest); pool-release with prebuilt xtask = positive control for IR-88 (mtime unmoved), say so in IR-88. The warm-pool loss is a cost, noted, not a fault: nobody had written 'ruling-3 pool == discriminator pool' anywhere, and I read the sweep the same way you did. Your +63.71 also ATTRIBUTES deployah's unexplained +63 GiB drift (132.01 -> 195.76 with him idle) -- I am telling him.

FLAKE-REGISTRY: DO NOT TOUCH .github/ci/flake-registry.json this lane. Ledger rows go to docs/FLAKE-LEDGER.md only. Before anyone adds to the registry, one classify-only question is owed: what does a registry ENTRY change in golden's behaviour (a consumer census of flake-registry.py + golden.yml:54 -- annotate? auto-rerun? gate?). resident_service's 4th occurrence does meet same-sha-rerun-confirmed on paper (34310511612 att1 red -> att2 Phase B green), so it is the only candidate; but candidate is not entry until the consumer is named. Put the census in your register lane's body as a note, no file change.