Commons demote pass: move mechanism out of the auto-loaded files behind lazy anchors
Commons demote pass, GO from pmmaster relaying Elazar's standing grant 2026-09-11 after blackboard bb-2026-09-11-fleet-token-priority. DEMOTE-ONLY: the auto-loaded file keeps the RULE, mechanism and measurements move behind a named anchor in the lazy-loaded evolutiva-rulings-detail.md. Nothing deleted, nothing summarised. WHY: 107,238 bytes of instruction text auto-load on every one of my responses, 50,960 of them (evolutiva-commons + evolutiva-coder-commons) on every coder and audit response fleet-wide. That is pure cache_read every turn whether or not the rule fires - roughly 90M cache_read tokens per 6h window across the fleet. Measured BYTES, not realised saving; cache residency across a session is not readable from here, so the bound is the direction and the leverage, not a percentage. ROUND 1 EDITED AND VERIFIED ON DISK, NOT YET PUSHED (venus-host.push-gate held by coder-pluto-cc since 14:38Z, 2400s TTL; registered for the release DM, not polling): 1. mars/shared/md/mars-commons.md - the markdown-only-push paragraph cut from ~1730 to ~840 bytes. Rule kept (freeze + branch publish, the git log origin/main..HEAD pre-read, and that the read narrows but cannot close the window). Both measured incidents (59c63ad/3763552, the de8dc45/96df60d race) and OPS-204 moved to evolutiva-rulings-detail.md#mars-git-push. 2. shared/md/evolutiva-coder-commons.md - Push-to-Done item 7's reroute bullet keeps the rule and sheds the mechanics; the verify-outside-the-lock paragraph sheds the mars/pluto independent-discovery provenance. Both landed in evolutiva-rulings-detail.md#push-to-done as two new subsections, IN FULL. PUSH ORDER WHEN THE GATE FREES, and it matters: the SHARED repo (pensanta-websites: evolutiva-coder-commons.md + evolutiva-rulings-detail.md) goes BEFORE mars-commons.md, because mars-commons now points at #mars-git-push content that only exists in the shared edit. mars-commons.md last, markdown-only. DISCIPLINE THIS PASS ENFORCES ON ITSELF: corrections are folded into the rule they correct, never appended as a second narrative layer - that append habit is what re-grows the file between passes. Verified after round 1 that both demoted blocks are present in the detail file; an earlier version of this pass DELETED rather than moved, which is the failure mode to check for on every subsequent round.
Activity
-
GATE YIELDED to coder-terra-cc, 14:54Z. I will not call lock() on venus-host.push-gate until terra confirms its push landed, so MARS-717 (two commons demote pushes) and MARS-677 (image-api.ts + tests) stay queued behind that. Why, recorded because it is a precedent and not a courtesy: terra is in an APPLIED-SCHEMA / UNPUSHED-APP split - terra_231 added terraPracticas.createdVia and is live as of 14:39:45Z, while the app half sits committed-unpushed at 7cd39a0, so every práctica created through the form is written createdVia NULL and becomes permanently indistinguishable from the six historical rows. That is unrecorded data accumulating with a clock on it. My three queued pushes are a telemetry discriminator and two doc demotes; nothing I hold loses data by waiting. The asymmetry decides it, not queue position. MECHANISM WORTH KEEPING, three agents hit it independently within one hour (coder-terra-cc, db-pluto-cc, coder-venus-cc all DM'd me): pmmaster's named-handoff ruling fixes the HEAD of the order and leaves the TAIL unnamed, and an unnamed tail is the same open race one position later. That is precisely how I took the gate at 14:48 while three agents with older claims were waiting - without breaching the ruling. A first-caller-wins primitive cannot carry an ordering that lives in messages, and the longest waiter loses because it has the most rounds in which to lose. Amendment proposed to pmmaster: the holder names a successor who has ASKED (the 409 already reports who is waiting), and an agent with an open apply-then-push split declares it and goes first. MY OWN ERROR IN THE SAME WINDOW, so it is not repeated: I piped gitpush.sh to head -40, which closed the pipe and killed it via SIGPIPE mid-run. The wreckage - ahead 0, version unbumped, --only files left STAGED - is byte-identical to a pre-hook abort, and I read it as one. No error text ever appeared because the process died before printing any. Files unstaged immediately (staged files on a shared tree are exposed to the next lane's commit). No version burned; origin/main remains 7e839a6. Redirect gitpush output to a file and tail it; never pipe it to head.
-
Round 1 shipped in two commits, ordered so the citation never dangled. Shared repo pensanta-websites 17c8723 (evolutiva-coder-commons.md + evolutiva-rulings-detail.md, ONE commit so a gate handoff could not split the pair), then mars 6e559a4 (shared/md/mars-commons.md). That order was load-bearing: mars-commons now cites evolutiva-rulings-detail.md#mars-git-push, whose content only exists in 17c8723, so landing the mars half first would have published a citation to text nobody could read. WHAT MOVED, demote-only, nothing deleted or summarised: - mars-commons.md Git / Push Rules, markdown-only paragraph: ~1730 bytes to ~840. Rule kept whole (freeze vs branch-publish, the git log origin/main..HEAD pre-read, and that the read narrows and cannot close the window because handoff commits are not taken under the push flock). Both measured incidents (59c63ad publishing coder02-mars-cc's handoff 3763552 under a frozen 2.79.32; the de8dc45/96df60d race) and OPS-204 now live behind #mars-git-push. - evolutiva-coder-commons.md Push-to-Done item 7 reroute bullet and the verify-outside-the-lock paragraph: mechanics and the mars/pluto independent-discovery provenance moved to #push-to-done as two new named subsections. CHECK RUN THIS ROUND AND OWED ON EVERY FUTURE ROUND: an earlier draft of this pass DELETED rather than MOVED - it cut two blocks from the lane file that never landed in the detail file. Caught before push. Demote-only means the text is PRESENT in the detail file, not merely absent from the auto-loaded one; verify the landing, not the removal. Doc-only push discipline applied to itself: ahead range read before 6e559a4 was 0 commits, so the version instrument is inert for that deploy by design and nothing rode out under a frozen version. Stated rather than reported as a clean version match - which is exactly what the paragraph being edited requires. BOUND, unchanged and not to be inflated: I measured BYTES, not realised saving. Cache residency across a session is not readable from here, so a 20 percent byte cut is not automatically a 20 percent token cut. pmmaster asked for the after-measurement on cost_index rather than file size; that is a separate read once the fleet has turns on the new files, and it is not claimed here. Round 2 candidates are not queued - this row closes on round 1. The monthly first-Monday prune pass owns the rest.