--- name: partial-milestone-golden-is-rejected-alternative description: "A tranche/partial-milestone golden head is ADR-0050's rejected eject-and-proceed wearing a new name; the intake gate caught it at LOCKSMITH t1 2026-08-03" metadata: node_type: memory type: project originSessionId: f4e862c6-83b6-41e5-b2f2-66fcf52663ac modified: 2026-08-03T16:25:42.830Z --- **2026-08-03, LOCKSMITH tranche 1** — doyle handed `golden/locksmith-t1 @28f8dd4` for golden. I held the push. Both findings stood; doyle upheld the hold and ruled the root was his. ## The mechanism: a rejected alternative re-enters under a new noun ADR-0050 rejects "eject-and-proceed" — shipping a batch minus a conflicting PR — with a GENERAL rationale, **"a milestone ships whole"**, not a conflict-specific one. A head carrying 4 of the milestone's 10 greenlit requests is that same shape from a different cause. doyle: *"I built the rejected alternative and did not notice because I called it a tranche."* ⭐⭐ **The rejection's RATIONALE, not its trigger, is what generalizes.** Search a ruling by the shape it forbids, never by the situation that produced it. A new noun ("tranche") is exactly how a ruled-out shape gets re-proposed without anyone re-litigating it. Stated at its real size: ADR-0050 is not a blanket ban on sub-milestone goldens — it carves out "an urgent fix is a golden batch of one". A PARTIAL milestone is a **third shape**: neither whole nor a batch of one. Unruled, and shaped like the rejected one. A gater does not mint a new process object to make his own head legal — it went to the operator, since milestone scope is an operator greenlight and board writes go through alchemy. ## The two checks that caught it (both cheap, both nearly skipped) 1. **Greenlit-form parity against the AUTHORITATIVE set** — `gh api repos///issues//sub_issues`, not the milestone's comments and not `alchemy list` (greenlit alone ≠ membership; #133 was greenlit and NOT a member). Head carried #70/#115/#117/#134; absent were #123 + five tranche-2. Zero drop comments existed. A dispatch comment that pre-records a tranche split records a sequencing of WORK — it is **not** a statement that the milestone's golden ships without them. 2. **Release-shape, and it is UNREPAIRABLE after the run.** Head `Cargo.toml` = `0.52.0` = `origin/main` = the already-PUBLISHED v0.52.0; `## [Unreleased]` empty; no bump/lockfile/ changelog in the delta. Under [[release-standard-bump-in-pr]] the bump RIDES THE BATCH because there are no post-golden edits — so adding it afterwards invalidates the run. Costs two `git show`s and kills a whole window's worth of wrong work. ## ⭐⭐ Check FULFILMENT, not presence I spot-checked #70 against its grill ruling rather than its commit subject: the monic-id re-key, `LegacyMonic` + `migrate_owner` (the W5 record migration the operator's ruling required), read-time migration, unreadable-record listing. doyle adopted this as the re-intake standard and named why: **"checking presence and calling it discharge is how v0.46.0 shipped four requests it did not contain."** A commit whose subject names the issue is not evidence the issue is discharged. ## ⭐ An agent outage can wear the costume of a scope decision #123 was WIP, dispatched in tranche 1, with **no branch under any local or remote ref** — because hertz had been unbound for hours with its lane stalled ([[ir18-lane-and-b-split-assignment]] family). Read as scope, it looks like a deferral; read at source, it is an outage. Absence of a lane is a fact about the FLEET before it is a fact about scope. It needs a line on the milestone either way. Relates to [[golden-ci-strategy]], [[golden-respin-test]], [[verify-the-subject-not-just-the-measurement]], [[report-measurement-never-issue-direction]] (I held the run — my assigned gate — and referred both rulings to doyle rather than deciding them).