backlog(#340): the merge-queue cost, re-derived — and two false statements in the item itself

[wshallwshall]

BACKLOG #340 was filed with its cost evidence agreed but not written. The hand account it was to be built from was wrong in every load-bearing figure, so this re-derives all of it from the Actions API rather than relaying it.

The claim under test, inherited via a handoff: "#132 took four full CI cycles to land, none from a failure — green → main moved → rebase, ×3. ~80 min of runner time for a change correct on the first pass."

What #132 actually did, from actions/runs, actions/jobs and commits:

	claimed	measured
CI cycles	4	9 runs / 10 attempts, across 11 head moves
from a failure	none	2 attempts failed, one on the head it opened at
rebases	3	5, four onto a tip landed 1–14 min earlier
runner time	~80 min	212.7 min run wall-clock / 442.5 min job wall-clock
correct first pass	yes	opened at 1 commit, merged at 5

The two quantities differ by 2.08x, so the unit is load-bearing. No billable figure exists — /timing reports 0 billable ms on every run (self-hosted legs), so none is quoted.

Two statements already in the item were false and are corrected:

• "A hand-coordinated merge freeze … did hold main still for a full window." It did not — main advanced four times between ADR 0154 increment B: synchronous captured-downstream reply (reply_from) #119's first fully-green head and its merge, one of them 8m26s after the freeze was recorded in a work claim.
• "ADR 0154 increment B: synchronous captured-downstream reply (reply_from) #119 still failed" / "the bounds that actually killed ADR 0154 increment B: synchronous captured-downstream reply (reply_from) #119" reads as the PR dying. ADR 0154 increment B: synchronous captured-downstream reply (reply_from) #119 merged at 2026-08-02T01:45:00Z, 12h15m after arming.

The caveat is stated level with the cost, since filing an overclaim inside the ticket about overclaiming would be its own instance. Chiefly: the overnight serialization is not evidence that hand coordination worked — it is entailed by strict = true plus a 20–25 min suite either way — and strict does not always serialize (#110/#111 merged 24s apart; mechanism explicitly not established).

One new actionable finding: no workflow carries a merge_group: trigger, so zero of the 13 required contexts would report on a queue ref. That makes Proposed step 2 a precondition, not a follow-up — enabling the queue first would wedge every open PR.

Also records the measurement-cost linkage to #344: supersession-and-re-run produces multi-attempt runs, and the default actions/runs/{id}/jobs endpoint returns only the latest attempt, so failed earlier attempts vanish from any duration table built from it — at exactly the tight end where a margin is decided.

Scope: docs/BACKLOG.md item #340 only. #344's figures are untouched — another session holds that item. Banner invariant re-checked (264 items, one status each); ledger gate passes.

Every figure above was produced under adversarial verification (independent refutation passes per claim, on units/population/overclaim lenses) rather than by a single author, because the six retractions in this repo last night were each caught by a peer and never by the author.