Verify the Actions spend actually dropped

[ExtraToast]

What to build

Sixteen pull requests were opened to cut the estate's private-repo Actions spend, with a predicted result: roughly 3,000 billable minutes a month down to roughly 800. Every one of those predictions is a model, built from job durations and a rounding rule. None has been observed.

That gap matters more than usual here. The changes were made while the org was out of Actions minutes, so not one of the private-repo pull requests was ever verified by CI — they were reviewed by reading and by local linting only. A collapse that silently dropped a check, or a shim removal that quietly stopped boarding issues, would look exactly like success.

Confirm what actually happened, once the changes have been live long enough for a full billing period to be meaningful.

Read the real consumption from the billing API and compare it against the predictions. Where a figure disagrees materially, find out why rather than adjusting the estimate to match — a workflow that did not get cheaper usually did not get changed the way the pull request claimed.

Then confirm the changes did not cost coverage, which is the failure mode the missing CI would have caught:

• every check that ran before a pipeline was collapsed still runs, and can still fail its pipeline
• issues and pull requests opened in the private repos still reach the project board, now via the sweep rather than a per-repo shim
• hygiene findings are still surfaced somewhere a human sees

Note that the earlier per-workflow estimates were derived from a single busy week and projected forward; the month's actual rate was around half that. Expect the shape of the saving to hold and the absolute figures to be optimistic, and correct the record rather than quietly restating it.

Acceptance criteria

• Actual private-repo consumption for a full month after the changes is recorded, from the billing API
• It is compared against the prediction, and any material gap is explained rather than absorbed
• Every collapsed pipeline is confirmed to still run all its original checks, by making one fail
• Board coverage is confirmed by opening an issue in a private repo and watching it arrive
• The corrected numbers are written where the original estimates were recorded, so the record does not stay wrong

Blocked by

• Restore Actions capacity so the pending fixes can merge #38 — nothing can merge, and therefore nothing can be measured, until Actions capacity is restored.