Intake gate follow-ups after PR #368 #369

The whole-branch review of PR #368 left five items that are real but do not belong in that pull request. They wait until the gate has run enforced for a while, because two of them change the arming set and the other three need live numbers to size.

• Split collection out of tools/pr_intake.py (about 1,900 lines) into its own module. That changes ARMED_MODULES, so it lands together with a re-arm.
• Attach the writes already made to WriteError, so a partial apply's writes reach the summary when the run goes red.
• Read the author's merged record lazily, or cache the closed-pull listing per run. Today every event run paginates every closed pull request before it decides anything, which is about twenty calls per event.
• Before arming, run the dry sweep twice a day apart and diff the would-queue lists. A pull request that flips with no human event between the two runs is a hash or seen defect the fixtures did not catch.
• Say in the design that the health renderer on main reads data lines the integration sweep writes, so an unknown key stays tolerated until promotion. The renderer already ignores unknown keys, and the sentence keeps it that way.

Part of #362.