[mforce]

Why

Issue #863 asks for evidence before splitting the serialized integration collection. The measurement found that DemoSeedTests.DemoSeed_PopulatesEveryScreen_AndIsIdempotent passed without performing its first seed: the boot fact had already populated the shared default account.

Change

• Require SeedStatus.Seeded on the first call and SeedStatus.AlreadySeeded on the second.
• Give the population and boot facts distinct Postgres fixtures, so each establishes its own default-account state regardless of fact order.
• Give AccountProvisioningTests a distinct fixture because its failed-provisioning checks compare database-wide counts.
• Record the per-class collision map and width-two through width-four timing model in docs/plans/863-integration-collection-split/01-measurement.md.

Evidence and decision

The first-call assertion went red with the old fixture: expected Seeded, actual AlreadySeeded. The focused green run passed 21 tests. The full local integration run and four PR CI runs each passed all 1,873 tests. Local process wall time changed from 416.7 to 395.3 seconds; the shared span changed from 403.5 to 383.0 seconds. PR CI test steps ranged from 390.4 to 605.9 seconds. Across the four runs, the shared-only tail was absent or at most 23.5 seconds. A width-two split has a conditional ceiling of about 4% before extra fixture cost; the document recommends do not split.

The staged documentation passed SchemaDocsTests.PostgresImagePin_IsOneIdenticalStringAcrossEveryTrackedFile.

Refs #863. The issue remains open for a later split decision if the CI schedule changes.