Function reset_is_not_commit_at_first
pub async fn reset_is_not_commit_at_first<F: ProjectionFixture>(
open: impl AsyncFn() -> F,
) -> RuleOutcomeExpand description
reset is distinguishable from commit(empty_batch, id, FIRST, Live),
in the checkpoint and in the replay that follows it.
PS-19 and PS-20. The substitute is the one every deployment reaches for —
six scenarios out of six reached for it and all six got it wrong
(RUNBOOK.md:3904-3907) — and it is wrong in the way nobody ever finds:
the checkpoint reads back as a position, the runner resumes strictly
after it, and event 1 is skipped permanently and silently
(spec/E2E-CASES.md:437-456).
Rejects: CommitAtFirstResetStore — a reset implemented as
commit(batch, id, FIRST, Live). It compiles, it returns Ok, and the
checkpoint moves, so every rule asserting only that something changed
passes it. TwoStatementResetStore fails it too, at the same assertion:
a checkpoint that never moved is Live as well.
§Both halves, and neither is sufficient alone
The state half compares the two checkpoints by variant —
NeverRun against Live { .. } — never by position: comparing a
checkpoint against any SequencePosition is simultaneously a CF-6
violation and the exact value the defective store writes, so a rule
written that way could not tell the two apart. Checkpoint is a
three-variant enum precisely so this assertion can be made without naming
a position (spec/SPECIFICATION.md:5008-5024). The consequence half
derives a resume point from each under the port’s own rule — in one
private helper shared by both arms, so the derivation is stated once —
and asserts the event at the store’s first position is applied in the
reset case and not in the substitute’s. The variant
check alone would not see the consequence; the resume check alone would
not see the state.
No runner is built. PS-20’s property is a claim about how a resume point is derived from a checkpoint, and that derivation is four lines the rule owns — the runner itself belongs to a later phase and a different crate.
§Why each checkpoint is read immediately after its own operation
So that a store which shares one checkpoint row between projections fails
reset_is_scoped_to_one_projection and
distinct_projections_advance_independently — the rules that own that
defect — rather than this one, which is about a single projection’s
state being told apart from another single projection’s.