Module projection
Expand description
The projection store rule family: its rules, its single enumeration, its three emitters and its entry macro.
§Why a projection rule may not live in suite.rs
cargo xtask spec-trace and no_orphan_rules both read suite.rs as the
event-store family’s rules module, matching pub async fn textually. A
projection rule written there is claimed by the event-store enumeration it is
absent from, and reported as an orphan of the wrong family. So this family
carries its own rules, its own enumeration and its own emitters, in one file
— which is the shape model.rs and concurrency.rs already have, and the
reason CF-22’s “exactly one place” is per rule family rather than per
crate.
§What the family costs, stated rather than discovered
Four exported macros:
for_each_projection_store_rule!,
three emitters, whose names CF-41 promises, and
projection_store_conformance! at the
crate root. happenstance-testkit carries its own version precisely because
a suite change can turn a passing adapter’s CI red, and a macro is the
hardest surface to walk back.
It reuses everything else unchanged: Capability,
RuleOutcome, block_on and
__rule_names. There is no projection-local skip type and no second
line shape, because an author reading one CI log must not have to learn two.
§What a green run here does and does not prove
Every rule below passes against MemoryProjectionStore, which is the oracle
and is supposed to pass. That the suite can fail a wrong store is a
different claim, and it is now discharged rather than carried: each rule
names the defect it rejects, and the store carrying that defect is registered
as data in tests/projection_mutation_coverage.rs, driven through this
enumeration, and asserted to fail exactly the rules it declares
(ADR-0010).
every_projection_rule_has_a_mutant is what makes that true of the next
rule as well: a rule added here without a store that fails it turns the
workspace red, and there is no exemption list.
What a green run still does not prove is that this family is complete.
Every rule §4.11 assigns to an adapter’s own suite is now written; what is
still owed is the six runner-dependent ones, which CF-36 moves to the
workspace e2e crate because they need a runner rather than a store
(spec/SPECIFICATION.md:6095-6104). A store that passes everything here has
not been observed under replay.
The seventeenth landed last and did not land quietly.
fresh_projection_has_no_checkpoint
was held out of this module for a day, because §4.11 assigned it to PS-19
and PS-19’s MUST is scoped after a successful reset — so writing the
rule would have widened a [FROZEN] clause by test, convicting adapters of an
obligation no sentence stated. The repair was a new clause rather than a line
edit: ADR-0030 minted PS-38, whose second sentence is the obligation, and the
rule cites that clause rather than the one it was filed under
(ADR-0030).
§Three rules here are answered by a skip against the reference fixture
failed_commit_leaves_both_unchanged
is PS-1’s second conjunct — the arm about a commit that reported failure —
and it is gated on
COMMIT_FAULT.
refused_reset_changes_nothing is
PS-18, and it is gated on
RESET_REFUSAL. Neither
capability is supported by any fixture outside the mutant harness, and
MemoryProjectionFixture declines both honestly and for reasons it states:
the reference store applies both halves of a commit under one write lock and
has no write that can be made to fail, and it holds no protection policy, so
there is no projection it could decline to reset. A run against the oracle
therefore prints two SKIP lines, and an adapter that wants either conjunct
checked has to supply what its own store can do.
The third is
committed_rows_and_checkpoint_survive_a_reopen,
PS-1’s durable read across a reopen, gated on
REOPEN. The reference store has no
medium outside the process, so MemoryProjectionFixture declines it, as
CF-17 requires, and the oracle run prints a third SKIP line.
InMemoryDatabaseStore and VolatileReadModelStore fail it by name in
tests/projection_mutation_coverage.rs, against a fixture whose reopen
keeps a modelled medium and discards the process state around it.
A third switch can produce a skip and is not on the fixture at all:
READS_THROUGH_BATCH
is a const on the store’s probe impl, because whether a batch can be
read through is a property of the batch type. An adapter declaring it false
— which PS-12 permits outright — gets
batch_reads_reflect_pending_writes
and rebuild_is_chunk_size_invariant
as skips naming that constant, with a reason the testkit writes rather than
the fixture, for NO_CEILING_REASON’s reason.
MemoryProjectionStore declares true, so a reference run does not print
those two.
Read that as the answer to “does a green run mean anything here”: for fifteen
of the seventeen landed rules it means the store was driven and asserted
about; for these two it means what the SKIP lines say. Neither rule is
decorative — PartialCommitStore, RefusalAsSuccessStore and
RefusalAfterTheFactStore fail them by name in
tests/projection_mutation_coverage.rs — but the demonstration lives against
a fixture that can arm a fault and protect a projection, and the reference
fixture is neither.
One stale sentence, named here because it cannot be repaired here. PS-1’s
Rejects: prose says “Today no rule can fail it”
(spec/SPECIFICATION.md:5122-5124), which stopped being true when
CheckpointOnlyStore landed and is further from true now. PS-1 is
[FROZEN], so the repair is a new decision atom under the repair-frozen-clause
discipline rather than a line edit, and it belongs to
unstable-projection-gate-and-clause-disposition along with the rest of the
PS-1 – PS-37 maturity sweep. Nothing in this crate may edit that sentence;
this paragraph exists so an adapter author reading the code is not the last
person to find out it is stale.
Modules§
- rules
- The projection conformance rules.