Retire @classified, @classifiedType and @commits: the corpus asserts relations, not leaves

@classified. The tuple axes (source wrapper, target wrapper, source shape, target shape) and the operation-member multiset each need a relation at the right grain. Some facts are assertable today (the corpus already asserts intent_resolved_field_claim, intent_bound_table, intent_authored_field_claim); the walk-side folds that own the rest (DeliveryFactRelation, OperationMemberRelation) are exactly the folds R682 re-sources into the store. As each relation lands, the corresponding axis becomes ordinary blocks and the axis argument leaves the directive. One grain rule, stated as the rule rather than a workaround: a block that asserts a relation at the relation’s own key cannot produce duplicate rows, so if a member block needs extra columns to keep duplicate arm tokens distinct (the operations: list is a multiset, and CorpusExpectations compares sets over the projected columns), that need is the signal the block is projecting below the relation’s grain. Each migrating commit derives the block’s columns from the relation’s key instead of remembering a caveat.

@classifiedType. The verdict it declares is a sealed leaf’s simple name, and the leaf zoo is what R682 deletes, so the replacement is deliberately not a store enum restating the leaf taxonomy: a type_verdict relation would be named for the walk’s question, at the walk’s grain, with the code being replaced as its only oracle. The default replacement is the relation a consumer actually forks on: where R682 lands a closed detection-verdict relation because a planner reads one, the document asserts that relation’s rows. Asserting the deciding facts individually (type backing, nodehood, synthesis, interface and union membership) is the fallback for types where no consumer composes them into a verdict; if the corpus reader would have to compose five facts into a verdict in their head, and a planner will evaluate the same predicate, the composite belongs in the model and the document should be asserting it there.

@commits. This is the one axis whose retirement is a design decision rather than a scheduling one, and the item says so rather than hiding it in a symmetric fork. The launcher command relation is derived by a planner and R682 keeps commands as in-memory plan-tier records ("each command relation’s declared arm set" is the surviving vocabulary its completeness gates re-key onto, not a store relation), so no relation for @expectEquals to name is ever scheduled to arrive. Two shapes work: (a) the corpus run lands each document’s produced command rows in the store before the expectation pass; (b) the document asserts the store facts the launcher verdict derives from, and the command relation’s own arm-census pins carry the derivation. Lean: (a), because it keeps the assertion at declared-equals-produced strength on the command row itself, per document, and keeps the launcher-production failure roster (launcherProductionFailureRosterIsExact) expressible; (b) weakens the corpus’s claim from "this command row exists" to "the inputs to the verdict exist", which output identity does not backstop per-document. Choosing (a) carries two obligations the fork would otherwise smuggle past the store’s gates. First, the rows are a function of the schema, not of captured facts, so they are scaffolding in the fact model’s sense, the shape rejection_ has: if they live as a declared relation they owe the full scaffolding charter (a meta_family roster row whose header states the writer, the cadence, the single reader and the clock it drains on, which here is this item’s own trio deletion plus R682’s command-vocabulary settlement); if that charter is not worth paying, the sanctioned per-reader shape is a LOCAL TEMPORARY table and the rows are not a declared relation at all. Second, @expectEquals’s worth rests on rows reaching the store only through capture (a document cannot state a shape capture never produces), and `expectationsRangeOverTheAssertablePopulationOnly enforces exactly that. Planner-produced rows must stay a distinguishable population, an explicitly named apparatus bucket that floor admits as its own third case, never merged into the captured-fact population; a reader of any block must be able to tell "this asserts a captured fact" from "this asserts a planner’s output".

In the prelude: the three directive declarations and the enums that exist only to type them (SourceWrapper, Member, TargetWrapper, SourceShape, TargetShape, TypeVerdict, LauncherSource, LauncherResult). Mint, SynthesisedType and @synthesises stay (see scope).

In the harness: ClassifiedHarness’s resolution and comparison halves for the trio (FieldCase, `TypeCase, CommitDeclaration, CommitCase, the enum-argument decoders and their unknown-value throws), ClassifiedDslTest’s per-document tuple assertion and the mirror floors for the retired enums. The mirror obligation does not vanish, and its successor is named per vocabulary rather than waved at: for a vocabulary the model owns, a closed CHECK plus a membership-binding test on the `rejection_validation_error.lsp_code pattern; for a vocabulary a surviving Java seal owns (the command arms), the mirror stays a Java meta-test against the seal and simply outlives the SDL enum it used to type, because the store’s own convention rejects a CHECK that hand-copies a compiler-enforced taxonomy. CorpusExpectations.membershipViolations validates blocks against CHECK vocabularies, which is a different claim and is not the successor.

The axis-census tests in ClassifiedDslTest (everyDimensionValueIsExercised, axisPairCensusIsDerivable, memberGrainCensusIsDerivable, launcherAxisCensusIsDerivable) recast as queries over the store relations that replace the directives, keeping their producer’s-arms grain, or retire where the replacing relation’s own census pins already state the same floor. Which of the two holds is a per-test decision made in the migrating commit, not a blanket deletion.

Depends on R682’s relations, axis by axis. This item does not build fact relations for the walk’s verdicts on its own schedule; that is R682’s deliverable and its law. The work here is the corpus-side conversion as each relation lands, plus the @commits materialization decision, plus the deletions. Axes whose relations exist already can convert immediately; the item ends when the last directive is deleted, which cannot precede the relations it waits on.

Coverage re-keying splits by who owns the vocabulary. For gates keyed on vocabularies R682 deletes, re-keying is R682’s: VariantCoverageTest’s output-and-type obligation reads `CorpusDocuments.coveredLeaves(), which resolves through the trio to sealed leaves, and the member-arms obligation is keyed on the walk-model OperationMember seal. This item coordinates with that work (the trio cannot be deleted while coveredLeaves() still reads it) and must not run ahead of it. The exception is this item’s own deliverable: the LAUNCHER_COMMITMENT obligation’s domain is LaunchSource plus ResultShape, command vocabulary that survives R682, and its covered set is today produced by corpusCommittedLauncherArms() reading @commits. Deleting @commits deletes that gate’s only witness channel, and nothing in R682 owes a successor to a gate over a vocabulary it keeps. The successor lands here: the covered-arm set derives from the replacing @expectEquals blocks, whose two-sided set equality per document is what keeps a declared row trustworthy as a witness.

@synthesises is out of scope. It has 2 applications, asserts minted type names for connection synthesis, and plausibly follows the same path (the store already declares graphitron_minted_type and graphitron_minted_field), but the user-visible ask is the trio; fold @synthesises in only if a migrating commit makes it free, otherwise file it separately when the trio is done.

Documents convert wholesale per axis, never per document: a corpus where the same fact is asserted two ways in two documents is harder to read than either endpoint.