[GEN-E3] ref chains, pure policies and schema-name resolvers get one owner each; an error-union refusal no longer cascades (PR 3/3, #436)

[Blind-Striker]

Binding's ref-chain walks, pure policies and schema-name resolvers each get one owner. A refused error union no longer cascades into misleading errors (#436). Generated output is byte-identical.

Closes: P44, P38, U7-30, U10-09, P47, P48, P50, #436. This is GEN-E3 PR 3 of 3, following the approved plan (faz4/GEN-E3/plan.md; amendments A.1–A.4, 14 and 15 bind this PR). PR 1 is #425 and PR 2 is #437. It was developed in parallel with GEN-E6 PR 2b (#455), with disjoint write sets. It was rebased onto 7bb2faef, past RT-E2 C, #451 and the preview.12 cut #454. This PR touches no CHANGELOG, and the full local check was rerun on that base.

What changes

• RefChain (P44) replaces the four ref-chain walks with one walk. Its result names the last successful key, including what QueryEnumShapePolicy needs. The document-based OperationFacetContext.Resolve is gone, and its two UpstreamClientQueryTypeTests callers use RefChain (amendment 14).
• The pure policies become static (P38, U7-30, U10-09):
  • comparer-only Binding types go from 10 to 0, and Binding _comparer fields from 11 to 0 (S4);
  • SchemaPlanBinder keeps its instance shape;
  • SchemaAliasApplier and PathTemplateValidator become static, along with the four Ingestion/Projection pure types (GraphKeyBuilder, ErrorStyleClassifier, EnumSchemaProjector, LiteralClassifier);
  • their constructor threading, DI registrations (ToolApp) and BindingTestHost's hand-wiring go too;
  • SpecIngestion.cs changes only in its GraphKeyBuilder construction and call lines (S1).
• SchemaNameResolver's four facet-root resolvers become their own static collaborators: RequestBodyRootNameResolver, QueryEnumRootNameResolver, EnvelopePayloadRootNameResolver and EffectStreamRootNameResolver (decision 9, S2). SchemaNameResolver goes from 410 to 205 lines.
• A duplicate-tag error union refusal cascades into misleading 'absent union' membership errors #436: a refused error union drops its membership claims before the graph is published, the same mechanism [GEN-E3] union binding returns its decisions; nested-union and object-or-array defects refuse or bind correctly (PR 1/3) #425 used for P36. A duplicate-tag refusal now reports one error instead of three. Amendment 15 makes Bind_Should_Refuse_A_Tag_Owned_By_Two_Closure_Schemas a deliberate, red-first change: three errors → the duplicate-tag error alone, as the full ordered list. Its response-collision row keeps its Naming error after it.
• Hygiene:
  • P47: the "in M1" milestone text is gone from the WebSocket and PUT refusals.
  • P48: the session-event wrapper is removed (plan §8.1). Removing it pushed SpecBinder.Bind past MA0051's 80 lines, so its orchestration got one private-method extraction inside SpecBinder.cs, keeping the pass order, arguments and results. That is mechanical fallout of a planned move, approved by the lead; there is no new type and no suppression.
  • P50: comment fixes.
• .editorconfig: the four CA1822/S2325 exemption sections go (4 → 0). The SuppressionLedger snapshot loses 8 deletions plus their aggregate count (editorconfig: 158 → 150), edited by hand with no auto-accept (S3, and the lead's OK on the count).

Proof (spec)

Metric	Result
Ref-chain walks	4 → 1
Comparer-only Binding types / _comparer fields	10 → 0 / 11 → 0
CA1822/S2325 exemptions	4 → 0
SchemaPlanBinder	154 lines (≤ 250)
OperationPlanBinder	252 (≤ 400)
SchemaNameResolver	205 (≤ 400)
New Binding files over 400	none
Methods with ≥ 6 parameters (decision 8, methods only)	0
generate --verify	byte-identical

Tests

• Red first:
  • RefChain's six contract rows: terminal, nullable, missing target, self-cycle, two-node cycle, last key;
  • A duplicate-tag error union refusal cascades into misleading 'absent union' membership errors #436's two rows;
  • a nullable-enum characterization;
  • two milestone-message rows.
• Mutation: three compiling mutants, all killed, each file restored with cp/cmp:
  • RefChain stops one hop early;
  • RefChain ignores the visited set;
  • a refused error union keeps its claims.
• Existing-test edits, each under the plan or an amendment:
  • CSharpNamePolicyTests' summary (comment only);
  • the two ToolAppTests marker strings and two PendingOperationBindabilityProbeTests messages, which lose only " in M1";
  • A duplicate-tag error union refusal cascades into misleading 'absent union' membership errors #436's two rows (amendment 15);
  • the S3 snapshot lines;
  • QueryRequestFacetBinderTests' two enum-refusal tests, folded into one data-driven test with every assertion unchanged;
  • otherwise only compile-forced arrange lines (new X().Bind → X.Bind, and constructor arguments removed), in BindingTestHost, SpecBinder/OperationPlanBinder/StabilizeDuplicatePolicy/ModelReachCollector/UnionMembershipValidator/UnionMemberHoistBinder tests, UnionHoistPlanData, EmitterPlanFixture, GraphKeyBuilder/LiteralClassifier/SchemaProjector tests and the two projection test hosts.
• E0 guards: full ordered error lists, and both TST-E1 site lists stay empty.

Clone baseline, every entry accounted for

Tools (1.98%): −5 stale, +2.

• The 5 removals: 3 are pairs freed by this PR's deduplications, and 2 are pairs earlier PR 3 work removed.
• Duplicates removed: the static conversions made five latent duplicates textually identical (they had differed only by instance access, e.g. _keys.Append against GraphKeyBuilder.Append). Each now has one owner, a same-file static helper:
  • LiteralClassifier's two marker prologues;
  • the media projection shared by RequestBody/ResponseProjector;
  • the array-node projection;
  • and two more.
• The 2 additions: ObjectSchemaProjector.cs:10-23 ↔ PrefixItemsAdapter.cs:12-25 and ObjectSchemaProjector.cs:12-23 ↔ UnionNormalizer.cs:23-34.
  • Why they stay: each is a one-field constructor plus an argument-guard prelude, with no shared logic. A helper would only wrap the guards, which is slop.
  • The change that removes them: jscpd sees the constructor + guard-prelude idiom as clones once statics strip the varying lines #452, which decides structure or detector there (quality-gates, Clone gate).

Tests (5.38%): −2, +1.

• b4bc3f07 is freed by the QueryRequestFacetBinderTests fold.
• The fixture pair EmitterPlanFixture.cs:1041-1067 ↔ UnionHoistPlanData.cs:149-175 is a fingerprint swap, 94b8d278 → e59cfed5. PR 3's compile-forced new X().Bind → X.Bind edits shifted it. With master's two fixture files restored at this head, the gate shows 0 new clones.

Budget (against the merge-base)

• Hand-written tools: +605 / −711, −106 net, against a target of ≤ −40. The run reached −123, and the review fix-up spent +17, mostly restored rationale comments.
• Tests: 169 added (−94), against a cap of 200.

Local check and review

• Local check at this head, on dev-01, after a root Release build: slopwatch 0 issues; the Release build with 0 warnings; both format passes; verify-clones passing; --help; generate --verify current; the whole Tools.Tests project (guards included) at 2070/0. This is the rerun after the review fix-up and the rebase, on a quiet host. The first check, on the pre-review head, ran beside a dispatch. CI is the full gate.
  • Bun 1.4.3, off the 1.4.2 pin.
• verify.log: each run's harness verify covered the guards, 28 touched classes, the whole project, verify-clones and generate --verify (manual §5.3).
• Process: the Sol guide pilot template (task 2).
  • The stops: run 1 stopped on MA0051, the in-file extraction above (over-triggered). Run 2 stopped on the S3 aggregate count, the lead's wording (justified). The clone fix-up's run 1 stopped on a restore-authority ambiguity in its prompt (over-triggered).
  • Review: the coordinator's Claude review and a fresh cross-family review, comparing git show 695c5557:<file> with the new code behaviour by behaviour. Neither found a behaviour change for any document they constructed: RefChain's four walks, the static conversions, resolver precedence and order, the SpecBinder pass order, and A duplicate-tag error union refusal cascades into misleading 'absent union' membership errors #436's scope. Fixed here:
    • the nullable characterization, which couldn't fail for the regression it names. It gains a direct ResolveModelKey(Ref(State), graph) row and a full-pipeline refusal, both red against an "unwrap Nullable inside the walk" mutant;
    • two helpers moved to their owners: the media-entry projection to MediaTypeProjector, and ProjectArrayItems to an instance method taking items explicitly, with no Debug.Assert;
    • the rationale comments the moves had dropped, restored;
    • one SchemaBindingContext factory, replacing two tests' hand copies.
  • A review ask that the analyzer overruled: reverting ReferencesMarkedUnion to IsMarkedUnionReference triggers S3218, because the static method shadows ProjectedBranch.IsMarkedUnionReference. So that rename stays: it was forced, not unrequested.

Provenance: Generator lane / opencode dispatch harness (openai/gpt-6.1-sol-1m), coordinated, reviewed and finished in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}