[Blind-Striker]

Union binding becomes named collaborators that return their decisions, and two latent union defects refuse by name instead of producing wrong code. Generated output is byte-identical.

Closes: P39, P36, P37, P45. This is GEN-E3 PR 1 of 3, following the approved plan (faz4/GEN-E3/plan.md, amendments included). PR 2 (operations, with #414) and PR 3 (graph and statics) follow serially.

What changes

• One binder per union kind. SchemaPlanBinder goes from 810 lines to 166. Each new binder returns its plan and its membership edges:

  • MarkedUnionBinder (321 lines);
  • ErrorUnionBinder (94);
  • ObjectModelBinder (129).

• The graph replaces the accumulators. UnionInheritanceGraph (70 lines) assembles base types and fixed markers from the returned edges. It replaces the shared inheritance and fixedMarkers accumulators and their post-pass, which are gone (git grep).

• Errors stay on the existing collector. Per amendment F3, no result-local error batches yet; those are PR 2's FacetResult. The new records are UnionBindingResult, ErrorUnionBindingResult, SchemaBindingContext and one UnionChildEdge, which carries a nullable fixed marker.

• The error union is not a schema node. The graph never resolves parents, stream-cause membership or a fixed marker for the synthetic error union. A real component keyed OpenCodeError therefore cannot wire itself onto IOpenCodeError. Without this, that component would make the error union inherit from itself, and the base-chain walk loops. A regression row pins it. Master never wired the error union, so this closes a gap the new graph would otherwise open.

• P36: two shapes refuse with one named error ("nested union has multiple parents 'A' and 'B'") instead of silently keeping the last parent:

  • a nested union under two parents;
  • a stream-cause union with a union parent.

  The refused union's membership claims are dropped before the graph is published, so no "absent union" errors follow from its leaves.

• P37: an unmarked union of a free-form object and an array of unrestricted values binds to JsonElement. Before, it bound to an unspecified object, which cannot hold the array. A typed object-or-array pair, such as Config.ModelEncoded.cost, stays a structural union. Nothing at the pin has the free-form pair. The maintainer was told about this direction and may reverse it (spec decision 1).

• P45: each error-spine spelling has one owner.

  • ErrorMarkerPolicy owns the C# spellings: Tag, IOpenCodeError and the unknown names.
  • The _tag wire literal is owned on the Ingestion side, by ErrorStyleClassifier.WireName, and ErrorMarkerPolicy.WireName refers to it. Nothing in Ingestion references Binding, so the Binding → Ingestion direction holds.
  • The two Emission "IOpenCodeError" literals now read the owner.

• Canon: protocol-and-generation.md gains the P37 sentence, as its own paragraph after the standalone-not rules.

Orders (byte-identical output depends on them)

Each order the emitted code depends on is kept, and a test pins it:

Order	Test	Mutation proof
Record base lists: graph-key order, then branch order, error union last	ModelEmitterTests: emitted base list	reversed order → red; restored (cmp) → 9/9
Error variants: by curated type name, not tag or key (three schemas where key, type-name and tag orders all differ)	SpecBinderTests	sort by Tag → red; restored → green
First-claim hoisted carrier: by depth, then name (the depth-zero union sorts last by name)	UnionMemberHoistBinderTests	name-only order → red; restored → green

The union variant order, registry order and diagnostic text and order are unchanged; both reviews checked them side by side against master.

Tests

• Red then green:

  • P37 (1 → 1).

  • Both P36 rows went red twice:

    • first because the old binder did not refuse;
    • then against the first draft, which produced 3 errors instead of 1.

    They now assert the full ordered Errors list, exactly one error.

  • The OpenCodeError-key regression: it was red with a false multiple-parent refusal plus two absent-union errors, and is green as a clean bind.

• Per-schema error decisions: one row asserts the exact Variants and Refusals for an admitted schema, a schema with no name and a schema with a bad marker. Deleting either recording call turns it red. PR 2's ErrorMapBinder consumes them.
• Existing-test edits: only the plan's §9 P37 row: unspecified object → JsonElement, and the test name …As_An_Unspecified_Object → …As_JsonElement. No other existing assertion or snapshot changed.
  • The order fixtures, the P36 rows and one compile-forced arrange line (a constructor that gained two set arguments) changed only inside this PR's own new tests, between its two runs.
• E0 guards:
  • every new collection assertion uses CollectionOrdering.Matching;
  • no Contains, and no IsTrue over Any/All;
  • both TST-E1 site lists stay empty.

Clone baseline (one fingerprint swap, proven)

• The swap: tests.baseline.json exchanges ccafdf2207f06e4f for 0555dcc5c34a8475. The pair is SpecBinderTests.cs:1673-1678 ↔ :1779-1784, two pre-existing arrange prefixes. P37's allowed test rename shifted the token window.
• The proof: with master's SpecBinderTests.cs restored at this head, verify-clones passes against master's baseline with 0 new clones. The clone count is the same, 664.
• Tests duplication is 5.66%, under the 6% ceiling. Tools has 0 new clones.

Budget (against the merge-base)

• Hand-written tools: +702 / −696, +6 net, against a cap of +120.
• Tests: 240 added, against a cap of 350.

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; whitespace format passing; verify-clones passing (with the swap above); --help; generate --verify current; the whole Tools.Tests project (guards included) at 2004/0. The style pass caught two using-order slips (ErrorUnionBinder, ModelEmitterTests). After the fix, both format passes, the all-framework build, ModelEmitterTests (9/0) and SpecBinderTests (157/0) were rerun green. Bun 1.4.3, off the 1.4.2 pin. CI is the full gate.
• verify.log: the harness ran the guards, 14 touched classes and generate --verify on the committed tree for both runs (manual §5.3). The only red line was verify-clones on the proven swap above.
  • The fix-up's first generate --verify hit a shell timeout mid-format. The rerun recovered to a byte-identical src/ through GEN-E8's re-runnable write, with no hand edits.
• Review: the coordinator's Claude review and a fresh cross-family review. Both found the extraction faithful. Fixed here:
  • the error-union key (rated High by the cross-family review);
  • the P36 cascade;
  • the two order fixtures that could not tell the wrong sort;
  • the untested per-schema decisions;
  • set lookups;
  • two comments that had lost their reason in the move;
  • member order;
  • doc comments;
  • a constants-restating test, deleted (SLOP-T03).

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

---

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