[Blind-Striker]

Operation binding becomes named collaborators that return what they decide. One generated-type-name inventory refuses every collision. Generated output is byte-identical.

Closes: P42, P46, P31, P43, #414. This is GEN-E3 PR 2 of 3, following the approved plan (faz4/GEN-E3/plan.md, amendments included). PR 1 is #425. PR 3 (graph and statics) waits for TST-E3's first half.

What changes

• The operation binders move out. SingleOperationBinder and ErrorMapBinder leave OperationPlanBinder, which goes from 593 to 254 lines.
• P42 (a characterization): each facet returns FacetResult<T>. A facet is bound, or refused with its own errors:
  • What it replaces: the shared error-count window and the partial-plan convention.
  • Independent facets stay independent. A refused query still feeds pagination and request-merge diagnostics, through its classified properties. A facet refusal still rejects the operation, MissingRow included.
• P46: ErrorMapBinder reuses the error union's verdicts. It does so only for closure members. ErrorStyle.None objects and stream causes still resolve directly. An error schema with no unique C# name reports once, from the error union.
• P31: a cursor list keeps its method and loses only the pagination companion when it declares a header, has a required member, or has an incompatible signature. Before, the whole operation refused.
• P43: one GeneratedTypeNameInventory, built once after the last declaring pass, feeds one collision wall. This is the lead's "deferred single wall" (plan amendment A.5).
  • Coverage: every generated name.
    • schema names, seeded from the same type-name map as before;
    • models;
    • unions, with their Kind and Unknown names;
    • hoisted interfaces;
    • clients;
    • envelopes;
    • streams.
  • The hoist composer records claims. It keeps only a structural first-claim guard on its own interfaces.
  • A hoisted name refused against another declaration never reaches placement. Before the fix-up review, it crashed in ModelOwnershipPolicy.
  • First claim wins: of two hoisted claimants, the first keeps its name and only the second reports, as on master.
  • Two derived names are now refused (for example a union's Unknown name equal to another union's Kind name). The old wall missed that.
• Generator: an invalid curated method name also reports a misleading 'does not pluralize naively' error #414: an invalid curated list method name reports once, by its own rule. The misleading "the group does not pluralize naively" error is gone.

Tests

• Red then green: every behaviour change.

  • P31: two expectation flips and three signature rows.
  • Generator: an invalid curated method name also reports a misleading 'does not pluralize naively' error #414: its two-row whole-list test (amendment F6).
  • P46: the unnamed-error regression.
  • P43's missed-collision rows:
    • hoisted against a client and against a Kind name, each exactly one refusal with the hoist's literal (the lead's A.5 condition);
    • hoisted against a placed model and against a placed union, through the full pipeline, red with ArgumentException;
    • hoist against hoist;
    • derived against derived;
    • a curated spine name on a non-emitted alias.
  • The generic model-collision message: four rows, proven by a mutation that deletes the branch (185/4 → restored with cmp → 189/0).

• Restored-behaviour rows: the fix-up derived each oracle from git show 04f18994:<file> and pins it as the full ordered Errors list:

  • the enumeration-name error behind a refused query;
  • the mixed body/query diagnostic;
  • a closure member's marker error;
  • a duplicate-tag error union that no longer drops its operations.

• Existing-test edits:

  • the plan's §9 PR 2 rows: P31's two expectations, Generator: an invalid curated method name also reports a misleading 'does not pluralize naively' error #414's two rows, and inventory argument wiring;
  • assertions strengthened from Contains or a category filter to the full ordered list: Bind_Should_Refuse_A_Tag_Owned_By_Two_Closure_Schemas and the mixed body/query row;
  • compile-forced arrange-only edits (lead's blanket rule): InternalNameCollisionCheckTests.Check, OperationPlanBinderTests.Bind_Should_Preserve_Type_Name_Claim_Order, UnionHoistPlanData.CheckNames, and P31's arrange moved into one CursorQuery helper.

  No existing expected error list was reordered.

• E0 guards:

  • full ordered lists with CollectionOrdering.Matching;
  • no Contains, and no IsTrue over Any/All;
  • both TST-E1 site lists stay empty.

• Found and kept as on master: A duplicate-tag error union refusal cascades into misleading 'absent union' membership errors #436. A duplicate-tag error union refusal still cascades into two "absent union 'IOpenCodeError'" errors. That is base behaviour, which this PR restored exactly and now pins. The fix is the same claim-drop PR 1 used for P36, and it lands in GEN-E3 PR 3, with a deliberate red-first change to this row.

Clone baseline (proven)

tests.baseline.json: −3 / +1, two fewer clones.

• The swap: 78811ccc → 94b8d278, for EmitterPlanFixture.cs:1041-1067 ↔ UnionHoistPlanData.cs:149-175. With master's UnionHoistPlanData.cs restored at this head, the gate shows 0 new clones.
• The removals: 04551f23 went with the OperationPlanBinderTests arrange, now the CursorQuery helper. 0555dcc5 is [GEN-E3] union binding returns its decisions; nested-union and object-or-array defects refuse or bind correctly (PR 1/3) #425's SpecBinderTests swap entry, gone with the Generator: an invalid curated method name also reports a misleading 'does not pluralize naively' error #414 rows that moved.
• Tools: 0 new. UnionHoistPlan.Claims became required in place, without being moved. The EmitPlan ↔ UnionHoistPlan property pair that the first draft showed is gone. Its owner, had it stayed, was GEN-E4 U7-36 (recorded in GEN-E4's plan).
• Totals: tests duplication 5.56%, under the 6% ceiling; tools 2.09%.

Budget (against the merge-base)

• Hand-written tools: +647 / −535, +112 net, against a cap of +120.
• Tests: 537 added (−50). PR 2's own work was 348, against a cap of 350. The rest is the review fix-up's rows, and the spec says caps bound scope, not review fixes.

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 2033/0. Bun 1.4.3, off the 1.4.2 pin. CI is the full gate.

• verify.log: for both runs, the harness ran the guards, 13 touched classes, the whole Tools.Tests project (failing only on the 14 packaging rows that need a root build; manual §5.3) and generate --verify.

• Review: the coordinator's Claude review and a fresh cross-family review, each comparing old against new code behaviour by behaviour. They found six behaviour changes that generate --verify could not see at the pin, all fixed here:

  • the placement crash;
  • the double hoist report;
  • missed derived collisions;
  • a narrowed claim set;
  • hidden independent diagnostics;
  • dropped operations.

  The fix-up rule was "restore 04f1899 exactly unless the plan lists the change; finish what is deliberate".

Provenance: Generator lane / opencode dispatch harness (openai/gpt-6.1-sol-1m) for 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.}