Filter a module's own coproduct-variant names out of the emitter's import candidates

[gunbai-bot]

The defect

v1.05_emit_rust reference_derived_use_lines filters import candidates against the names the module declares. Its own note names the hazard exactly: the bare-name registry is last-write-wins across the closure, so a homonym elsewhere steals the row and a self-colliding pub use is synthesized (the E0255 class).

The filter was written over items — one authored name per TOP-LEVEL item. A coproduct's variants are CHILDREN of the item, so they never entered it. The guard was at the wrong grain, not missing.

A bare self-reference to a module's own variant therefore fell through to map_get(registry, name), resolved to whichever module last wrote that global name, passed export proof, and got a use-line for a symbol the module already binds itself (via use self::Enum::*;, or natively for a grounded alias).

The fix

local_coproduct_variant_names reads the variant set from each coproduct item's own children, and the result joins the already-filter.

Deliberately not via emit_info.variant_to_enum or type_summaries: both are keyed by bare name and carry the same closure-global last-write-wins, so answering a module-local question through them would reintroduce the defect one layer down.

Measurement — prediction stated first, then measured

Predicted delta: only REMOVALS of pub use lines whose imported name is a variant of a coproduct declared in the emitting module. Nothing else.

Two arms, each built from its own stage0 mirror (distinct sha256), over the emitted closure of src/v2/compiler/00_compile.dag — 175 files:

• Membership delta: exactly one file. v2_std_node.rs loses

  pub use crate::std_constructors::{Cardinality};
  use crate::std_constructors::Cardinality::*;
  

  Cardinality is a variant of v2.std.node's own Connective; the homonym that stole the registry row is std.constructors' type Cardinality = Required | Optional. The measured set is the predicted set.
• v1 seed: zero bytes. With the mirror installed, claim_executor --required-regen reports first_generation_equal=true planned=136 executed=136.

READ THIS BEFORE TRYING TO REPRODUCE ON A SMALL ENTRY — you will get zero, and the fix is not inert

The class is closure-size-dependent. Measured: --entry dag/std/algebra.dag (7 files) and --entry dag/std/node.dag (15 files) both show a zero delta on both arms.

That is the mechanism, not a weakness. The escape requires the homonym to be present in the same closure to win the last-write-wins registry row; an entry-scoped closure usually excludes it, and then the pre-existing info.module_name != this_module_name check already skips the name. So a per-entry emission can be clean while a wide one fabricates — and any population of this class measured at entry grain is complete over the wrong subject while looking complete.

Control failed, and that is a separate finding

Run-to-run over the same binary, BOTH the unmodified and the modified compiler emit v2_lens_enforcement_vocab.rs and v2_std_cross_tree_resolution.rs with import blocks in varying order (identical membership) — three runs of one binary differ. This instability is present on main at 38a127b, is not caused by this change, and is why the arm comparison above is stated at membership grain for those two files. It is an unordered-iteration defect (an unordered collection where an ordered one belongs); it is routed to its own owner and is not chased here. Note the scope: it makes empirical byte-comparison controls unreadable on those two files; it says nothing about structural byte-identity arguments.

Re-measured after merging main

The branch was merged onto 3bbd53c05a (which includes #9439's four-armed disposition refactor of this same function; the conflict in 05_emit_rust.dag was resolved by keeping main's function and re-applying this change on top, and the stage0 mirror was regenerated, never hand-resolved). The two-arm delta was re-run against that base:

• v2_std_node.rs loses the same two lines — unchanged result.
• v2_std_cross_tree_resolution.rs shows an order-only change with identical membership. That is one of the two files named in the control finding above, so it is instrument noise, not a delta.
• claim_executor --required-regen → first_generation_equal=true planned=138 executed=138, and --required-regen-fixed-point → fixed_point_equal=true referenced_at=3bbd53c05a.

First measurement was taken against 38a127bd60; the re-measurement above is against 3bbd53c05a.

[gunbai-bot]

CI red here is main's red, not this PR's. Same failure signature, both phases, all six errors identical:

main run 33095906314 (4e5486487) and this PR's run 33097895476 (913d073) both report:

required-ci: lane=witnesses phases_run=3 failed=2
required-ci: FAILED PHASE declarations (1 finding(s))
required-ci: FAILED PHASE floor refused: modules_resolved=4136 modules_excluded=4
  dag/gunbc/fleet_fan_wiring_witness.dag:754  duplicate declaration 'srv3_wiring_with_a_duplicated_header'
  dag/product/fabric/contention.dag:23        'grant_duration_seconds' not found in module 'product.fabric.supply'
  dag/product/fabric/contention.dag:280       non-exhaustive match: missing UnobservedGrantDuration
  dag/product/fabric/contention.dag:421       non-exhaustive match: missing QuoteNotPriceableWithoutDuration
  dag/product/fabric/contention.dag:463,469   function 'grant_duration_seconds' not found in scope

The witnesses lane has been failing on main for the last five commits (4e5486487, 353529a35, b37bb5c30, 26eee4b8e, f820985cc). None of the named files is in this PR's diff, which is src/v1/05_emit_rust.dag and its generated stage0 mirror only. Nothing here is fixable from this branch.

[gunbai-bot]

Two updates now that this is out of draft.

The reordering class is independently confirmed. The control finding in the body — that both the unmodified and the modified compiler emit v2_lens_enforcement_vocab.rs and v2_std_cross_tree_resolution.rs with varying import-block order — has been measured independently over six runs on one binary and one corpus: two distinct outputs per file, sorted lines identical, every changed line a pub use, and rustfmt normalizes the difference away. So the order-only change this PR's arm comparison shows on v2_std_cross_tree_resolution.rs is instrument noise, not a delta, and the membership-grain framing in the body stands as measured. The class is owned elsewhere and is not chased here.

The CI red is main's, and both halves are owned. Per the comparison above: the duplicate declaration in dag/gunbc/fleet_fan_wiring_witness.dag was fixed by #9497, and the dag/product/fabric/contention.dag break came in with #9397 and is being repaired by #9488. Note the floor phase is currently dark on main rather than merely failing — it refuses at strict preparation, so no claims execute and no ledger is written; a green on this branch is not currently producible from main's state. Nothing in this PR's diff touches either file.

[gunbai-bot]

Executed-specimen gap from node adhoc-0a3a31ae-4ee: I ran this PR head (96000c2) through the 175-file Rust emission of src/v2/compiler/00_compile.dag (BuildBuddy invocation). The compile completed with 0 blocking errors and 175 emitted files, but the node’s exact std.algebra / Optional / empty_map specimen survives in emitted src/std_algebra.rs:

pub use crate::v2_std_collection::{empty_map};
pub use crate::v2_std_optional::{Optional};

So the direct local-coproduct-child filter fixes the PR’s measured Cardinality case but does not reach this assigned specimen. I am not opening a competing implementation PR; this case should join this PR’s construction or receive an explicit disposition here. The same emitted tree had zero Option<()> / Optional<()> matches, consistent with the unit-collapse face having been repaired independently before this PR.