R3 P0 — FieldProject dual-authority dissolution (Class B; illegal-state-representable)

[briansrls]

Auto-opened by session-dashboard for session bright-carp-806.
Pushing to session/bright-carp-806 advances this PR.

Closes #2462

Worker attestation

Before flipping this PR to ready for review, confirm each item:

• Title describes the change (not the session id or branch).
• PR body summarises what and why (replace the TODO below).
• Tests run: name the command (e.g. npm test, cargo test) and the result.
• If this closes a work item, the body contains a Closes #N directive.
• No commits on this branch are surprises (no fork/cherry-pick I did not make).
• No secrets / credentials / large binaries staged.

Summary

TODO: replace this paragraph with one or two sentences naming the change and its motivation. Reviewers read this first.

Test plan

• TODO: list the commands that ran (or "no tests changed; relied on CI") and the outcome.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 311f7726 · Trigger: schedule
• Thinking: 214s wall

---

BLOCKING (3)

Root Cause

• src/v3/compiler/src/dag.rs 3544|ResolvedFieldProject is modeled as { field_ref: DeclarationId } only, so downstream re-derivation uses type equality instead of a stored field identity, which is not injective for duplicate-typed fields.
• src/v3/compiler/src/dag.rs 3544|The dual-authority split drops field-label identity in the resolved state, so equivalence cannot distinguish provenance of a field_ref when multiple fields share the same type slot in a Conj.
• src/v3/lenses/cost.dag 102|cost.dag defines entry_for with b.result_port and l.result_port and loop_cost(..., l.body); the generated file appears to have drifted from that authoritative source contract.

⚠️ The PR has blocking correctness issues in resolved field projection identity and symbolic-cost lens entry generation and should be fixed before landing.

[briansrls]

Violations (could not place on specific lines):

• src/v3/compiler/src/emit.rs:3689 This breaks ILLEGAL STATES UNREPRESENTABLE because resolved_field_project_label resolves label by children.iter().find(|field| field.ty == field_ref), so two distinct Conj fields with the same declaration type can map to the same projected label and emit the wrong field accessor.
• src/v3/compiler/src/infer.rs:4024 This violates FACTS FLOW FORWARD and the “distinct facts” invariant because transform_targets_equal_under_subst treats resolved projections equal when their field_ref declarations are equivalent, ignoring source-field identity and allowing x.a and x.b (same type) to be conflated.
• src/v3/compiler/src/lens_cost_symbolic_generated.rs:4400 This violates FACTS FLOW FORWARD in DB-3: entry_for now keys Branch/Loop entries by input/source, and loop_cost calls body_cost(..., &l.id), so lookups are generated for the wrong ports/nodes and can miss costs that dimension.rs expects at behavior result ports.