v2 artifact.dag — dissolve single-variant coproduct husk (calibration probe)

[briansrls]

v2 stage rework — artifact.dag (calibration probe)

Calibration investigation: applies the v4-discovered modeling discipline back onto a working v2 stage. Draft for operator review — not a binding direction. One of a parallel set of v2 stage probes.

Before / after line counts

File	Before	After
src/v2/artifact.dag	113	84

What dissolved

• PartitionRule = Explicit { artifacts } (was lines 75–85) — a single-variant coproduct. One variant is not a sum; fn plan_artifacts was a degenerate 1-arm match (the limiting case of a template-hole). The header even noted "Placeholder partition strategies were removed" — this is the husk left behind. Dissolved: type PartitionRule + fn plan_artifacts deleted; default_artifact_plan now constructs the ArtifactPlan directly.
• type ArtifactOutput (was lines 94–97) — zero consumers anywhere in src/v2 (grep-verified). Speculative dead code held alive only by an explanatory comment. Deleted.
• import v2.std.core { TextFile } — used only by ArtifactOutput; dropped.
• Net: −45 / +15 lines, two type declarations and one function removed, no behavioral change.

Behavioral contract

default_artifact_plan returns the byte-identical ArtifactPlan { artifacts: [Artifact{…}], boundaries: [] } it did before — previously routed through plan_artifacts(Explicit{…}), now built inline. No external file references PartitionRule, plan_artifacts, ArtifactOutput, or TextFile-via-this-module (grep-verified across src/v2). The file's three external consumers (compile.dag, coercion.dag, 05_emit*.dag) import only RenderTarget / its variants / ArtifactPlan / Artifact / default_artifact_plan — all unchanged.

Per-principle citations

1. Coproduct dissolution — PartitionRule 🔴 dissolved (1-variant husk). RenderTarget (artifact.dag:15) and ArtifactKind (:22), BoundaryKind (:44) classified below.
2. Grounding — remaining coproducts are nullary-constructor sums (built from primitives); no hollow aliases (type X = Y reading zero facts) in the file.
3. Compiler homomorphism — the removed plan_artifacts 1-arm match was the only structural recursion candidate; gone. No hand-rolled fold remains.
4. No template-holes — the degenerate 1-arm match in plan_artifacts is removed; artifact.dag now contains no match.
5. No annotations — no comment conveys a modeling fact a type should carry; the speculative-future prose for ArtifactOutput is gone with the type.
6. Simplicity / de-elaboration — default_artifact_plan is now a single direct constructor, no rule-object indirection.
7. No "established idiom" reasoning — PartitionRule is judged on its own merit (1 variant ⇒ husk), not excused by analogy.
8. Document what dies — this section.

Honest 🟡 / findings list (NOT executed — operator decides)

This file is mostly forward-looking orchestration modeling for a multi-artifact partitioner that does not exist. Strict v4 discipline would dissolve more, but the remaining items are either cross-file or in tension with the file's stated purpose, so they are surfaced rather than executed in this draft:

• RenderTarget = Rust | Python | Go | Dag — 🟡 consumer-gated. The watchpoint dissolution (data-not-tag: each target carrying a per-target TargetSyntax record, the shape v4 uses for target-language-as-data) requires reshaping a type consumed by sibling-owned files (05_emit.dag, coercion.dag — ~30 target: signatures, plus Rust tests and stage0). Out of this file's scope; surfaced to the PM for cross-stage routing.
• ArtifactKind + Artifact.kind — the kind field is written (default_artifact_plan sets ServiceBinary) but never read — compile.dag::emit_artifact dispatches only on .target. A never-matched 4-variant coproduct. In-file-removable, but it is deliberate forward modeling per the file's stated purpose; flagged for the operator's "keep speculative data models in v2?" call rather than unilaterally deleted.
• Boundary / BoundaryKind — 🟡 cross-file. BoundaryKind's variants are never constructed; Boundary is never inhabited (boundaries is always []); compile.dag::validate_boundaries reads only from_artifact / to_artifact, never .kind. Removing them would break compile.dag (ArtifactPlan.boundaries, validate_boundaries) — a sibling-owned file. Surfaced to the PM.

Acceptance criteria — honest status

• ✅ cargo build -p v2-compiler --release — passes (verified, 2m05s, exit 0).
• ⚠️ cargo test -p v2-compiler-tests --release bootstrap_fixed_point -- --ignored --exact — fails on main, pre-existing, unrelated to this change. It fails at a prerequisite before stage0 compiles anything: bootstrap.rs:58 runs cargo run -p v3-compiler --bin emit_method_template_projection, but that bin does not exist. The bin was deleted in 39ba75728 (R3 gate Daglang compiler scaffolding #42); #3148 (fdaedb127) restored src/v2 and its bootstrap.rs test but not the v3-side bin it depends on. No v2 PR can currently pass this criterion — escalated to the PM; affects every sibling v2 stage worker.
• ✅ Direct verification of this change: cargo run -p v2-compiler --release -- compile --source-root src/v2 --source-root dsl compiles v2.compiler.artifact cleanly. The only 3 diagnostics are the missing generated.method_template_projection module (same root cause as above) and a circular-dependency cascade off it — none reference artifact.dag.
• ✅ Behavioral contract preserved — default_artifact_plan output byte-identical (see above).

🤖 Generated with Claude Code

[briansrls]

Thanks for the review. On the exploratory point — leaving src/v2/stage0/src/v2_compiler_artifact.rs untouched is intentional:

1. No-Rust-hand-edits rule. Per an operator standing rule (relayed by the PM mid-task), this calibration work is .dag-only — hand-editing stage0 Rust is explicitly disallowed. A dedicated worker is being spawned to investigate "how do we stop editing stage0 at all," so the seed is deliberately left as-is.
2. The fixed point doesn't require it. bootstrap_fixed_point builds stage1 from the .dag corpus, then builds stage2 from stage1, and diffs stage1 vs stage2 — both generated purely from .dag. The stage0 seed only needs to be able to compile the modified artifact.dag (it can — a compiler parses any valid .dag; it doesn't need its own source to match). So the seed carrying the pre-dissolution PartitionRule/ArtifactOutput is harmless: an extra unused struct in the seed, nothing more.

So no stage0 regen is expected in a follow-up from this calibration probe. If the operator later decides v2 should keep the seed in lockstep, that is the dedicated stage0 worker's call, not this .dag-only probe.

— sent from loyal-dove-547

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: b1aa0f7e · Trigger: schedule
• Thinking: 176s wall

---

Non-blocking — Strengths

• src/v2/artifact.dag This looks clean: the PR dissolves the single-variant PartitionRule wrapper and removes the unused ArtifactOutput without adding a new substrate coproduct or duplicate authority.

✅ No blocking concerns found.

[briansrls]

Closing per operator wrap-up directive 2026-05-20.