{T-9+T-10} infer/emit/compile interface-freeze + Wave-1 scaffold + witness anchor

[briansrls]

Summary

T-9/T-10 interface freeze + Wave-1 fail-closed scaffold + Tier-1 v2-compile witness anchor.

• compiler/04_infer.dag (T-9): declares type InferredTree = Node (IR-1: not a new recursive type; A1/B4) and type AlgebraRef = Symbol (T-9 Theme-A audit Codex/graph viz test helpers #2 — Symbol name-reference to algebra inhabitance, NOT a type std/algebra.dag declares). fn infer(tree: ResolvedTree) -> Outcome<InferredTree> returns the fail-closed Rejected branch with a stage-locus infer_coercion_fold_not_realized diagnostic — synthesizing a candidate would leave the decidable Find fragment (DECISIONS C1). The InferredFacts 4-coordinate (resolved_type, inhabits, cost, descent) is documented as co-landing with lens/cost.dag's SymbolicCost declaration (T-12); 04_infer stays a boundary index per IR-1.
• compiler/05_emit.dag (T-10): fn emit(tree: InferredTree, target: TargetModel) -> Outcome<TargetSource> returns fail-closed Rejected with emit_inverse_grammar_walk_not_realized. Wave-1's fail-closed branch upholds the no-templating constraint structurally (TASKS T-10 operator 2026-05-17).
• compiler/00_compile.dag (T-10 orchestrator): single declaration site for type Source = String, type TargetModel = Node (TASKS T-10 Theme-A audit Remove LLM response caching module #9 — a model IS a Node per B2-OMNI), type TargetSource = String. fn compile(source: Source, target: TargetModel) -> Outcome<TargetSource> returns fail-closed Rejected with compile_pipeline_not_realized. Header documents Result vs Outcome literal alignment per DECISIONS item I.
• test/claim/manual/infer_emit_compile_anchor.dag (witness-Acceptance): Tier-1 v2-bootstrap compile anchor mirroring the resolve_compile_anchor.dag shape — three anchor functions that exercise each Wave-1 boundary and the new types. v2-compiler run deferred to T-22 per test/v2_run_preflight/MOVE1_COVERAGE.txt.
• STRUCTURE.md: enumerate the new manual anchor + bump checksum 73→74; align 00_compile / 04_infer / 05_emit scope-line prose with the typed Owns (TargetModel, Outcome<TargetSource>, bounded Find phase).

CP-1b/T-8-close coordination — TargetModel vs TargetSpec

The brief explicitly forbade unilateral cement on the TargetSpec / TargetModel single-spelling reconciliation. Per TASKS T-10:431-439 the operator-ratified Owns / B2-OMNI signature spelling has already moved to TargetModel; the legacy TargetSpec scope-line prose reconciliation is owned by gentle-heron-61's CP-1b/T-8-close train. This PR uses TargetModel exclusively (matching the brief signature and the operator-ratified Owns spelling) and:

• Surfaced the choice up front to gentle-heron-61 via dashboard-message before authoring any code.
• Confined TargetModel declaration to a single site (00_compile.dag) — 05_emit.dag imports it; if T-8-close picks TargetSpec instead, the rename touches one declaration site + three import lines.
• Calls out the coordination in each touched header.

Test plan

• cargo run --quiet --bin v2-compiler -- compile --source-root src/v4 --output-dir /tmp/v4-out → indexed 74 modules from 1 source roots / resolved 74 sources / compiled: 78 files emitted, 0 diagnostics. The interface-freeze types and Wave-1 boundary signatures type-check on the v2 bootstrap graph.
• v2-compiler run against the new anchor functions: deferred until T-22 (matches resolve_compile_anchor.dag posture; DECISIONS CP-1b item 10).

Acceptance criteria (witness, T-9/T-10 interface freeze)

• infer(ResolvedTree), emit(InferredTree, TargetModel), compile(Source, TargetModel) all type-check against their frozen signatures under v2-compiler compile --source-root src/v4 ✅
• Each anchor returns Outcome::Rejected { diagnostic: … } carrying the stage-locus not_realized diagnostic — fail-closed Wave-1 boundary, never a fabricated Produced value (DECISIONS C1 / I) ✅
• v2-compiler run execution is deferred until T-22.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: fc785302 · Trigger: schedule
• Thinking: 308s wall

---

BLOCKING (1)

Root Cause

• src/v4/compiler/04_infer.dag IR-1 is cited but not encoded → declare the InferredFacts coordinate on the infer boundary, or explicitly encode where those four facts live on Node before freezing emit/eval consumers.

⚠️ The interface-freeze needs to preserve the IR-1 inferred facts before this lands.

[briansrls]

Both BLOCKING reviews flagged the same root cause — IR-1 was cited but not encoded; type InferredTree = Node silently dropped the four-field facts coordinate. Addressed in 5506f69 (pushed before this relay arrived): 04_infer.dag now declares InferredFacts { resolved_type, inhabits, cost, descent } flat per IR-1 and InferredTree { node, facts } carries it. CostRef = Symbol mirrors the AlgebraRef = Symbol K-1 name-ref pattern (the SymbolicCost lattice itself stays at lens/cost.dag per T-12). Cardinality and effects remain non-fields per IR-1 (carried by resolved_type's Node structure). v2 compile: 74 modules, 78 emitted, 0 diagnostics. — sent from crisp-cat-892

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 3ab0aea6 · Trigger: schedule
• Thinking: 325s wall

---

BLOCKING (1)

Root Cause

• src/v4/compiler/04_infer.dag IR-1's flat per-node coordinate was encoded as a root wrapper -> encode the facts coordinate for every Node before freezing emit/compile.

⚠️ The prior field set is present, but the interface still drops inferred facts below the root.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 804f479d · Trigger: schedule
• Thinking: 198s wall

---

⚠️ No new blocking findings; request changes remains for the unresolved prior per-node InferredFacts propagation blocker.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 9f151a50 · Trigger: schedule
• Thinking: 318s wall

---

BLOCKING (1)

Root Cause

• src/v4/compiler/00_compile.dag Target boundary carrier ownership is placed on the orchestrator → move TargetModel/TargetSource authority to the emit target contract or an existing non-orchestrator substrate carrier, then have compile consume emit.

⚠️ One new boundary-direction issue should be fixed before freezing this interface, and prior comment 3264313100 remains unresolved.

[briansrls]

BLOCKING: 05_emit imports TargetModel and TargetSource from the compile orchestrator, which inverts the pipeline dependency that compile will need to realize emit ∘ core ∘ ingest and violates INVARIANTS P2 boundary discipline.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 03745206 · Trigger: schedule
• Thinking: 278s wall

---

BLOCKING (1)

Root Cause

• src/v4/compiler/04_infer.dag IR-1/T-9 records the AlgebraRef Symbol disposition but not the cost coordinate's authority → either point CostRef at a declared cost carrier with a named reference target and dissolution receipt, or make the cost coordinate structural before freezing the boundary.

⚠️ The remaining interface-freeze gap is narrow, but it should be fixed before downstream emit/eval/lens consumers depend on this shape.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 1407aa3b · Trigger: schedule
• Thinking: 294s wall

---

BLOCKING (1)

Root Cause

• src/v4/compiler/04_infer.dag IR-1 freezes per-node facts as a partial lookup beside the root Node rather than a total coordinate over the root's Node fold -> make fact coverage structural on InferredTree before freezing the T-9 boundary.

⚠️ The infer boundary still admits incomplete inferred trees, so the interface should be tightened before downstream emit/eval/lens consumers depend on it.

[briansrls]

BLOCKING: InferredTree.facts is an unconstrained Map<Node, InferredFacts>, so a Produced InferredTree can omit facts for the root or children and still type-check, leaving IR-1 per-node fact coverage enforced by convention rather than INVARIANTS P2/API-level enforcement.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 84f99e50 · Trigger: schedule
• Thinking: 253s wall

---

⚠️ No new independent findings, but prior blocking comments on emit's compile import and total InferredTree fact coverage remain unresolved.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: 47cabe98 · Trigger: manual
• Comparison: main @ 020b25cb ... session/crisp-cat-892 @ 47cabe98
• Conversation: View conversation

---

1. Story of the diff

This PR turns the v4 infer/emit/compile seam from mostly documented placeholders into typed .dag interfaces. 04_infer.dag now imports ResolvedTree, exposes infer(...) -> Outcome<InferredTree>, and preserves the IR-1 per-node fact coordinate through InferredTree.facts; 05_emit.dag exposes emit(InferredTree, TargetModel) -> Outcome<TargetSource>; and 00_compile.dag exposes the top-level compile(Source, TargetModel) -> Outcome<TargetSource>. All three stage functions currently fail closed through typed Diagnostic/Rejected values rather than pretending implementation exists, and src/v4/test/claim/manual/infer_emit_compile_anchor.dag gives the v4 whole-tree compile path a manual anchor that constructs enough witness data to typecheck those seams until T-22 can execute them. STRUCTURE.md records the new manual anchor and updates the v4 checksum.

2. Invariant categories

1. LAYER MODEL (substrate vs implementation). — Finding

src/v4/compiler/05_emit.dag:12 imports TargetModel and TargetSource from v4.compiler.compile, while 00_compile.dag is the higher-level orchestrator that says its contract is emit ∘ core ∘ ingest at src/v4/compiler/00_compile.dag:2 and owns compile(...) at src/v4/compiler/00_compile.dag:50. That freezes the dependency direction backwards: the stage that compile must eventually call already depends on the orchestrator module for its interface carriers, so realizing compile by importing emit would introduce an orchestrator↔stage cycle or force another bridge. Boundary Discipline wants each fact in one authoritative place and lower layers not coupled to their caller; move the shared target/source carriers to a lower/common interface authority, or let the stage own the stage-specific carriers and have the orchestrator import them.

1. INVARIANTS.md + modeling-discipline.md. — Finding

src/v4/test/claim/manual/infer_emit_compile_anchor.dag:59 constructs UnknownCost { reason: "" }. This is a C-9-style fabricated empty string for a missing reason, and it is especially telling because the file has just introduced anchor_stub_cost_reason_symbol at src/v4/test/claim/manual/infer_emit_compile_anchor.dag:56 but does not use it. The fail-closed shape elsewhere is good, but the witness anchor should still carry a real typed or at least non-empty reason rather than making “unknown because empty string” a plausible modeled value. Use the named reason carrier or a non-empty String reason that identifies the anchor stub.

1. CODING.md. — Compliant

The changed implementation is .dag, not Rust, and the new stage bodies are small free functions with structured inputs/outputs: compile(...) -> Outcome<TargetSource> at src/v4/compiler/00_compile.dag:50, infer(...) -> Outcome<InferredTree> at src/v4/compiler/04_infer.dag:65, and emit(...) -> Outcome<TargetSource> at src/v4/compiler/05_emit.dag:40. The comments are mostly terse headers/status tags rather than explanatory prose; no Rust method/object style issue is introduced.

1. TESTING.md. — Compliant

The PR adds a .dag manual anchor rather than a Rust harness, and STRUCTURE.md:122 registers it under test/claim/manual. The anchor is explicitly scoped as “compile-only until T-22” at src/v4/test/claim/manual/infer_emit_compile_anchor.dag:5, which matches the current v4 parse/compile witness pattern; no new hand-authored Rust test debt is introduced.

1. LOCKED DESIGN DECISIONS. — Compliant

The diff references the locked/planned seams rather than redefining them: 04_infer.dag:2 ties infer to DECISIONS C1 / IR-1, 05_emit.dag:2 ties emit to DECISIONS C5, and STRUCTURE.md:151 records T-22 as the execution deferral. I do not see a changed line that contradicts those named decisions; the dependency-direction finding above is an implementation/interface placement issue, not an explicit locked-design divergence.

1. TRACKED vs UNTRACKED DEBT. — Compliant, except for the C-9 finding above

The stage scaffolds are documented and bounded: 00_compile.dag:5, 04_infer.dag:5, and 05_emit.dag:5 all mark the functions as T-9/T-10 Wave-1 fail-closed scaffolds, and STRUCTURE.md:151 names T-22 as the run deferral for the manual anchor. The remaining untracked shape I found is not a scaffold-boundary problem but the fabricated empty cost reason at infer_emit_compile_anchor.dag:59, already called out under fail-closed modeling.

2.5. Top-down PM intent review

Compliant. At the PM level, this PR is directionally faithful: it moves compiler pass contracts into v4 .dag, keeps not-yet-realized behavior as typed fail-closed Outcome rejections, and adds a manual witness anchor instead of new hand-Rust. I do not see a concrete diff-cited change that postpones a must-have target, turns dissolution into permanent scaffolding, or asks a worker to execute the wrong plan. The two findings are pre-merge interface hygiene issues: one prevents the frozen seam from creating a future compile↔emit cycle, and the other removes a fabricated witness value.

3. Verdict

REQUEST_CHANGES. The PR has the right fail-closed scaffold direction, but the emit→compile carrier import freezes an inverted layer dependency at the exact interface-freeze seam, and the manual anchor fabricates an empty cost reason. Both are small, localized fixes that should land before this becomes the stable Wave-1 boundary.

[briansrls]

codex #15025 finding (STRUCTURE.md inventory not updated for the new anchor / checksum stale at 75): MOOT per operator-direct directive 2026-05-19 (relayed via vivid-carp-207 → neat-hawk-87, msg_e4531efc). Operator verbatim: "please also delete STRUCTURE.md" — same class as the DECISIONS.md nuke. The STRUCTURE.md manifest-update obligation and "stale/missing-entry" finding-class are explicitly retired; the hand-maintained enumerated-file manifest is killed as over-machinery. The coherent delete+INVARIANTS.md-dangling-ref reconcile is hub-program-dispatched via vessel node://adhoc-93a2d8b7-075, not authored from this lane. No commit needed on this finding. — sent from crisp-cat-892