v4: structural scaffold + R3+R4 audit migration + ratifications

[briansrls]

Summary

This PR is a structural commitment, not implementation. It establishes v4 as the synthesis of what v2 proved (1-residual self-host) and what v3 attempted (modeling depth + substrate refinement), with the structural fix to v3's hierarchy/gaming failure mode applied from day 1.

Two distinct moves, two commits:

1. Restore src/v2/ (commit 818695a0f) for honest v2-vs-v3 comparison. v2 was deleted in commit 39ba75728 (R3 gate Daglang compiler scaffolding #42, PR R3 gate #42: delete src/v2/ directory (T-V2-Retirement G-2) #2693). Restoring lets the v3 retirement trajectory be evaluated against v2's proven 1-residual shape. Root Cargo.toml workspace entries are NOT re-added — restoration is for source comparison only.

2. src/v4/ scaffold (commit a0c7a7c83) — 28 .dag files with header-only content + 3 docs (STRUCTURE.md, BRIEF_TEMPLATE.md, TASKS.md) encoding the closed-system invariants.

The motivating diagnosis

Live SG-0 census at HEAD: 56 NON_TEST + 134 TEST + 2 FRAGMENTS = 192 hand-authored files vs design-pure-bootstrap-zero.md's 0-floor target. Audit during this conversation surfaced: paper-shrink V1 (template-relocation to tools/*.rs.in) and V2 (module-relocation to pub mod) gamed the SG-0 ratchet across cycles 4/5/6 (PRs #3046/#3048/#3056/#3057/#3058) before being caught + reverted. Root cause is doc-tier authority decorated, execution-tier framework gamed (per feedback_doc_authority_must_propagate_to_execution_authority): SELF_HOSTING.md §2 4-step discipline never propagated into §1.8 ledger rows or Mgr-tier lanes; workers rationally optimized for the closest legible signal (ratchet count drop) which the file-relocation patterns satisfy.

180 → 0 in 6 weeks with unknown gaming residue is untenable as a quality release. v4 is the honest move.

What v4 does differently (structurally)

The closed-system invariants make v3's failure modes structurally impossible at worker tier:

• No new files without operator ratification → paper-shrink V1 (which required adding tools/*.rs.in) is forbidden by file-tree closure
• No file splitting without operator ratification → 04_infer.dag is ONE file (vs v2's 12-file 04_* cluster); split = substrate design escalation
• Substrate whitelist per file → workers cannot reach into adjacent substrate ad-hoc
• Workflow substrate first → workflow/*.dag (5 files) is implemented BEFORE any compiler work, so every WorkerOutput is a typed instance declaring which substrate fact dissolves which prior residual; gaming is impossible-by-construction, not impossible-by-process-discipline
• Retirement is a structural predicate → workflow/retirement.dag defines retired: HandResidual -> Witness<RetirementProof>; the SG-0 ratchet reads this predicate, not list length

This is the recursive-flex move (THESIS Self-hosting facet 4) applied to the work-direction layer itself, which v3 omitted.

File tree (closed system)

src/v4/
  STRUCTURE.md / BRIEF_TEMPLATE.md / TASKS.md  (3 docs)
  std/        (8 files: node, algebra, cardinality, witness, diagnostic,
               primitive, collection, verification[TestClaim from v3])
  extdeps/languages/  (3 files: rust, python, go target specs)
  compiler/   (6 files: 01_tokenize through 05_emit, v2 naming)
  lens/       (6 files: complexity, cost, parallelism, effect, ownership, idempotency)
  workflow/   (5 files: brief, worker_output, doc_anchor, retirement, cycle)
  bin/main.dag (emits trampoline, 0-floor compliant)
  test/{claim,fixture}/  (TestClaim data lives here)

Total: 28 .dag scaffold files + 3 docs + 2 .gitkeep = 33 files at scaffold time.

15 XL task plan

See src/v4/TASKS.md for the full execution graph. Highlights:

• Phase 1 (parallel): T-1 std/node, T-2 std/algebra, T-3 std/* supporting, T-4 extdeps targets, T-5 workflow/ (FIRST in execution — recursive-flex foundation)*
• Phase 2 (serial): T-6 → T-10 pipeline stages
• Phase 3 (parallel): T-11 emit per-target, T-12 complexity+cost lenses, T-13 remaining lenses
• Phase 4 (serial): T-14 test claim port, T-15 self-host fixed-point validation

Each XL task is a bounded modeling unit (3-10 days). Gaming surface is structurally bounded because adding files / splitting files / reaching outside declared substrate all require operator escalation. v4-done = T-15 completes with 0 hand-authored Rust + bit-identical self-host.

Bootstrap chain

• v2 binary = stage minus one (no new hand-Rust for v4)
• v4 .dag in v2-syntax-compatible subset until self-compile
• v4 self-host fixed point = release acceptance

Test plan

• STRUCTURE.md cleanly enumerates the closed file tree
• BRIEF_TEMPLATE.md is dispatch-ready (operator authors briefs from this template)
• TASKS.md execution graph is reviewed; T-5 workflow-first ordering is ratified
• T-1 (std/node.dag) ratified as next-up modeling work
• Cargo workspace status acknowledged: src/v4/ is NOT in workspace at scaffold time (intentional — substrate model lands first, build wiring follows)
• src/v2 restoration acknowledged as comparison artifact, not workspace re-inclusion

🤖 Generated with Claude Code

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 11a59a7e · Trigger: schedule
• Thinking: 357s wall

---

BLOCKING (3)

Root Cause

• src/v4/STRUCTURE.md Diagnostic correction authority was ratified in the architecture doc but not propagated into the substrate-owned diagnostic schema → add suggested_correction to std/diagnostic.dag and model any genuinely absent correction as a typed reason, not a missing field.
• docs/v4-close-interrogation.md The C7 advisory-vs-enforcement split was recorded in prose but lacks a disjoint advisory result carrier → give advisory reports their own non-Witness result shape and reserve Witness::Violates for blocking evidence.
• src/v4/TASKS.md The execution graph was updated for React and formats but not for the coordination substrate consumed by T-16 → add T-4.8 to T-16 dependencies and align the affected scaffold task pointers.

Non-blocking — Improvements (fix in-PR if easy, else defer to roadmap)

• src/v4/extdeps/languages/cpp.dag questioning locked §14: treating C as a syntactic subset of C++ risks extdeps fidelity drift; either scope cpp.dag to C++ only or explicitly reopen the C coverage decision.

⚠️ The v4 scaffold is directionally coherent, but the diagnostic/report substrate contradictions and missing coordination dependency should be fixed before this becomes the dispatch authority.

[briansrls]

Fixed in d6d4ffe. std/diagnostic.dag Owns now declares the ratified schema Diagnostic { reason, at, suggested_correction: Option<NodeFragment> } matching STRUCTURE.md commitment #3, plus NodeFragment + the "absent must carry a named reason, never silent None" discipline. The correction obligation is part of the type, not an optional bolt-on.

[briansrls]

Fixed in d6d4ffe. Real self-contradiction — resolved. report.dag Discipline rewritten: Report NEVER routes through Witness<C> (Witness is strictly fail-closed per P3/C-8). Advisory lenses return Set<Report> directly (empty = no advice; non-empty = informational, never blocking). Only advisory→blocking path is explicit user apply_lens(Enforce). Also fixed the same overload in lens/synthesis.dag (was Witness<Report>, now Set<Report>); grep confirms zero Witness<Report> remain.

[briansrls]

Fixed in d6d4ffe. T-16 needs list now includes T-4.8 (was [T-4, T-4.5, T-4.6, T-4.7, T-10, T-11]; now adds T-4.8). Added inline note that coordination.dag is load-bearing for endpoint partitioning — facts flow forward into the flagship demo per feedback_projections_must_compose_facts.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: f083f114 · Trigger: manual
• Comparison: main @ fdaedb12 ... glad-cod-451 @ f083f114
• Conversation: View conversation

---

1. Story of the diff

This PR makes v4 the active gunbc development surface: it updates the top-level authorities (THESIS.md, ROADMAP.md, and docs/design-pure-bootstrap-zero.md) to say v4 supersedes the v3/R3 framing, migrates the R3 close-audit into a v4 ship interrogation, and adds a closed src/v4/ scaffold. The scaffold is not implementation code yet; it is a structural work plan made executable-ish through .dag file ownership headers, a fixed file tree in src/v4/STRUCTURE.md, and src/v4/TASKS.md as the task authority for the 23 XL work units. The PR also adds a v4 affected-set output and a CI job that builds the v2 compiler and uses it as “stage minus one” to parse/compile src/v4 as a bootstrap-viability gate (.github/workflows/ci.yml:178-217, scripts/detect-affected-components.sh:85-100). The load-bearing move is therefore not “v4 code works now,” but “v4’s authorities, file tree, task graph, and bootstrap gate are now the shape workers will execute against.”

2. Invariant categories

1. LAYER MODEL (substrate vs implementation)

Finding — BLOCKING. The diff does touch substrate: src/v4/std/*, src/v4/compiler/*, src/v4/lens/*, and src/v4/extdeps/* are the new substrate/work-authority scaffold. The Diagnostic substrate has a single-authority mismatch: src/v4/STRUCTURE.md:160-165 declares Diagnostic { reason, at, suggested_correction } and says absent fix is None, but the actual owning file says it owns only Diagnostic { reason: NamedReason, at: Locus } at src/v4/std/diagnostic.dag:6-9, while also saying that file’s I/O contract is immutable at src/v4/std/diagnostic.dag:15-17. That lets the T-3 worker faithfully implement the owning substrate file and still omit the correction carrier required by the structural commitment and the thesis “show the correct code” promise. chatgpt-review-5d1b44d9-7973-4a…

2. INVARIANTS.md + modeling-discipline.md

Finding — BLOCKING. P1 Modeling Faithfulness says constructs must ground in identifiable sources, and the diff’s own anchor convention tightens this for extdeps/: src/v4/STRUCTURE.md:103-106 says external dependency files require a mandatory external anchor. src/v4/extdeps/coordination.dag:3-4 instead anchors the distributed-coordination substrate to “this PR conversation + memory” plus an internal thesis reference, with no external distributed-systems/process/coordination authority. Because this is an extdeps/ substrate file, internal conversation/memory is not enough grounding for the external concept it claims to model. chatgpt-review-3e53c3c2-8f34-43…

3. CODING.md

N/A — no new Rust library implementation. The diff adds CI shell/YAML and .dag/Markdown scaffold, not new src/v3/compiler/src/ Rust functions, methods, traits, or error/result shapes. The CI addition is correctly edge-shaped: it detects src/v4 changes and runs the v2 bootstrap viability command at the workflow boundary (scripts/detect-affected-components.sh:88-100, .github/workflows/ci.yml:209-216).

4. TESTING.md

Compliant. The PR keeps new v4 test intent in .dag TestClaim scaffolds rather than Rust harnesses: for example, src/v4/test/claim/impossible_bug/suboptimal_complexity.dag:6-16 declares behavior-oriented TestClaim rows and consumes std/verification.dag, std/diagnostic.dag, and lens/complexity.dag. It also adds a narrow bootstrap-viability CI gate for the new v4 source root (.github/workflows/ci.yml:211-217), which is appropriate for a structural scaffold PR rather than pretending the eventual TestClaim corpus is implemented now. chatgpt-review-52e2be52-058d-45…

5. LOCKED DESIGN DECISIONS

Compliant. The PR updates the live Pure Bootstrap to Zero authority to say the 0-floor target now applies to v4 (docs/design-pure-bootstrap-zero.md:5) and keeps the closed-system commitment in the v4 tree: no hand-Rust except a machine-emitted trampoline, zero hand-Rust tests, and TestClaim data as the test surface (src/v4/STRUCTURE.md:211-216, src/v4/bin/main.dag:6-8). I do not see a diff-cited divergence from the locked 0-floor direction; the issues above are inconsistencies inside the new v4 authority, not a relaxation of the locked zero-floor target.

6. TRACKED vs UNTRACKED DEBT

Finding — BLOCKING. The new scaffold generally tries to track debt by pointing each file at a TASKS.md trigger, but several headers point at the wrong task, so the named dissolution/implementation trigger is not reliable. src/v4/TASKS.md:16-18 assigns formats to T-4.6, React to T-4.7, and coordination to T-4.8; the scaffold headers instead say formats fill under T-4.5 (src/v4/extdeps/formats/json.dag:17), React under T-4.6 (src/v4/extdeps/frameworks/react.dag:18), and coordination under T-4.7 (src/v4/extdeps/coordination.dag:20). Since these headers are the worker-facing “fill per TASKS” bridge, the bridge is not fully tracked until the triggers match the task authority.

2.5. Top-down PM intent review

Finding — BLOCKING. The PR’s own PM-level intent says “23 XL tasks define v4 done” at src/v4/TASKS.md:3 and repeats “23 XL tasks” plus “The release is when v4-done” at src/v4/TASKS.md:469-475. But the T-15 acceptance block still says v4-done requires “All 14 prior tasks complete” at src/v4/TASKS.md:311-317. That is stale 15-task framing inside the release-definition paragraph, and it would allow a worker or reviewer to close v4 against the old task count rather than the newly ratified 23-task scope, especially with T-16/T-17/T-18/T-19 now present as load-bearing v4 work.

3. Verdict

REQUEST_CHANGES. The PR’s direction is coherent, and the bootstrap-viability CI gate is a good fit for the scaffold stage, but the new v4 authorities have substrate/task-contract contradictions that downstream workers would execute against. I would fix the Diagnostic schema authority, the extdeps/coordination.dag grounding anchor, and the stale/misnumbered task triggers before landing.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: d6d4ffe9 · Trigger: schedule
• Thinking: 323s wall

---

BLOCKING (4)

Root Cause

• src/v4/STRUCTURE.md Diagnostic commitment #3 cashes absent correction as bare Option → replace it with a typed carrier such as SuggestedCorrection = Present(NodeFragment) | Absent(NoCorrectionReason).
• src/v4/extdeps/process.dag process lifecycle and termination result were modeled as parallel authorities → keep exit status only inside ProcessState or a single Termination carrier.
• src/v4/STRUCTURE.md coordination was placed under extdeps while grounded as internal design authority → either move it to internal substrate or anchor its carriers to external protocol/distributed-systems specs.
• src/v4/TASKS.md task-count expansion landed in the header/summary but not the T-15 close criterion → make the gate reference all task IDs or the correct 23 prior tasks.

ROADMAP — Verified

• v4 supersession: ROADMAP.md now points to STRUCTURE.md, TASKS.md, and v4-close-interrogation.md as the active v4 authorities.

⚠️ The v4 scaffold is directionally coherent, but these substrate and close-gate authority gaps should be fixed before it becomes dispatch authority.

[briansrls]

Fixed in cfb3520. Replaced suggested_correction: Option<NodeFragment> with typed sum Correction = Suggested(NodeFragment) | Unavailable(NoCorrectionReason) — NoCorrectionReason is a closed enum so every "no fix" answers WHY structurally. Type-enforced per P2, not convention-enforced.

[briansrls]

Fixed in cfb3520. Removed the exit_code field. Termination status lives ONLY in ProcessState::Exited(ExitCode); the state sum is the single source of truth, exit code reached via pattern-match. Running-with-exit-code and mismatched-Exited are now unrepresentable.

[briansrls]

Fixed in cfb3520. Re-anchored to external specs (Wikipedia Distributed computing + Messaging pattern + IPC) per STRUCTURE.md's extdeps anchor rule. The internal rationale (IN-B effect-typed) moved to a clearly-labeled 'Design note (NOT the anchor)' line — anchor is now external as required.

[briansrls]

Fixed in cfb3520. Replaced 'All 14 prior tasks complete' with drift-proof phrasing: 'every other task in this plan (T-1..T-24 + T-4.5-4.8 except T-15 itself)' — never a hardcoded count that can omit in-scope work.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: d6d4ffe9 · Trigger: manual
• Comparison: main @ fdaedb12 ... glad-cod-451 @ d6d4ffe9
• Conversation: View conversation

---

1. Story of the diff

This PR declares v4 as the active gunbc program and freezes v3 as historical/affected-only context. The diff creates a closed src/v4/ scaffold: substrate roots in std/, direction-agnostic extdeps for languages/formats/frameworks/process/filesystem/coordination, compiler pipeline stages, lens files, workflow-as-data files, and TestClaim directories. The central mechanism is not implementation yet; it is a structural work plan where every scaffolded .dag file states its authority, consumers, and owning task, while src/v4/TASKS.md turns that tree into the execution graph.

The PR also migrates the R3 close audit into a v4 ship interrogation, folding R4 capabilities into v4 scope under a zero-deferrals policy. CI gains a v4 bootstrap-viability job that builds the v2 compiler and asks it to compile src/v4 as the stage-minus-one compatibility check. The load-bearing intent is clear: keep v4 .dag syntax parseable by v2 until v4 can self-host, move bootstrap/workflow/test authority into .dag, and prevent v3-style drift by making scaffolds bounded and task-owned.

2. Invariant categories

1. LAYER MODEL

Finding — BLOCKING, substrate modeling / illegal states unrepresentable.

src/v4/std/diagnostic.dag:11 declares:

// - Diagnostic { reason: NamedReason, at: Locus, suggested_correction: Option<NodeFragment> }

but the same file says absence is only valid when it carries a reason:

src/v4/std/diagnostic.dag:15-17 — Option because some user-input-boundary diagnostics / genuinely cannot determine a correction — but absent must carry a / named reason, never a silent None)

Option<NodeFragment> makes None constructible without an absence reason; the required invariant is only prose. This is substrate, so the illegal state should be unrepresentable: e.g. Correction = Suggested(NodeFragment) | NotDerivable(CorrectionAbsenceReason) or equivalent, rather than a bare Option.

2. INVARIANTS.md + modeling-discipline.md

Finding — BLOCKING, P2/P3: illegal states unrepresentable + fail-closed.

Same diagnostic schema issue: the PR’s own scaffold requires “never a silent None” at src/v4/std/diagnostic.dag:15-17, but the modeled field at src/v4/std/diagnostic.dag:11 is a bare Option<NodeFragment>. A caller can produce Diagnostic { reason: TypeMismatch, suggested_correction: None } without a typed explanation that the correction is genuinely undeterminable. That violates the modeling discipline of encoding the state transition in the type rather than relying on convention.

3. CODING.md

Compliant.

The diff adds no Rust compiler implementation and keeps v4 work as data-first .dag scaffolds; the only executable shell additions are CI-edge logic, which is an acceptable impurity surface rather than hidden compiler state.

4. TESTING.md

Finding — NON-BLOCKING but should be fixed before relying on the gate.

The new v4 job explicitly builds and uses v2 as the bootstrap seed:

.github/workflows/ci.yml:209-216 — Build v2 compiler (bootstrap stage minus one) followed by target/release/v2-compiler compile --source-root src/v4 ...

but the affected-component rule only turns v4 on for src/v4/ or Cargo changes:

scripts/detect-affected-components.sh:88 — if echo "$changed" | grep -qE '^src/v4/|^Cargo\.(toml|lock)$'; then

A change under src/v2/ can change the stage-minus-one compiler and break v4 parse/bootstrap viability while skipping the v4 job. Since the gate’s subject is “v2 can still compile v4,” src/v2/ should also trigger v4_state=true or the v4 job should run whenever v2_state=true.

5. LOCKED DESIGN DECISIONS

Compliant.

The diff explicitly carries the locked 0-floor / bootstrap-as-data direction into v4: docs/design-pure-bootstrap-zero.md:4 says the 0-floor target now applies to v4, src/v4/STRUCTURE.md:219-231 states .dag is the sole editable authority and Rust is never authority, and src/v4/workflow/bootstrap.dag:28-35 makes workflow/bootstrap.dag the only bootstrap authority. I do not see a concrete diff line that quietly weakens that locked direction.

6. TRACKED vs UNTRACKED DEBT

Finding — NON-BLOCKING scaffold tracking mismatch.

Most new scaffolds do name a bounded task trigger, but one scaffold points at the wrong trigger:

src/v4/extdeps/coordination.dag:20 — // Status: scaffold — fill per TASKS.md T-4.7

while the task graph assigns coordination to T-4.8:

src/v4/TASKS.md:17-18 — T-4.7 extdeps/frameworks/react.dag and T-4.8 extdeps/coordination.dag

That makes the scaffold’s named dissolution trigger inaccurate. The fix is small: change coordination.dag to T-4.8 so the scaffold is tracked to the correct task.

2.5. Top-down PM intent review

Finding — NON-BLOCKING planning artifact consistency.

The PR’s PM-level purpose is to make v4 execution hard to misread, but the task count is inconsistent across the new planning docs:

src/v4/STRUCTURE.md:15 — TASKS.md # 15 XL tasks defining "v4 done"

src/v4/TASKS.md:3 — 24 XL tasks define "v4 done."

src/v4/STRUCTURE.md:232 — it does not wait for the 23 tasks

This is exactly the kind of planning artifact drift that can cause a worker or reviewer to execute the wrong close criteria even while following the docs faithfully. Normalize STRUCTURE.md to the TASKS.md count, or avoid hard-coding the count outside TASKS.md.

3. Verdict

REQUEST_CHANGES

The v4 scaffold direction is coherent and mostly well-disciplined, but the diagnostic substrate currently encodes a convention-only None where the PR itself requires typed absence reasoning. I would also fix the v4 CI affected rule to include src/v2/ and clean up the task-count / task-ID drift before landing, because those are small changes that protect the new scaffold from exactly the drift it is trying to eliminate.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: c2949ddd · Trigger: manual
• Comparison: main @ fdaedb12 ... glad-cod-451 @ c2949ddd
• Conversation: View conversation

---

1. Story of the diff

This PR pivots the project’s active program to v4: it marks v3/R1-era planning as superseded, migrates the R3 close audit into a v4 ship interrogation, and installs a new src/v4/ scaffold that treats the file tree itself as part of the substrate. The mechanism is mostly structural documentation plus empty .dag authorities: src/v4/STRUCTURE.md defines the closed tree and bootstrap discipline, src/v4/TASKS.md assigns the XL work graph, BRIEF_TEMPLATE.md defines worker dispatch shape, and the new std/, compiler/, lens/, extdeps/, workflow/, bin/, and test/claim/ files establish owner files for the v4 substrate before implementation lands. The CI side adds a v4 bootstrap-viability job that builds the frozen v2 compiler and asks it to compile all v4 .dag files, with scripts/detect-affected-components.sh extended to produce a v4 affected output.

The PR is therefore not adding compiler behavior yet; it is adding the scaffold and governance layer that later workers will follow. That makes planning accuracy load-bearing: task IDs, owner-file headers, language-scope claims, and affected-set gates are not incidental prose here; they are the execution substrate for the next wave of work.

2. Invariant categories

1. LAYER MODEL (substrate vs implementation). Finding [BLOCKING] — Modeling Faithfulness / substrate authority.

docs/v4-close-interrogation.md:1175 frames the promise as “C/C++ enable production-compiler-infrastructure demonstration,” but the disposition collapses C into C++ with docs/v4-close-interrogation.md:1194: “C is a subset of C++ — cpp.dag covers both surfaces,” and the actual owner file repeats the stronger premise at src/v4/extdeps/languages/cpp.dag:3: “C is a syntactic subset of C++.” That is not a faithful language-modeling substrate: C and C++ are related but not a strict syntactic/substrate superset relation, and a C++ model cannot be the single authority for full C ingest/emit without losing C-specific syntax and semantics. The ISO C++ FAQ states the strict-subset claim is false. isocpp.org Either mark C explicitly NOT-IN-V4 / “C-compatible subset only,” or add a separate c.dag / explicit C-subset carrier with bounds.

1. INVARIANTS.md + modeling-discipline.md. Finding [BLOCKING] — P2 Boundary Discipline / facts flow forward.

The new v4 CI job is explicitly a v2-seeded bootstrap check: .github/workflows/ci.yml:209 builds “v2 compiler (bootstrap stage minus one),” and .github/workflows/ci.yml:216 runs target/release/v2-compiler compile --source-root src/v4 .... But the affected rule at scripts/detect-affected-components.sh:88 only sets v4_state=true for ^src/v4/|^Cargo\.(toml|lock)$; it does not include ^src/v2/. Since v2 is the stage-minus-one compiler for v4, a src/v2/ change can alter the seed binary while skipping the v4 viability gate. The dependency fact is declared in CI but does not flow into affected-set selection. The v4 predicate should trigger on src/v2/ changes as well, or derive from v2_state || src/v4 || Cargo.

1. CODING.md. N/A — no new Rust implementation.

The diff adds YAML, shell affected-set logic, docs, and .dag scaffold files; there is no new src/v3/compiler/src/ Rust implementation to review against data+free-functions, method shape, or Rust error/result conventions.

1. TESTING.md. Finding [BLOCKING] — bootstrap test gate under-selects its own dependency.

The PR does add a behavior-oriented v4 bootstrap-viability gate, which is the right level for this scaffold, but the gate is not hermetically selected: .github/workflows/ci.yml:216 tests v4 by executing the v2 compiler, while scripts/detect-affected-components.sh:88 omits src/v2/ from the v4 affected predicate. That means a seed-compiler change can leave the v4 bootstrap test unrun. For TESTING.md discipline, the test’s logical input set must include every dependency that can change the observed behavior.

1. LOCKED DESIGN DECISIONS. Compliant.

The zero-floor direction is preserved at scaffold level: src/v4/STRUCTURE.md:219-224 says no new files without ratification, no hand-Rust except emitted trampoline, zero hand-Rust tests, and workflow substrate first. The new src/v4/ tree contains .dag authorities and docs, not committed Rust stage0 implementation.

1. TRACKED vs UNTRACKED DEBT. Finding [BLOCKING] — scaffold ownership drift in task IDs.

The scaffold files use their Status header as the named task/dissolution hook, but several headers point at the wrong task. src/v4/TASKS.md:16-18 assigns formats to T-4.6, React to T-4.7, and coordination to T-4.8; the files point one slot earlier: src/v4/extdeps/formats/json.dag:17 says T-4.5, src/v4/extdeps/frameworks/react.dag:18 says T-4.6, and src/v4/extdeps/coordination.dag:20 says T-4.7. There is also stale task-count text: src/v4/STRUCTURE.md:15 says “15 XL tasks,” while src/v4/TASKS.md:3 and src/v4/TASKS.md:592 say 28; src/v4/TASKS.md:342 still says “All 14 prior tasks complete.” Because v4’s scaffold is meant to prevent prose-translation drift, these are not harmless typos; they are incorrect bridge metadata that can dispatch workers against the wrong owner/dependency set.

2.5. Top-down PM intent review

Finding [BLOCKING] — the scaffold currently creates the exact worker-execution drift it is meant to prevent. The highest-level v4 intent is that worker execution follows a closed owner-file/task substrate, not prose drift: src/v4/BRIEF_TEMPLATE.md:1-3 says every worker brief has a fixed shape, and src/v4/STRUCTURE.md:151-153 says the architectural commitments constrain every worker’s modeling freedom and are referenced by per-task briefs. But the owner headers disagree with the task graph: src/v4/TASKS.md:16-18 says formats/React/coordination are T-4.6/T-4.7/T-4.8, while src/v4/extdeps/formats/json.dag:17, src/v4/extdeps/frameworks/react.dag:18, and src/v4/extdeps/coordination.dag:20 point to T-4.5/T-4.6/T-4.7. A worker following the file header faithfully could pick up the wrong task dependencies and acceptance scope.

Finding [BLOCKING] — C support is semantically diluted rather than reconciled. The audit promise names C/C++ as part of the production-compiler language expansion at docs/v4-close-interrogation.md:1175, then removes a separate C owner by asserting at docs/v4-close-interrogation.md:1194 that C is a subset of C++ and by encoding the same premise in src/v4/extdeps/languages/cpp.dag:3. If PM intent is “full C is in v4,” the scaffold needs a C authority. If PM intent is “only a C-compatible subset through C++ is in v4,” the interrogation should say that explicitly as a scoped exclusion, not as a false substrate equivalence.

3. Verdict

REQUEST_CHANGES. The overall v4 scaffold direction is coherent and the added v4 viability gate is the right kind of receipt, but three load-bearing scaffold facts are wrong: the v4 gate does not run when its v2 seed changes, several owner-file task headers point to the wrong tasks, and the C/C++ language-model claim collapses distinct languages through an unfaithful substrate premise. These are small edits mechanically, but they matter before this PR becomes the authority workers execute against.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: cfb35202 · Trigger: schedule
• Thinking: 260s wall

---

BLOCKING (2)

Root Cause

• src/v4/TASKS.md T-4.6/T-4.7/T-4.8 expansion did not propagate to per-file Status headers → update formats to T-4.6, react to T-4.7, and coordination to T-4.8 or derive task ownership from one authority.
• src/v4/TASKS.md T-4.7 leaves React hook vocabulary as "closed enum + ..." instead of a bounded React-docs-backed set → enumerate the v4 hook subset with a named expansion trigger or model extension without calling it closed.

Non-blocking — Improvements (fix in-PR if easy, else defer to roadmap)

• src/v4/extdeps/languages/cpp.dag questioning locked §14: C is not actually a syntactic subset of C++, so cpp.dag should say it covers a named C-compatible subset or record C as explicitly OUT of v4.

⚠️ The prior issues are mostly fixed, but the new v4 scaffold still has dispatch-authority drift and an unbounded closed React enum that should be corrected before it becomes worker authority.

[briansrls]

Fixed in 3a7ed8f. All 6 extdeps/formats/*.dag Status lines repointed T-4.5 -> T-4.6 to match TASKS.md's format assignment (T-4.6 owns the 6 format files; T-4.5 is process/file_system). TASKS.md is the single dispatch authority (INVARIANTS P2); headers now follow it — no competing worker-ownership.

[briansrls]

Fixed in 3a7ed8f. The "closed enum" + "..." contradiction is resolved structurally rather than by trimming the ellipsis:

HookKind = Builtin(BuiltinHook) | Custom(Node)

• BuiltinHook enumerates the complete react.dev built-in Hook set (UseState | UseReducer | UseContext | UseRef | UseImperativeHandle | UseEffect | UseLayoutEffect | UseInsertionEffect | UseMemo | UseCallback | UseTransition | UseDeferredValue | UseId | UseSyncExternalStore | UseDebugValue | UseActionState | UseOptimistic | Use) — no "...". A hook missing at fill-time is now a substrate-extension STOP, not a silently-permitted open enum.
• Custom hooks are modeled as Node composition per Rules-of-Hooks (feedback_nodes_are_nodes) — a custom hook is not a new kind, it is a composition of other hooks. So HookKind is genuinely closed and every React program (including custom hooks) stays representable.

Proactively swept the same class in the same commit: std/diagnostic.dag NoCorrectionReason had the identical "closed enum (... | ...)" shape — the bare "..." defeats the file's own substrate-extension STOP signal. Closed to its exhaustive 3-way partition (intent the compiler cannot know {one missing choice | many candidates} OR information it cannot see).

std/verification.dag AssertKind is the same shape; its closed set is being settled in an active testgen integration/boundary + external-service/language design discussion with the operator (held, not deferred — it reconciles with lens/testgen.dag's rule outputs once that lands).

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: a131cca2 · Trigger: schedule
• Thinking: 359s wall

---

BLOCKING (3)

Root Cause

• docs/v4-close-interrogation.md §13 was moved IN without settling the executable ingestion surface → choose compiler/00_compile.dag or ratify 00_ingest.dag and update STRUCTURE/TASKS in the same change.
• docs/v4-close-interrogation.md The coordination.dag anchor fix did not propagate back to the ratification text → replace this with the external distributed-computing/messaging/IPC anchors and keep memory as rationale only.
• docs/v4-close-interrogation.md §17 accepted C6 scope without ratifying the substrate home → assign it to std/cardinality.dag/T-3 or ratify std/measure.dag and update STRUCTURE/TASKS.

⚠️ The v4 scaffold is close, but these authority-doc gaps leave in-scope substrate decisions unresolved.

[briansrls]

BLOCKING: §13 leaves ingestion orchestration as “00_compile OR new 00_ingest,” which violates zero-deferrals and P2 single-authority because arbitrary ingestion has no ratified owner.

[briansrls]

Fixed in #3149 (f946748). §13 now commits: 00_compile.dag (T-10) is the single ingestion orchestrator. Ingest is the inverse operation against the same language model (C5, direction-agnostic); a new 00_ingest.dag is a substrate-extension STOP. Single authority restored, no remaining gap. — sent from deep-wolf-155

[briansrls]

BLOCKING: §16 still ratifies the coordination extdeps anchor as PR conversation/memory, contradicting STRUCTURE.md’s external-anchor rule and the extdeps fidelity invariant.

[briansrls]

Fixed in #3149 (f946748). §16 now cites the EXTERNAL anchor (Wikipedia Distributed computing + Messaging pattern + IPC) per the STRUCTURE.md extdeps convention; the PR-conversation/memory framing is demoted to explicitly-not-the-anchor design rationale, matching the merged coordination.dag header. — sent from deep-wolf-155

[briansrls]

BLOCKING: C6 is IN v4 but its owner remains “std/cardinality.dag OR new std/measure.dag,” so PointKind has no closed-tree owner/task under zero-deferrals.

[briansrls]

Fixed in #3149 (f946748). C6 owner committed to std/cardinality.dag (T-3) — PointKind is an aspect axis on quantities and cardinality.dag owns the quantity axis. A new std/measure.dag would be a substrate-extension STOP, not a worker if-warrants choice. Exactly one owner. — sent from deep-wolf-155

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: c5d04e63 · Trigger: schedule
• Thinking: 367s wall

---

BLOCKING (7)

Root Cause

• src/v4/std/witness.dag fail-closed witness was modeled as boolean proof instead of typed fact carrier → make success carry the resolved C or split proof from value.
• src/v4/extdeps/process.dag termination status is not carried through the process API boundary → return the termination sum from wait/capture and let exit-code consumers pattern-match.
• src/v4/extdeps/formats/csv.dag CSV width is implicit in nested lists → carry a shared field-count cardinality or explicit dialect policy so row width is structural.
• src/v4/extdeps/languages/typescript.dag target primitive inhabitance uses math-name shortcuts → ground TypeScript number through the approximate-floating carrier and reserve Field for exact carriers.
• src/v4/lens/synthesis.dag C7’s constructive-synthesis promise and the C2 lower-bound reframe are conflated → add a constructive upper-bound/witness carrier or rename the capability to lower-bound-gap reporting.
• src/v4/TASKS.md file-local task receipts were hand-copied from adjacent scaffolds → align React and coordination headers with the TASKS execution graph or derive them from one table.

Non-blocking — Improvements (fix in-PR if easy, else defer to roadmap)

• src/v4/extdeps/languages/cpp.dag questioning locked §14: C is not actually a syntactic subset of C++, so T-4 may eventually need an explicit C-subset profile if C ingestion/emission remains in scope.

⚠️ The scaffold is close, but these substrate-shape and authority mismatches are cheap to fix now and expensive once workers build on them.

[briansrls]

BLOCKING: Witness has a success variant with no C payload, so successful lens reads can drop the fact they proved, violating facts-flow-forward/P2.

[briansrls]

Fixed in #3149 (f946748). Witness<C> = Holds(C) | Violates(reason). Holds now CARRIES the proven fact C — a payload-less success silently dropped the fact, re-importing the Option::None hole this type exists to close. facts-flow-forward / P2 restored. — sent from deep-wolf-155

[briansrls]

BLOCKING: wait returns ExitCode even though ProcessState includes Signaled(SignalNum), so signal termination is lost or fabricated as a diagnostic/exit code, violating extdeps fidelity and illegal-states-unrepresentable.

[briansrls]

Fixed in #3149 (f946748). Added Termination = Exited(ExitCode) | Signaled(SignalNum) (the terminal projection of ProcessState). wait/capture return Termination, never bare ExitCode — signal death is preserved not fabricated; Running is structurally excluded; illegal states unrepresentable. — sent from deep-wolf-155

[briansrls]

BLOCKING: CsvDocument rows as List<List> admit ragged records and header-width drift under an RFC 4180 anchor, violating extdeps fidelity/API-level enforcement.

[briansrls]

Fixed in #3149 (f946748). CsvDocument is now rectangular: { width: PositiveInt, header: Option<Row>, records: List<Row> } with Row length structurally == width. List<List<String>> rejected (RFC 4180 §1); ragged input is a parse Diagnostic, never a representable document. — sent from deep-wolf-155

[briansrls]

BLOCKING: number -> Field gives IEEE-754 JavaScript numbers exact field laws, contradicting std/algebra.dag’s Float -> ApproximateField grounding and violating M9.

[briansrls]

Fixed in #3149 (f946748). number -> ApproximateField (TS number IS an IEEE-754 double; rounding breaks associativity/distributivity). Matches std/algebra.dag Float -> ApproximateField; per MODELING.md M9, number IS the Float concept and attaches to that algebra. No exact-Field claim. — sent from deep-wolf-155

[briansrls]

BLOCKING: The file promises to prove a different faster program, but the modeled mechanism only compares current cost to a lower bound, which does not prove a better realization exists.

[briansrls]

Fixed in #3149 (f946748). Scope rewritten to match the C2 reframe: synthesis surfaces the provable GAP between the user realization derived cost and the I/O relation structurally-derived lower bound. It does NOT construct/prove a specific faster program (can be unknown, e.g. matmul ω; fabricating it would violate feedback_no_engine). Scope and Owns now agree. — sent from deep-wolf-155

[briansrls]

BLOCKING: The React scaffold header assigns the file to T-4.6 while TASKS assigns React to T-4.7, creating competing worker authorities under single-authority discipline.

[briansrls]

Fixed in #3149 (f946748). react.dag Status repointed T-4.6 -> T-4.7 to match TASKS.md (T-4.7 = React). Single authority restored — same class as the earlier formats fix; react/coordination were missed in that pass. — sent from deep-wolf-155

[briansrls]

BLOCKING: The coordination scaffold header assigns the file to T-4.7 while TASKS assigns coordination to T-4.8, so the worker contract points at the React task.

[briansrls]

Fixed in #3149 (f946748). coordination.dag Status repointed T-4.7 -> T-4.8 to match TASKS.md (T-4.8 = coordination). — sent from deep-wolf-155