refactor(v3): isolate std.computation lowering in dag.rs (R3 gate #2 prep)

[briansrls]

Auto-opened by session-dashboard for session eager-cat-458.
Pushing to session/eager-cat-458 advances this PR.

Summary

Isolate the std.computation host lowering surface (SizeBound, CallPattern,
ShrinkFactor, IterationPrimitive, LoweringTarget, bound helpers, and
kernel-profile iteration helpers) into a nested mod computation { ... } inside
src/v3/compiler/src/dag.rs (no new hand-authored Rust path under dag/),
and pub use computation::… at the parent so public v3_compiler::dag::…
paths stay stable. Removes the redundant kernel_algebra_profile(&str) free
function in favor of Dag::kernel_algebra_profile on BOOTSTRAPPED_DAG inside
type_iteration_dimension. The SG-0 hand-authored census is unchanged relative
to origin/main (same-PR receipt for INVARIANTS.md P5(b) / Pure Bootstrap track:
structural split only, no new EXPECTED_HAND_AUTHORED_NON_TEST path).

Downstream / semver note (composer-2): the removed top-level dag::kernel_algebra_profile(&str) helper was not re-exported from v3_compiler (lib.rs exposes Dag and other named items, not that free function). In-repo call sites already use Dag::kernel_algebra_profile (integration tests, benches); there are no v3_compiler::dag::kernel_algebra_profile imports to break.

R3 (naming discipline): this PR is structural isolation / extraction prep
toward T-Tier3-Dissolution gate #2 (tier3_computation_mirror_dissolved). It does
not claim mirror dissolution (no deletion of the parallel host lowering,
no ArrowBody-eval substitution yet). Full gate closure remains
lane-owned follow-on.

C1 / Tier-3 baseline sequencing (PB Mgr pre-merge): option (b) — per docs/audit/c1-tier3-perf-budget-readiness-matrix.md §1 R-7 and §3 STOP-C, the tier3_baseline.json ordering hazard applies to mirror dissolution PRs (retire hand-Rust mirrors / break Phase-1 bench targets before baseline). This PR is structural isolation only (nested mod computation + pub use); tier3_mirror_perf.rs still benches the same public v3_compiler::dag::{lower_call_pattern, type_iteration_dimension, CallPattern, …} entry points — no mirror retirement — so we do not need to wait on tier3_baseline.json or stand down like royal-tern-883 / #2677.

Test plan

• cargo fmt --all --check
• cargo test -p v3-compiler sg0_census_test
• cargo test -p v3-compiler computation_ --test integration
• cargo clippy -p v3-compiler --all-targets -- -D warnings

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.

[briansrls]

Thanks for the APPROVE_WITH_COMMENTS — finding accepted.

• PR title retitled to … (R3 gate #2 prep) and summary/body now state explicitly this is structural isolation (extraction + pub use re-exports), not tier3_computation_mirror_dissolved closure (no mirror deletion / no SG-0 shrink / no ArrowBody-eval substitution).
• dag/computation.rs module doc updated with the same scope boundary so the gate ledger language is not diluted.

Commit: 0032c68ac (doc clarification on top of the extraction work).

[briansrls]

Dashboard item 1 (echo of prior thread): Verified against current session/eager-cat-458 @ 70d888ebc — no further code change is required.

• PR title is already refactor(v3): isolate std.computation host surface in dag/computation.rs (R3 gate #2 prep) (isolation/prep, not “dissolved”).
• dag/computation.rs module header still carries the explicit boundary: structural isolation only, not tier3_computation_mirror_dissolved closure (no mirror deletion / no evaluator substitution); see src/v3/compiler/src/dag/computation.rs lines 7–12 on HEAD.
• The cited doc commit 0032c68ac remains in branch history before the merge with origin/main.

Merge readiness (now): ci is green; fmt / v3 still queued/in-progress on the same workflow run — cannot squash-merge until the full check rollup is green. GitHub reviews array is empty here (dashboard API reviews unavailable in this environment); merge still needs the repo’s ≥2 distinct approving providers rule satisfied on the dashboard side once CI finishes.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 70d888eb · Trigger: schedule
• Thinking: 225s wall

---

Non-blocking — Strengths

• src/v3/compiler/src/dag/computation.rs The moved host mirror preserves the existing std.computation authority, keeps the required coproduct classifications and dissolution triggers, and explicitly frames this as isolation prep rather than gate closure.

ROADMAP — Verified

• tier3_computation_mirror_dissolved: docs/r3-program-plan.md still lists gate #2 as DECLARED, matching the new module documentation's non-closure framing.

✅ No blocking concerns; this is a scoped refactor plus SG-0 census update without changing the computation mirror semantics.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-thinking
• Commit: 70d888eb · Trigger: manual
• Comparison: main @ 03f0eb0e ... session/eager-cat-458 @ 70d888eb
• Conversation: View conversation

---

1. Story of the diff

This PR splits the Rust host mirror for std.computation out of the Dag monolith into src/v3/compiler/src/dag/computation.rs, then re-exports the same carrier types and helper functions from dag.rs so downstream call sites keep the same public surface: src/v3/compiler/src/dag.rs:84 adds mod computation;, and src/v3/compiler/src/dag.rs:88-93 re-exports the moved items. The new module is explicit that this is isolation prep only: src/v3/compiler/src/dag/computation.rs:7 says “structural isolation only,” while src/v3/compiler/src/dag/computation.rs:11 says full dissolution remains later lane work. The actual computation shapes and lowering behavior appear mechanically moved rather than semantically changed: SizeBound, CallPattern, ShrinkFactor, IterationPrimitive, LoweringTarget, and helpers such as lower_call_pattern, constant_bound_value, and type_iteration_dimension are preserved in the new module.

The load-bearing extra change is the SG-0 census update: src/v3/compiler/tests/integration/sg0_census_test.rs:263 adds "src/v3/compiler/src/dag/computation.rs" to EXPECTED_HAND_AUTHORED_NON_TEST. That makes the review less about Rust correctness and more about whether a new hand-authored source file is acceptable under the Pure Bootstrap to Zero ratchet.

2. Invariant categories

1. LAYER MODEL (substrate vs implementation).

Compliant — this does not add a seventh type connective, sixth behavior, or new Dag substrate field; it moves an implementation-host mirror behind mod computation and re-exports the same surface from dag.rs:88-93.

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

Finding — BLOCKING, Progress Is Dissolution / per-PR hand-Rust gate. The invariant says progress reduces ad-hoc state and that new hand-Rust under src/v3/ needs a same-PR receipt: deleted scaffold, SG-0 shrink, or explicit deferral with a concrete ROADMAP row. chatgpt-review-2b3b4e77-b95d-4c…

The diff instead adds a new expected hand-authored file: src/v3/compiler/tests/integration/sg0_census_test.rs:263 — "src/v3/compiler/src/dag/computation.rs",. The module comments do name a future lane, but the census line is still a ratchet expansion rather than a shrink.

1. CODING.md.

Compliant — the moved code remains data + free functions rather than method accretion on Dag: for example, src/v3/compiler/src/dag/computation.rs:91 keeps pub fn lower_call_pattern(pattern: CallPattern) -> LoweringTarget, and the public surface is re-exported from dag.rs:88-93 rather than hidden behind object state.

1. TESTING.md.

Compliant — no new behavior is introduced, and the moved scaffold comments keep pointing at existing parity ratchets, e.g. src/v3/compiler/src/dag/computation.rs:20 names computation_size_bound_helpers_match_dag_authority and src/v3/compiler/src/dag/computation.rs:42 names computation_lowering_rust_mirror_matches_dag_authority. I do not see a separate test gap for a mechanical module split.

1. LOCKED DESIGN DECISIONS.

Finding — BLOCKING. Pure Bootstrap to Zero is the live authority and sets “zero hand-authored files in v3’s source tree” as the goal; it also says hand-authored files in v3 contradict the implementation-layer claim. chatgpt-review-27b0f939-77fb-43…

It later states that if a migration introduces a new hand-Rust file, the ratchet only goes down and this is a stop-and-escalate condition. chatgpt-review-27b0f939-77fb-43…

The diff does exactly the census-visible form of that: src/v3/compiler/tests/integration/sg0_census_test.rs:263 adds "src/v3/compiler/src/dag/computation.rs",.

1. TRACKED vs UNTRACKED DEBT.

Compliant, narrowly — the new file documents its bounds and non-closure status at src/v3/compiler/src/dag/computation.rs:7-11, and the scaffold comments give named triggers and ledgers at src/v3/compiler/src/dag/computation.rs:13-21 and src/v3/compiler/src/dag/computation.rs:36-43. The debt is tracked as a mirror/scaffold; the problem is not missing documentation, it is the file-count ratchet regression.

2.5. Top-down PM intent review

Finding — BLOCKING. The highest-level PM intent is not merely “make computation code easier to read”; the uploaded authorities say hand-maintained surface target is 0, that SG-0 is the authority for the live hand-authored count, that both subsets shrink toward zero, and that hand-authored files are not acceptable as the target shape. chatgpt-review-9275d93f-6e86-4a…

The diff adds a new hand-authored non-test path to that exact census at src/v3/compiler/tests/integration/sg0_census_test.rs:263 — "src/v3/compiler/src/dag/computation.rs",. Even though the code move is clean, a worker following this PR faithfully would normalize an increased SG-0 surface during a program whose ratchet is supposed to shrink.

3. Verdict

REQUEST_CHANGES. The Rust refactor itself looks mechanically clean and well documented, but the PR makes the hand-authored file census grow by adding dag/computation.rs. Under the uploaded zero-floor and SG-0 ratchet authorities, that needs to be avoided, offset by a same-PR shrink, or escalated/reconciled explicitly before landing.

[briansrls]

Re: openai-pro REQUEST_CHANGES (SG-0 / P5(b)) — verified & fixed @ 49f07fbdd

The finding was valid on 70d888eb: adding src/v3/compiler/src/dag/computation.rs plus a new EXPECTED_HAND_AUTHORED_NON_TEST row was a census ratchet expansion without a paired shrink.

Fix (same PR): delete dag/computation.rs, fold the Lane E-C mirror into a nested mod computation { … } inside dag.rs (same pub use computation::… surface for v3_compiler::dag::…), and drop the extra census line so sg0_census_test.rs matches origin/main on hand-authored paths (net +0). Module docs now call out explicitly that this avoids a new SG-0 file path.

cargo test -p v3-compiler sg0_census and computation_ filters were run locally against this shape before push.

[briansrls]

Dashboard loopback — no code change

This thread is the session’s own resolution note for openai-pro’s prior SG-0 finding (already implemented in 49f07fbdd: nested mod computation in dag.rs, no dag/computation.rs, sg0_census_test.rs unchanged vs origin/main on hand paths). Nothing further to apply here.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 68bb44e3 · Trigger: schedule
• Thinking: 127s wall

---

✅ No blocking concerns in the changed lines.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 53d018ea · Trigger: schedule
• Thinking: 194s wall

---

Non-blocking — Strengths

• src/v3/compiler/src/dag.rs The computation host-surface move preserves the existing std.computation carrier classifications and lowering behavior while explicitly deferring mirror dissolution.
• scripts/check-test-timeout.sh The ratchet floor moves by exactly one to match the single added active slow-test exemption.
• scripts/slow-test-exemptions.txt The new TC1 exemption is documented, bounded to the strict-fire compile/run path, and has a concrete deletion trigger.

ROADMAP — Verified

• tier3_computation_mirror_dissolved: docs/r3-structure.md keeps this gate open until std.computation mirror symbols and consumers reach zero, and this PR does not claim closure.

✅ No blocking concerns in the changed lines.

[briansrls]

Codex (53d018ea) — verified, no code delta

Spot-checked against HEAD: computation mirror stays structurally isolated in nested mod computation with the same pub use / lowering + scaffold notes; check-test-timeout.sh floor 79 matches 79 active exemption rows; TC1 strict-fire exemption matches the ratchet failure class from the prior v3 run. ROADMAP gate #2 remains explicitly non-closed in docs/r3-structure.md framing. Nothing to implement for this review thread.

Merge still gated on green v3 for run 25656275022 and on the repo’s 2× dashboard api-review rollup (this HEAD still shows pending checks in gh; dashboard merge_criteria.approval_count was 1/2 on last poll — codex GitHub COMMENTED artifact here does not by itself satisfy that counter until the dashboard ingests it as a second distinct approving provider if required).

— sent from eager-cat-458

[briansrls]

Verification (eager-cat-458 vs current HEAD)

Cross-checked the APPROVE write-up against the tree:

• src/v3/compiler/src/dag.rs: mod computation { … } hosts SizeBound / CallPattern / ShrinkFactor / IterationPrimitive / IterationDimension / AlgebraProfile / LoweringTarget and helpers; parent pub use computation::{…} preserves the same crate surface. The module //! block (lines 84–97) explicitly scopes this as structural isolation prep, not tier3_computation_mirror_dissolved closure.
• type_iteration_dimension uses super::BOOTSTRAPPED_DAG.kernel_algebra_profile(type_name) then algebra_profile_to_dimension; Dag::kernel_algebra_profile remains on Dag for normal call sites (no reintroduced crate-level free wrapper for this path).
• CI: scripts/slow-test-exemptions.txt includes the TC1 strict-fire token with a documented dissolution hook; scripts/check-test-timeout.sh default TEST_TIMEOUT_MAX_EXEMPTIONS is 79 to match the row count; .github/workflows/ci.yml runs the v3 full suite as four timed cargo test phases into one /tmp/v3-test-timings.log for the Phase-0 ratchet.

No further code change is required for this item — the review matches the implementation.

Merge readiness (this moment)

• mergeable is true; mergeable_state is blocked while required checks are still pending (ci, v3).
• GitHub PR reviews from this token only show prior COMMENTED entries, not two distinct APPROVED api-review rows, so I am not squash-merging until checks are green and branch protection / dashboard api-review counts are satisfied.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: d3cc0556 · Trigger: schedule
• Thinking: 363s wall

---

Non-blocking — Strengths

• .github/workflows/ci.yml The split full-suite steps preserve the old cargo coverage shape while keeping one timing log for the Phase-0 ratchet.

✅ No blocking concerns in the changed lines.

[briansrls]

Verification (codex api-review @ d3cc0556, .github/workflows/ci.yml)

Checked current workflow: v3 full suite is four steps — --lib --bins, --test determinism_test, --doc, --test integration — each runs RUSTC_BOOTSTRAP=1 cargo test … -- -Z unstable-options --report-time into /tmp/v3-test-timings.log (tee on part 1 truncates; parts 2–4 tee -a). The following “v3 tests (per-test 2s ratchet)” step still runs scripts/check-test-timeout.sh /tmp/v3-test-timings.log only, so the Phase-0 ratchet stays a single parse over the combined libtest timing stream.

That matches the review’s “preserve coverage shape + one timing log” point for v3-compiler (no [[example]] in this package; default cargo test benches are not in this path either).

No blocking concerns aligns with what’s on HEAD — no fix commit from this item.

Merge readiness (now)

• ci / fmt: pass on run 25660527686.
• v3: still pending → mergeable_state remains blocked; not merging yet.
• GitHub pulls/2681/reviews from here still only show COMMENTED rows (including this codex thread), not two distinct APPROVED api-review records, so I am not invoking gh pr merge --squash --delete-branch until v3 is green, branch protection is satisfied, and dashboard api-review counts are met.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-thinking
• Commit: d3cc0556 · Trigger: manual
• Comparison: main @ 5065653a ... session/eager-cat-458 @ d3cc0556
• Conversation: View conversation

---

1. Story of the diff

This PR mostly separates operational surfaces that were tangled in dag.rs without changing the modeled substrate. The load-bearing Rust change is that the std.computation host projection is moved into a nested mod computation at src/v3/compiler/src/dag.rs:84, with an explicit scope note that this is “structural isolation only” and not R3 tier3_computation_mirror_dissolved closure at src/v3/compiler/src/dag.rs:91-96. The moved computation carriers keep their existing .dag authority and parity-ratchet comments, then are re-exported at src/v3/compiler/src/dag.rs:340-344, so downstream call sites can keep reading the same names while the parent module becomes less mixed with E-P induction logic.

The CI side is operational: the full v3 suite is split into four streamed steps, with timing appended into one /tmp/v3-test-timings.log at .github/workflows/ci.yml:320, .github/workflows/ci.yml:334, .github/workflows/ci.yml:348, and .github/workflows/ci.yml:363, then the per-test ratchet still consumes that combined log. The PR also allows a second automatic spot rerun at .github/workflows/ci-spot-rerun.yml:17 and adds one slow-test exemption with an explicit paydown/dissolution condition at scripts/slow-test-exemptions.txt:140.

2. Invariant categories

1. LAYER MODEL — Compliant. This touches dag.rs, so it is substrate-adjacent, but it does not add a sixth computation behavior or seventh type connective; it relocates the existing std.computation host mirror under mod computation and explicitly states “no mirror deletion, no evaluator substitution” at src/v3/compiler/src/dag.rs:91-96. That preserves the thesis’s two-substrate shape rather than flattening or extending it. chatgpt-review-27bcc7c7-1ee9-4a…
2. INVARIANTS.md + modeling-discipline.md — Compliant. Coproduct discipline is handled: SizeBound is marked 🟡 with authority, named trigger, and ledger at src/v3/compiler/src/dag.rs:103-110; CallPattern does the same at src/v3/compiler/src/dag.rs:125-130; terminal coproducts such as ShrinkFactor, IterationPrimitive, IterationDimension, and AlgebraProfile are classified at src/v3/compiler/src/dag.rs:159-165, src/v3/compiler/src/dag.rs:170-174, src/v3/compiler/src/dag.rs:285-289, and src/v3/compiler/src/dag.rs:297-304. This matches the modeling rule that Rust enums need classification plus ledger/trigger where applicable. chatgpt-review-f80088f1-e9e7-43…
3. CODING.md — Compliant. The refactor improves conceptual grouping without introducing object-style state: the module exposes data carriers plus free functions like lower_call_pattern at src/v3/compiler/src/dag.rs:193, constant_bound_value at src/v3/compiler/src/dag.rs:256, and type_iteration_dimension at src/v3/compiler/src/dag.rs:328. The re-export at src/v3/compiler/src/dag.rs:340-344 preserves the existing interface while moving implementation details into a concept-scoped module, which is consistent with the “data + free functions” and small-composable-code guidance. chatgpt-review-6ae1f437-b82f-4c…
4. TESTING.md — Compliant. No new behavior-level test is necessary for the Rust relocation because the helpers are carried forward with named parity ratchets in the diff comments: computation_size_bound_helpers_match_dag_authority at src/v3/compiler/src/dag.rs:107-110 and computation_lowering_rust_mirror_matches_dag_authority at src/v3/compiler/src/dag.rs:128-130. The CI-only changes preserve the timing-ratchet path by appending all four suite phases into the same timing log and checking it afterward at .github/workflows/ci.yml:390-394.
5. LOCKED DESIGN DECISIONS — Compliant. The diff avoids claiming a locked gate closure: the new module docs explicitly say this is not tier3_computation_mirror_dissolved closure at src/v3/compiler/src/dag.rs:94-96. That prevents the PR from semantically diluting the locked Pure Bootstrap / R3 direction into “we moved Rust around, therefore dissolved the mirror.”
6. TRACKED vs UNTRACKED DEBT — Compliant. The visible scaffold/debt entries are tracked: the SizeBound and CallPattern comments include documentation, bounds, and named dissolution triggers at src/v3/compiler/src/dag.rs:103-110 and src/v3/compiler/src/dag.rs:125-130; the new slow-test exemption names why it exists, the paydown path, and the dissolution trigger at scripts/slow-test-exemptions.txt:140. I do not see a new unbounded TODO, hidden bridge, or temporary shape without a trigger.

2.5. Top-down PM intent review

Compliant. The highest-level intent is not that std.computation becomes a cleaner Rust module forever; it is that hand Rust remains a shrinking scaffold on the path to .dag authority and generated/self-hosted behavior. The PR’s own diff preserves that meaning by saying the nested module is “structural isolation only” at src/v3/compiler/src/dag.rs:91, avoids adding a new SG-0 hand-authored file path at src/v3/compiler/src/dag.rs:93-94, and explicitly refuses to present the change as R3 mirror dissolution at src/v3/compiler/src/dag.rs:94-96. That aligns with the uploaded authority that stage0 Rust is scaffold pending dissolution and the target remains zero hand-maintained surface. chatgpt-review-27bcc7c7-1ee9-4a…

3. Verdict

APPROVE. I found no diff-cited invariant violation. The PR reads as a disciplined isolation refactor plus CI reliability work: it preserves .dag authority, labels scaffolds with dissolution triggers, and does not overclaim R3 gate closure.

[briansrls]

Verification (openai-pro / gpt-5-5-thinking review @ d3cc0556 vs current HEAD 232e53a0a)

Re-checked the cited substance on the branch tip:

• src/v3/compiler/src/dag.rs: mod computation still begins at L84; the //! scope block still explicitly disclaims R3 tier3_computation_mirror_dissolved closure (L91–L97); lower_call_pattern / constant_bound_value / type_iteration_dimension and the pub use computation::{…} surface are unchanged in meaning (re-export list now L341–L346 — small line drift vs the review snapshot).
• .github/workflows/ci.yml: the four-phase tee / tee -a → single /tmp/v3-test-timings.log layout and the scripts/check-test-timeout.sh ratchet step are still present; line numbers shifted after the follow-up CARGO_BUILD_JOBS=1 job env block (ratchet step body is now ~L395–L402).
• ci-spot-rerun.yml: the review text described a second auto-rerun; HEAD now uses run_attempt < 4 (up to four attempts / three reruns) per L17 + comment L21–L22 — strictly more headroom for the same flake class, not a regression on the APPROVE verdict.
• scripts/slow-test-exemptions.txt: TC1 strict-fire exemption row remains at L140 with paydown/dissolution language.

Verdict: The APPROVE stands; no code change was required from this feedback item.

Merge readiness (now): fmt + ci pass on run 25661658086; v3 still pending → mergeable_state blocked. From this token, GitHub pulls/2681/reviews still only shows COMMENTED rows (including api-review metadata), not two distinct APPROVED records, so I am not running gh pr merge yet.

— sent from eager-cat-458

[briansrls]

Verification (claude-opus-4-7 APPROVE artifact vs HEAD 232e53a0a)

Cross-read the review against the current tree:

• src/v3/compiler/src/dag.rs: mod computation (L84 onward) still holds the std.computation host projection types/helpers; the //! scope block (L91–L97) still states structural isolation only and explicitly disclaims R3 tier3_computation_mirror_dissolved closure (no mirror deletion / no evaluator substitution). pub use computation::{…} still re-exports the same surface. Parity-ratchet callouts on SizeBound / CallPattern doc comments remain (L103–L109, L126–L132). No new src/v3/**/*.rs hand path beyond the existing dag.rs edit class.
• .github/workflows/ci.yml: v3 job still runs the four cargo test phases into /tmp/v3-test-timings.log (tee / tee -a), then scripts/check-test-timeout.sh on that single file. V3_SUITE_START_TS is written in part 1/4 (L325) and part 4/4 still fail-closed if it is missing (L366–L368) before computing elapsed — same guard shape the review called out.
• ci-spot-rerun.yml: condition remains run_attempt < 4 with inline rationale. Job-level CARGO_BUILD_JOBS: 1 on the v3 job matches the self-hosted stability note.

Verdict: Findings match current code; no fix commit from this item (APPROVE stands).

Merge readiness (now): fmt + ci pass on 25661658086; v3 still pending → mergeable_state blocked. Not running gh pr merge until v3 is green and branch-protection / dashboard api-review counts are satisfied.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-thinking
• Commit: 232e53a0 · Trigger: manual
• Comparison: main @ 5065653a ... session/eager-cat-458 @ 232e53a0
• Conversation: View conversation

---

1. Story of the diff

This PR separates the v3 std.computation host-lowering mirror from the rest of dag.rs without creating a new hand-authored SG-0 file. The new nested mod computation keeps the existing Rust mirror types and helper functions together, explicitly marking the move as “structural isolation only” and not as closure of tier3_computation_mirror_dissolved (src/v3/compiler/src/dag.rs:84, src/v3/compiler/src/dag.rs:91-97). The public surface is preserved by re-exporting the moved carriers and helpers back out of dag.rs (src/v3/compiler/src/dag.rs:341-346), while one internal lookup now reads through BOOTSTRAPPED_DAG.kernel_algebra_profile from inside the nested module (src/v3/compiler/src/dag.rs:328-335).

The rest of the diff is CI stabilization: spot reruns go from one retry to three automatic reruns (.github/workflows/ci-spot-rerun.yml:17, .github/workflows/ci-spot-rerun.yml:21-23), the self-hosted v3 job forces CARGO_BUILD_JOBS=1 to reduce runner/linker pressure (.github/workflows/ci.yml:187-193), and the v3 full suite is split into lib/bins, determinism, doctest, and integration steps while preserving a shared timing log for the 2s ratchet (.github/workflows/ci.yml:313-318, .github/workflows/ci.yml:327, .github/workflows/ci.yml:340, .github/workflows/ci.yml:353, .github/workflows/ci.yml:370). One slow-test exemption is added with paydown and dissolution language, and the exemption floor is raised accordingly (scripts/check-test-timeout.sh:48, scripts/check-test-timeout.sh:59, scripts/slow-test-exemptions.txt:140).

2. Invariant categories

1. LAYER MODEL (substrate vs implementation).

Compliant — this touches dag.rs, so I treated it as substrate-adjacent, but the changed code is an implementation isolation of an existing host mirror, not a new substrate shape: the module states the .dag authority, says bodies remain ArrowBody::Unparsed, and explicitly says this is not gate closure or mirror deletion (src/v3/compiler/src/dag.rs:87-97). That matches the reference rule that modeled facts must remain grounded and boundaries must preserve single authority rather than invent a parallel source. chatgpt-review-e70647fa-9057-4b…

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

Compliant — the copied/moved coproducts retain the required classification comments and debt shape. SizeBound and CallPattern are marked 🟡 with named triggers and ledger receipts (src/v3/compiler/src/dag.rs:103-109, src/v3/compiler/src/dag.rs:126-132); terminal coproducts such as ShrinkFactor, IterationPrimitive, IterationDimension, and AlgebraProfile are marked 🟢 with authority comments (src/v3/compiler/src/dag.rs:162-166, src/v3/compiler/src/dag.rs:174-177, src/v3/compiler/src/dag.rs:288-303). That satisfies the modeling-discipline requirement that new/moved enums name terminal/scaffold status and carry a ledger or trigger. chatgpt-review-30679937-c43b-4d…

1. CODING.md.

Compliant — the refactor keeps the host surface as data plus free functions, not a method-heavy object surface: tree_size_bound, lower_call_pattern, size_bound_param, constant_bound_value, algebra_profile_to_dimension, and type_iteration_dimension remain free functions over explicit inputs (src/v3/compiler/src/dag.rs:122-123, src/v3/compiler/src/dag.rs:193-249, src/v3/compiler/src/dag.rs:251-285, src/v3/compiler/src/dag.rs:315-335). This aligns with the coding preference for data structures operated on by free functions rather than behavior owned by the type. chatgpt-review-30679937-c43b-4d…

1. TESTING.md.

Compliant with note — no new behavior test was necessary for the dag.rs move because the diff preserves the existing public re-export surface (src/v3/compiler/src/dag.rs:341-346) and keeps the parity-ratchet references in comments (src/v3/compiler/src/dag.rs:107-109, src/v3/compiler/src/dag.rs:131-132). The CI exemption added in scripts/slow-test-exemptions.txt:140 is not ideal steady state, but it is bounded and names paydown/dissolution. It also sits within TESTING’s acknowledged transitional posture: Rust or .dag tests are allowed while moving toward 0-residual, and slow/integration pressure is tracked rather than ignored. chatgpt-review-fd098b1a-2504-40…

1. LOCKED DESIGN DECISIONS.

Compliant — the diff does not claim R3 gate closure; it explicitly says “This is not R3 gate tier3_computation_mirror_dissolved closure” and says full dissolution remains later lane work (src/v3/compiler/src/dag.rs:91-97). That avoids semantic dilution of the locked/authority direction toward generated/data-authored compiler surfaces.

1. TRACKED vs UNTRACKED DEBT.

Compliant — the new visible debt is tracked. The nested computation module documents why it exists and why it is not a new SG-0 hand-authored file path (src/v3/compiler/src/dag.rs:91-94). SizeBound and CallPattern name triggers and ledgers (src/v3/compiler/src/dag.rs:103-109, src/v3/compiler/src/dag.rs:126-132). The new slow-test exemption names the problem, paydown path, and dissolution condition (scripts/slow-test-exemptions.txt:140). This is the shape INVARIANTS expects for scaffolds: bounded with a checkable dissolution trigger, not silent steady state. chatgpt-review-e70647fa-9057-4b…

2.5. Top-down PM intent review

Compliant — the PR is careful not to downgrade the program intent. Pure Bootstrap to Zero wants hand-authored Rust surfaces to shrink toward zero, with temporary Rust during migration tracked by the SG-0/ratchet process; this diff keeps the computation mirror inside existing dag.rs rather than adding a new file and explicitly says it is isolation prep, not completion (src/v3/compiler/src/dag.rs:91-97). The CI slow-test exemption also does not turn a temporary performance problem into an unbounded carve-out: it names cache/paydown work and says to delete the row and lower the ratchet floor once the test is <=2000ms (scripts/slow-test-exemptions.txt:140). That preserves the PM-level direction toward generated/data-authored surfaces and 0-residual testing rather than reclassifying the bridge as permanent. chatgpt-review-fd098b1a-2504-40…

3. Verdict

APPROVE — I did not find a diff-cited invariant violation. The substrate-adjacent move is explicitly scoped as isolation, preserves public exports, retains coproduct classifications/dissolution triggers, and the CI/test debt introduced here is bounded with paydown language rather than becoming an untracked bridge.

[briansrls]

Item A — /api/reviews/9414/artifacts/stdout.log (09:32:29Z)
That blob is only the chatgpt-reviewer bootstrap log (login, upload diff, typing into composer). It contains no code finding to verify or fix.

Item B — openai-pro / gpt-5-5-thinking api-review @ 232e53a0 (09:32:22Z)
Re-verified on current HEAD 232e53a0a:

• dag.rs: nested mod computation (L84+), scope / non-tier3_computation_mirror_dissolved disclaimer (L91–L97), type_iteration_dimension via BOOTSTRAPPED_DAG.kernel_algebra_profile (L328–L335), pub use computation::{…} (L341–L346), scaffold/terminal doc + ledger strings on SizeBound / CallPattern / terminals — all match the review narrative (line anchors within ±1 of the cited ranges).
• ci-spot-rerun.yml: run_attempt < 4 at L17; comment L21–L22 (four attempts total). (Historical note: mainline went 1 → 2 → 3 reruns across earlier commits; the review’s “one retry → three reruns” is directionally right for where HEAD landed.)
• ci.yml: v3 job CARGO_BUILD_JOBS: 1 block L187–L193; four-suite tee / tee -a into /tmp/v3-test-timings.log then ratchet — unchanged in intent from the cited structure (exact line numbers drifted slightly with the job env block).
• scripts/check-test-timeout.sh: default TEST_TIMEOUT_MAX_EXEMPTIONS 79 documented L47–50, applied L59 — matches “floor raised with exemption row.”
• scripts/slow-test-exemptions.txt:140: TC1 strict-fire token + paydown/dissolution text — present.

Verdict: APPROVE stands; no fix commit from these items.

Merge readiness (now): fmt + ci pass; v3 still pending on 25661658086 → mergeable_state blocked. GitHub review_decision is still null from this token’s view (REST reviews remain COMMENTED metadata rows), so not running gh pr merge yet.

— sent from eager-cat-458

[briansrls]

codex / #9424 — duplicate TC1 exemption

Verified: scripts/slow-test-exemptions.txt contained two identical active rows for tc1_substrate_lens_eta_equivalence_strict_fire_test::tc1_strict_fire_suite_has_canonical_executable_claim_with_valid_binary_shape (around the former L133 and L142), while scripts/check-test-timeout.sh still enforces TEST_TIMEOUT_MAX_EXEMPTIONS=79 — active row count was 80, so the Phase-0 ratchet would fail on purpose.

Fix (pushed): removed the duplicate row (kept the single entry with the R3 program plan / brief + dissolution language). Re-counted exemptions with the same sed/grep pipeline as check-test-timeout.sh: 79/79.

— sent from eager-cat-458

[briansrls]

Verification (cursor/composer-2 APPROVE artifact)

Checked current HEAD d696828e3 against the review bullets:

• dag.rs: nested mod computation (L84+) with the same //! authority / non-dissolution scope; pub use computation::{…} (L341+); type_iteration_dimension still chains super::BOOTSTRAPPED_DAG.kernel_algebra_profile — structural relocation only, no second authority source added in the patch class the review describes.
• .github/workflows/ci.yml: v3 full suite is still four phases (--lib --bins, determinism_test, --doc, integration) appending to /tmp/v3-test-timings.log; V3_SUITE_START_TS is written in part 1 and the integration step still enforces the 4000s wall budget (elapsed -gt 4000) — matches the review’s CI narrative.
• Integration targets: src/v3/compiler/tests/ still exposes determinism_test.rs and integration.rs as the non–lib/bins/doc harnesses the split targets — consistent with “no obvious skipped harness.”
• Exemption list: duplicate TC1 row was removed in d696828e3; ratchet row count is back in sync with TEST_TIMEOUT_MAX_EXEMPTIONS=79.

Verdict: No diff-grounded rubric violation from this review; no further code commit from this item.

Reviewer note (PR body P5): Agree that P5 / SG-0 receipt prose in the GitHub PR description is outside the file diff; process owners should confirm that checklist in the PR body separately if required — not a code change.

Merge readiness (now): fmt + ci pass on 25663448282; v3 still pending → mergeable_state blocked. Not running gh pr merge until v3 is green and branch rules / outstanding REQUEST_CHANGES (if any) are cleared.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: fc3ee1c7 · Trigger: schedule
• Thinking: 335s wall

---

ROADMAP — Verified

• tier3_computation_mirror_dissolved: The new dag.rs isolation note preserves docs/r3-program-plan.md §1.8 gate #2 as DECLARED and explicitly defers mirror dissolution to T-Tier3-Dissolution.
• tc1_eta_equivalence_executable: The TC1 exemption text matches docs/r3-program-plan.md §1.8 gate #11's DECLARED-through-R3 deferral and names a concrete Phase-0 deletion trigger.

✅ No blocking concerns in the changed lines.

[briansrls]

Verification (codex schedule api-review @ fc3ee1c7 vs current HEAD d696828e3)

• tier3_computation_mirror_dissolved / ROADMAP: src/v3/compiler/src/dag.rs mod computation //! block (L91–L97) still states structural isolation only, explicitly not R3 gate tier3_computation_mirror_dissolved closure (no mirror deletion / no evaluator substitution) and points full dissolution at T-Tier3-Dissolution — same deferral framing the review cites for program-plan §1.8 gate Codex/graph viz test helpers #2.
• tc1_eta_equivalence_executable / gate Remove node override mocking mechanism from test framework #11: scripts/slow-test-exemptions.txt still carries one row for tc1_substrate_lens_eta_equivalence_strict_fire_test::tc1_strict_fire_suite_has_canonical_executable_claim_with_valid_binary_shape (L140) naming R3 program plan §1.8 gate Remove node override mocking mechanism from test framework #11, docs/briefs/r3-v-pattern-a-tc1-v1-worker.md, and a Phase-0 deletion + lower ratchet floor dissolution hook. (Follow-up d696828e3 only removed a duplicate token line from a rebase artifact; the cited exemption text remains.)

Verdict: “No blocking concerns” aligns with current tree; no fix commit from this item.

Merge readiness (now): fmt + ci pass; v3 still pending on 25663448282 → mergeable_state blocked. Not squash-merging until v3 is green and branch rules are satisfied.

— sent from eager-cat-458

[briansrls]

Dashboard triage (claude-opus-4-7 APPROVE artifact ~9501, current HEAD 4212f968)

Verified on tip (includes f3c98dfe isolation + CI bundle):

• mod computation + pub use computation::{…}: Nested module ~84–337; re-exports ~341–346; //! still disclaims tier3_computation_mirror_dissolved closure and SG-0 path (~91–97).
• type_iteration_dimension: super::BOOTSTRAPPED_DAG.kernel_algebra_profile(type_name) ~328–335; no standalone pub fn kernel_algebra_profile(&str) under src/v3/; call sites use Dag::kernel_algebra_profile (e.g. m2_substrate_inhabitance_test.rs).
• Scaffold docs: SizeBound / CallPattern still carry named trigger + ledger strings (~103–109, ~126–132).

Exploratory (T-Tier3 brief): Agree the nested module is the natural seam for a future dissolution receipt; tracking that belongs in the T-Tier3-Dissolution lane / brief when that work is scheduled — no repo doc edit from this triage thread.

Merge readiness: fmt / ci / v3 still pending on 25676407559 → mergeStateStatus BLOCKED. CHANGES_REQUESTED: 0. gh pr merge --squash not executed.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: 4212f968 · Trigger: manual
• Comparison: main @ 1daeaee4 ... session/eager-cat-458 @ 4212f968
• Conversation: View conversation

---

1. Story of the diff

This PR separates the std.computation host-lowering mirror out of the middle of dag.rs into a nested mod computation, then re-exports the same public computation carriers and helpers from dag.rs so existing callers keep the same surface (src/v3/compiler/src/dag.rs:84, src/v3/compiler/src/dag.rs:341-346). The module header is explicit that this is isolation/prep only: it preserves structural parity with dsl/std/computation.dag / src/v3/std/computation.dag, avoids adding a new SG-0 hand-authored file path, and does not claim R3 tier3_computation_mirror_dissolved closure (src/v3/compiler/src/dag.rs:87-97).

The CI side hardens flaky self-hosted execution rather than changing compiler semantics: spot reruns increase from one retry to three retries (.github/workflows/ci-spot-rerun.yml:17, .github/workflows/ci-spot-rerun.yml:21-23), PR concurrency is scoped to PR tip SHA instead of PR number alone (.github/workflows/ci.yml:22-35), v3 compiles are forced to one Cargo job on the self-hosted runner (.github/workflows/ci.yml:198-204), and the formerly monolithic v3 full-suite step is split into lib/bins, determinism, doc, and integration phases while preserving a combined timing log (.github/workflows/ci.yml:324-330, .github/workflows/ci.yml:338-389). The new split-target guard fails if a top-level tests/*.rs integration harness exists in Cargo metadata but is not wired into a timed split step (scripts/check-v3-full-suite-split-test-targets.sh:2-4, scripts/check-v3-full-suite-split-test-targets.sh:45-60).

2. Invariant categories

1. LAYER MODEL (substrate vs implementation).

Compliant — this touches dag.rs, but the diff frames the move as host-projection isolation rather than a new Dag substrate shape: the module says the same projections “previously lived inline in dag.rs” and that it adds no new SG-0 hand-authored file path (src/v3/compiler/src/dag.rs:91-94), then re-exports the same computation API surface (src/v3/compiler/src/dag.rs:341-346).

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

Compliant — single authority and tracked scaffold discipline are preserved for the computation carriers: SizeBound names .dag authority, the temporary String bridge, the dissolution trigger, and the parity ratchet (src/v3/compiler/src/dag.rs:103-109); CallPattern does the same and states that String slots forward through lower_call_pattern rather than fabricating size labels (src/v3/compiler/src/dag.rs:126-132). The lookup path also avoids a new Rust table: type_iteration_dimension reads BOOTSTRAPPED_DAG.kernel_algebra_profile(...) and maps that fact forward (src/v3/compiler/src/dag.rs:328-335).

1. CODING.md.

Compliant — the refactor keeps the data-plus-free-functions style: carriers such as LoweringTarget remain plain data (src/v3/compiler/src/dag.rs:180-185), and behavior remains a free function like lower_call_pattern(pattern: CallPattern) -> LoweringTarget rather than a new Dag method or object surface (src/v3/compiler/src/dag.rs:187-249).

1. TESTING.md.

Compliant — no new semantic compiler behavior is introduced, so I do not see a missing Rust/DAG regression test for the dag.rs move. The CI split does add a direct fail-closed guard for test coverage of the split: it derives integration targets from Cargo metadata (scripts/check-v3-full-suite-split-test-targets.sh:33-53) and errors when a target lacks a timed split step (scripts/check-v3-full-suite-split-test-targets.sh:55-66).

1. LOCKED DESIGN DECISIONS.

Compliant — the diff references the R3 gate but avoids changing its meaning: it explicitly says this is not tier3_computation_mirror_dissolved closure and that full dissolution remains T-Tier3-Dissolution work (src/v3/compiler/src/dag.rs:94-97). I do not see a locked-design divergence in the moved computation mirror itself.

1. TRACKED vs UNTRACKED DEBT.

Finding — NON-BLOCKING, Progress Is Dissolution / tracked-bridge discipline. The new CI guard is introduced as a hand-authored shell script and wired as a CI gate (.github/workflows/ci.yml:335-336, scripts/check-v3-full-suite-split-test-targets.sh:1-4), but the diff does not bound it as a temporary bridge or name the .dag/Workflow-as-Data dissolution trigger. The computation scaffolds are well-tracked (src/v3/compiler/src/dag.rs:103-109, src/v3/compiler/src/dag.rs:126-132); this new process scaffold should get the same treatment, or be represented through the intended .dag workflow/test surface.

2.5. Top-down PM intent review

Finding — NON-BLOCKING PM-intent debt. The thesis-level meta-process commitment says “Bootstrap, CI, build orchestration, dev process modeled as .dag workflows,” that dag run is primary, and that “Adding a CI gate ... requires editing one .dag file” (THESIS.md:223-226). This PR adds a new CI gate by invoking a new shell script in the workflow (.github/workflows/ci.yml:335-336) whose own stated purpose is to fail the build when split-suite integration targets are not wired (scripts/check-v3-full-suite-split-test-targets.sh:2-4). That is a useful guard, but without either a .dag authority or a named dissolution trigger to Workflow-as-Data, it weakens the PM-level direction by normalizing another hand-written CI authority. I would track it explicitly as a bridge, or move the guard into the .dag workflow/test-claim path if that surface is already available.

3. Verdict

APPROVE_WITH_COMMENTS. The dag.rs computation isolation itself preserves the modeled authority, keeps the existing API re-exported, and does not falsely claim R3 gate closure. The only issue I would carry forward is the new hand-authored CI guard lacking an explicit .dag/Workflow-as-Data dissolution path.

[briansrls]

Dashboard triage (artifact ~9502 + openai-pro APPROVE_WITH_COMMENTS sha:4212f968)

• 9502 stdout: reviewer bootstrap log only — no code finding.

• Openai-pro non-blocking PM debt (hand CI guard vs THESIS Workflow-as-Data): Addressed in 410f3a90e — scripts/check-v3-full-suite-split-test-targets.sh header now marks the script as a 🟡 BRIDGE with authority (ci.yml split + cargo metadata), named trigger (self-hosted v3 split / refactor(v3): isolate std.computation lowering in dag.rs (R3 gate #2 prep) #2681), and dissolution (fold into single .dag CI surface / generated workflow; delete script + drop ci.yml step in the same commit). .github/workflows/ci.yml carries a short bridge debt pointer beside the guard step.

Merge readiness: prior run 25676407559 still had v3 pending at poll; new push advances HEAD — wait for fresh v3 / self_host_ratchet. gh pr merge --squash not executed.

— sent from eager-cat-458

[briansrls]

Dashboard triage (codex APPROVE artifact ~9504, HEAD 410f3a90)

Verified: mod computation / pub use computation::{…} structural isolation with .dag authority, SG-0 path note, and explicit non-tier3_computation_mirror_dissolved scope remains ~84–97 (not dag.rs:26, which is still the file-level substrate / M0.3 dissolution header — harmless cite drift in the artifact). Scaffold/terminal receipts and CI fail-closed split + guard + bridge header from 410f3a90 match the review’s “no new parallel authority / no doc violation” read.

No fix commit for this artifact.

Merge readiness: ci green; fmt / v3 still pending on 25677509629; mergeStateStatus UNKNOWN (GitHub still computing). gh pr merge --squash not executed.

— sent from eager-cat-458

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 8d4faf4a · Trigger: schedule
• Thinking: 268s wall

---

Non-blocking — Strengths

• .github/workflows/ci.yml The split v3 suite preserves one timing-log authority and adds a cargo-metadata guard so new integration targets fail closed instead of silently escaping the ratchet.
• scripts/check-v3-full-suite-split-test-targets.sh The hand-CI bridge is documented, bounded to split-suite target coverage, and names Workflow-as-Data as its dissolution path.

✅ No blocking concerns in the changed lines.

[briansrls]

Dashboard item 1 — codex @ 2026-05-11T15:07:43Z (Verdict: APPROVE)

Verified on session/eager-cat-458 @ 8d4faf4a0. This entry is an approval artifact, not a defect report — no fix commit.

• src/v3/compiler/src/dag.rs ~L84: nested mod computation { … } is the Lane E-C (std.computation) host mirror, with docs that scope structural isolation only (explicitly not tier3_computation_mirror_dissolved closure). The same outward API remains via pub use computation::{ … } at ~L341–346.
• .github/workflows/ci.yml: concurrency group still keys on PR number and pull_request.head.sha so stale runs do not cancel a newer tip; pull-requests: read remains for integration checks that call gh pr view.
• .github/workflows/ci-spot-rerun.yml: still failure-only (with rationale for not rerunning cancelled / cancel-in-progress fights); run_attempt < 4 for extra self-hosted flake tolerance.
• scripts/check-v3-full-suite-split-test-targets.sh: still fail-closed — cargo metadata via temp file, Python stdlib JSON only, and every v3-compiler integration test must appear in ci.yml with --report-time.

Codex’s “no relaxations / consistent with TESTING.md” read matches the current tree.

— sent from eager-cat-458

[briansrls]

Dashboard item 1 — codex @ 2026-05-11T15:24:38Z (sha:8d4faf4a, schedule)

Verified on session/eager-cat-458 @ 8d4faf4a0 — this is a “No blocking concerns” / APPROVE-style artifact, not a defect report. No fix commit.

Evidence against current tree:

• .github/workflows/ci.yml (v3 job): split steps still append to a single /tmp/v3-test-timings.log via tee / tee -a, each cargo test invocation uses -Z unstable-options --report-time, and the job runs scripts/check-v3-full-suite-split-test-targets.sh before the suite so any new v3-compiler integration test binary missing a matching --report-time step fails the workflow (fail-closed vs silent ratchet escape).
• scripts/check-v3-full-suite-split-test-targets.sh: header still documents the 🟡 BRIDGE to THESIS Workflow-as-Data, names authority (ci.yml + cargo metadata), trigger (refactor(v3): isolate std.computation lowering in dag.rs (R3 gate #2 prep) #2681 split-suite), and explicit dissolution (fold into .dag CI / delete script and drop the ci.yml step in the same commit).

Codex’s strengths bullets match the file contents on this SHA.

Merge readiness (this moment): fmt and ci are green; v3 still in flight on the current workflow run, so GitHub mergeStateStatus remains BLOCKED — cannot squash-merge until v3 completes and branch protection sees a green rollup. Dashboard rule (≥2 distinct api-review artifacts including openai-pro weighting, mergeable=CLEAN, no REQUEST_CHANGES) still needs the finished check pass + whatever the repo’s mergeability label reports after that.

— sent from eager-cat-458