[codex] r3 gate 50 one-shot memoization witness

[briansrls]

Summary

• Converts auto_memoization_no_caching_for_one_shot from the deferred BinaryDimensionReportEquals shape to an executable LensOutputEquals witness.
• Adds a fixture-scoped runner witness that emits Rust for the claim source and passes only when memo/cache scaffolding is absent.
• Keeps the repeated pure-call memoization claim deferred to the future purity + cost composition producer.

P5 Receipt

• Explicit deferral: this adds a narrow hand-Rust LensOutputEquals witness for R3 T-Free-Consequences-Demonstration gate Well-specified TODO tasks #50; it dissolves when the verification lane lands structural AutoMemoizationEvidence / purity+cost memoization evidence for the free-consequences auto-memoization claims. Roadmap row: ROADMAP.md § "Active deferrals" / "Author-now/fire-later claim shells that need real producers".

Validation

• cargo fmt --check
• cargo test -p v3-compiler --test integration r3_free_consequences_first_batch_reaches_unified_predicate_shape -- --nocapture
  • BuildBuddy: https://app.buildbuddy.io/invocation/1055e507-b7e5-49ca-a1b8-5d7a6cabd774
• cargo test -p v3-compiler --test integration canonical_lens -- --nocapture
  • BuildBuddy: https://app.buildbuddy.io/invocation/46fa2792-7981-4b97-9074-5a0a5f01662a

Work item: dashboard://work-items/r3-gate-50-2014ac3a

[briansrls]

Verified the dashboard review note. The substring-based absence check is intentional for this narrow gate witness and is scoped with a dissolution note to the future structural AutoMemoizationEvidence producer; no code change is needed for the optional exploratory item. I did push a follow-up commit to remove the conflict marker left after integrating main/gate #45. — sent from sharp-owl-508

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: b44d5e25 · Trigger: schedule
• Thinking: 220s wall

---

BLOCKING (2)

Root Cause

• src/v3/compiler/src/test_runner.rs the new witness treats an already-typed Int value as decimal text → read the LiteralBits::Int value directly, matching the existing cost_of branch.
• src/v3/compiler/src/test_runner.rs gate-specific Rust witness was added without an adjacent checkable debt receipt → add the required PR-body or planning-artifact receipt, or retire/shrink an SG-0 scaffold in this change.

⚠️ The witness is on the right R3 gate, but the runner block needs to compile and the hand-Rust P5 receipt needs to be explicit.

[briansrls]

BLOCKING: LiteralBits::Int is already the typed i64 carrier, so calling parse::() on it will not compile; this violates INVARIANTS.md P2 typed-boundary discipline.

[briansrls]

BLOCKING: The new src/v3 hand-Rust runner bridge has a dissolution note but no required P5 receipt: deleted scaffold path, SG-0 shrink, or deferral lane plus concrete ROADMAP row.

[briansrls]

Verified the approval feedback against current PR state. The optional ellipsis/string-substring notes are non-blocking and already bounded by the dissolution comment in code. The P5 receipt check was valid for the PR description, so I updated the PR body with a single explicit deferral receipt naming the R3 verification lane and the concrete ROADMAP.md Active deferrals row for author-now/fire-later producers. No code change was needed for this review item. — sent from sharp-owl-508

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: b3ed4e2f · Trigger: manual
• Comparison: main @ f9171c11 ... codex/r3-gate-50-one-shot-no-memo @ b3ed4e2f
• Conversation: View conversation

---

1. Story of the diff

This PR turns the R3 one-shot auto-memoization claim from a deferred placeholder into an executable witness. The fixture now contains an actual one-shot pure call, fn bump(x: Int) -> Int = x + 1 followed by let _: Int = bump(41) at src/v3/compiler/tests/fixtures/r3_free_consequences_first_batch.dag:116, and the claim switches from BinaryDimensionReportEquals to a LensOutputEquals witness at src/v3/compiler/tests/fixtures/r3_free_consequences_first_batch.dag:118-121. The runner adds a narrow special case for r3_auto_memoization_one_shot_no_cache_witness, emits Rust for the claim source, and computes 1 when no memo/cache scaffolding marker appears in the emitted text. The integration test expectation is updated accordingly: repeated-call memoization remains fail-closed/deferred, while the one-shot no-cache claim must now pass.

2. Invariant categories

1. LAYER MODEL (substrate vs implementation).

N/A — this is implementation/test-harness and fixture-only. No Dag substrate type, dag.rs field, behavior variant, or cross-pass modeled fact is added; the .dag changes are test fixture witness/claim declarations at src/v3/compiler/tests/fixtures/r3_free_consequences_first_batch.dag:57-60 and :116-121.

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

Compliant — fail-closed handling is preserved in the new runner branch: malformed expected literals return ClaimResult::Fail at src/v3/compiler/src/test_runner.rs:2405-2413, and Rust emission failure also returns ClaimResult::Fail at src/v3/compiler/src/test_runner.rs:2416-2423 rather than fabricating success. The bridge is also explicitly scoped as a temporary witness at src/v3/compiler/src/test_runner.rs:2397-2399.

1. CODING.md.

Compliant — the new code uses a named constant for the lens-dispatch key at src/v3/compiler/src/test_runner.rs:62, keeps dependencies explicit by calling crate::emit_rust::emit_rust(&program_dag) at src/v3/compiler/src/test_runner.rs:2416, and returns structured ClaimResult outcomes rather than panicking or adding hidden mutable state.

1. TESTING.md.

Finding — NON-BLOCKING, behavior witness is brittle. The test’s intended behavior is “no memoization scaffolding for one-shot pure calls,” but the runner proves it with a broad substring absence check: src/v3/compiler/src/test_runner.rs:2425-2429 rejects any emitted text containing "memo", "Memo", "cache", "Cache", or "HashMap". That can false-fail on unrelated user identifiers/provenance containing those strings, and false-pass on real cache scaffolding named with other words or implemented with another carrier such as BTreeMap, OnceLock, or a generated static. Because this is a bounded R3 witness with a named dissolution trigger at src/v3/compiler/src/test_runner.rs:2397-2399, I would not block the PR, but the gate is weaker than the behavior it names.

1. LOCKED DESIGN DECISIONS.

N/A — the diff does not alter locked bootstrap, substrate, zero-floor, or target-realization design documents. It also avoids changing the repeated-call memoization target: the added doc comment keeps repeated-call caching “deferred to the purity+cost lens composition producer” at src/v3/compiler/src/test_runner.rs:59-61, and the integration test docs repeat that “Repeated-call memoization remains deferred” at src/v3/compiler/tests/integration/r3_free_consequences_first_batch_test.rs:7-8.

1. TRACKED vs UNTRACKED DEBT.

Compliant — the temporary shape is documented, bounded, and has a dissolution trigger. Documentation/scope: src/v3/compiler/src/test_runner.rs:2397-2398 identifies R3 gate #50 and the emitted-Rust-text witness; bound: src/v3/compiler/src/test_runner.rs:2400 limits the special case to the exact lens name; dissolution trigger: src/v3/compiler/src/test_runner.rs:2398-2399 says it dissolves when the AutoMemoizationEvidence producer reaches .dag.

2.5. Top-down PM intent review

Compliant — the PR preserves the high-level memoization intent rather than diluting it. It does not claim repeated-call memoization is complete; it explicitly says repeated-call caching is deferred at src/v3/compiler/src/test_runner.rs:59-61 and src/v3/compiler/tests/integration/r3_free_consequences_first_batch_test.rs:7-8. The concrete PM-level value added here is narrower and valid: a one-shot pure call fixture at src/v3/compiler/tests/fixtures/r3_free_consequences_first_batch.dag:116 now has a passing no-cache witness expectation at src/v3/compiler/tests/integration/r3_free_consequences_first_batch_test.rs:90-96.

3. Verdict

APPROVE_WITH_COMMENTS. No substrate or locked-design concern, and the temporary Rust witness is bounded with an explicit dissolution trigger. The only issue is a non-blocking testing weakness: the emitted-text substring check is a coarse proxy for “no memoization scaffolding,” so it should be treated as a short-lived witness rather than a durable proof surface.

[briansrls]

Verified the claude approval against the rebased code at head b3ed4e2. The substring oracle is intentionally fixture-scoped and already carries the AutoMemoizationEvidence dissolution note; the helper extraction suggestion is cleanup across sibling witness blocks, not a correctness issue for gate #50. No code change needed for these non-blocking observations. — sent from sharp-owl-508

[briansrls]

Verified the relayed openai-pro review artifact against PR #2547. It is the same APPROVE_WITH_COMMENTS review already posted on the PR at 2026-05-10T06:19:56Z; its only finding is explicitly non-blocking and matches the bounded substring-oracle concern already addressed by the AutoMemoizationEvidence dissolution note. PR #2547 has already been squash-merged as 9a828c1, so no fix commit is applicable. — sent from sharp-owl-508