Substrate T-E-P P1 Slice 4: multi-arg composition cementing tests

[briansrls]

Summary

T-E-P-Producer-Broadening Phase 1, Slice 4 of N. Pure cementing slice — no producer extension. Pins the per-arg evidence vector contract that Slices 1-3 (#2167, #2178, #2182) built but only single-arg tests exercised. The contract aligns with v2's ExprCall.descent_evidence: List<SubValueRelation>? shape (src/v2/00_core.dag:199) — one relation per argument port.

Two new behavioral tests

1. emits_per_arg_relation_for_multi_arg_self_call

Three-arg recursion over EpListA with mixed arg shapes:

fn ep_count_acc(xs: EpListA, acc: Int, limit: Int) -> Int =
  match xs {
    EpConsA(tail) => ep_count_acc(tail, acc + 1, limit),
    EpNilA => acc
  }

• arg 0 (tail): match-payload positional → StrictSubValue (Slice 1 classifier)
• arg 1 (acc + 1): non-descending arithmetic → SubValueUnknown (sound fail-closed: accumulator grows, can't be sub-value of itself)
• arg 2 (limit): unchanged forwarded param → PreservedValue

2. emits_per_arg_arithmetic_descent_on_non_first_param

Two-arg recursion exercising param_1 arithmetic descent (vs the conventional param_0-only recursion):

fn ep_two_param_descent(x: Int, n: Int) -> Int =
  if n == 0 then x else ep_two_param_descent(x, n - 1)

Pins per-arg classifier independence: arg N is classified against caller.params[N], not always against params[0]. ArithmeticDescent's ordinal label tracks the per-arg index (param_1).

Cross-slice invariants reaffirmed

• ✓ No producer code changes — pure cementing slice
• ✓ No new CallPattern variant
• ✓ No TransformNode widening
• ✓ No INVARIANTS.md P1 substrate-fact-introduction trigger
• ✓ No termination-prover changes

Discipline value

Future producer broadening MUST preserve per-arg classification independence. This slice cements the contract with two behavioral pins so a regression that drops any of the three classifiers (StrictSubValue / ArithmeticDescent / PreservedValue) on multi-arg shapes fires immediately.

Gate progress

• e_p_per_call_descent_evidence_full_coverage (gate Lane B tasks #76, P1): partial — Slice 4 cements. No additional shapes classified, but the per-arg contract (one relation per port) is now behaviorally pinned.
• Other gates: unchanged.

Tests

• 8 e_p_per_call_descent_evidence_* tests now (Slices 1-3's 7 + 2 new − 1 already-existing pattern overlap = 8 passing). Previously: 7 green on Slice 3.
• All Slices 1-3 tests preserved.
• Local CI shim is unable to reach the remote runner from this session; relying on PR CI to verify.

Test plan

• Multi-arg fixture exercises StrictSubValue + Unknown + Preserved in one call
• Non-first-param arithmetic descent exercises per-arg parameter independence
• Slice 1-3 tests preserved
• CI green (verifying via PR pipeline)

Authority

• Brief: docs/briefs/r3-t-e-p-producer-broadening-worker.md (PR R3 Substrate #1782, MERGED)
• Issue: Substrate T-E-P P1 re-dispatch: nested-match scrutinee + termination (post-#1865 abandonment) #2165
• Slices 1-3 (predecessors): Substrate T-E-P P1 first-slice broadening: match-payload self-call → StrictSubValue #2167 / Substrate T-E-P P1 Slice 2: match-field-projection self-call → StrictSubValue #2178 / Substrate T-E-P P1 Slice 3: nested-match scrutinee + termination #2182 (all MERGED)
• Substrate Mgr planned-slices table cite: gunbc#2167 ("Slice 4: Multi-arg composition")

🤖 Generated with Claude Code

[briansrls]

Review metadata

• Provider / model: claude / claude-opus-4-7
• Commit: e9e80ba0 · Trigger: schedule
• Comparison: origin/main @ 4ac0cfd3 ... review/pr-2192-e9e80ba0 @ e9e80ba0
• Thinking: 11s wall

---

APPROVE — Two new behavior-driven tests cementing the per-arg SubValueRelation contract for multi-arg self-calls. Hermetic (compile-to-dag fixtures), tightly scoped, and the assertions match v2's documented descent_evidence: List<SubValueRelation>? shape. No invariants, modeling, or coding-style concerns; pure test additions with no production code touched.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: e9e80ba0 · Trigger: schedule
• Thinking: 166s wall

---

Non-blocking — Strengths

• src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs Rust test-only PR; the added fixtures cleanly ratchet the E-P per-argument SubValueRelation contract without adding substrate state.

ROADMAP — Verified

• T-E-P-Producer-Broadening: The tests advance the documented producer-broadening lane by covering multi-arg evidence vectors and non-first-parameter arithmetic descent.

✅ No blocking concerns.

[briansrls]

CI fix pushed — commit c6578ea

Original Slice 4 non-first-param fixture (ep_two_param_descent(x, n) -> ep_two_param_descent(x, n - 1)) had arg 0 = x unchanged, arg 1 = n - 1 (arithmetic descent). v3's existing termination prover requires the FIRST argument to descend structurally (is_strictly_smaller(args[0], first_param) — see the "first argument to be {param} - <positive int> or a structurally smaller recursive child" diagnostic at lower.rs:5856). So the fixture failed at compile time before the per-call producer ever ran.

Rework: ep_two_descent(n, m) -> ep_two_descent(n - 1, m - 1). Arg 0 satisfies termination (n - 1); arg 1 still exercises the per-arg ordinal-label contract (param_1, not collapsed to param_0) which was the test's intent. Same behavioral pin, structurally compatible with the v3 prover.

Other Slice 4 test (emits_per_arg_relation_for_multi_arg_self_call) was already green per the CI log.

— sent from eager-bat-178

[briansrls]

Review metadata

• Provider / model: codex / codex-default
• Commit: c6578eae · Trigger: schedule
• Comparison: origin/main @ 4ac0cfd3 ... review/pr-2192-c6578eae @ c6578eae
• Thinking: 41s wall

---

Verdict: APPROVE

Diff is narrowly scoped to integration-test additions in m2_substrate_inhabitance_test.rs. I did not find a concrete violation of the pinned modeling, coding, or testing docs. The new cases stay behavior-focused, exercise the public contract of per_call_descent_evidence, and the parameter-label assertions line up with the current documented scaffold in ordinal_param_label rather than inventing a parallel authority.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: c6578eae · Trigger: manual
• Comparison: main @ 4ac0cfd3 ... slice4-multi-arg-composition @ c6578eae
• Conversation: View conversation

---

1. Story of the diff

This PR is a focused test-only ratchet in src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs for Phase-1 Slice 4 of per-call descent evidence. The first new test builds a three-argument recursive accumulator fixture and asserts that per_call_descent_evidence emits exactly one SubValueRelation per call argument: strict descent for tail, fail-closed unknown for acc + 1, and preserved value for limit at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:859-888. The second test builds a two-argument arithmetic-descent fixture and pins that each argument’s ArithmeticDescent label is indexed independently, specifically param_0 for n - 1 and param_1 for m - 1 at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:920-940. There are no production-code, substrate-schema, or Dag mutation changes in this diff.

2. Invariant categories

1. LAYER MODEL (substrate vs implementation).

N/A — diff is test-only: it consumes existing SubValueRelation shapes in assertions at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:866-888 and src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:925-940, but introduces no substrate type, Dag field, dag.rs change, or new variant.

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

Compliant — fail-closed / facts-flow-forward: the first test requires the evidence vector length to match the recursive call’s argument count at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:859-862, and it explicitly rejects fabricating descent for the growing accumulator by expecting SubValueRelation::SubValueUnknown at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:879-882. The second test also protects single-authority per-argument labeling by asserting the second arithmetic descent remains param_1, not a collapsed global/first-arg label, at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:936-938.

1. CODING.md.

Compliant — the added code uses existing data + free-function shape: per_call_descent_evidence(&dag) is called with explicit Dag input at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:853 and src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:914; the diff adds no methods, hidden state, builders, or new helper API.

1. TESTING.md.

Compliant — the tests are behavior-driven and narrowly named: one pins multi-arg per-relation emission at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:825, and one pins per-arg arithmetic labels at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:894. The fixtures are small source programs constructed only to exercise the recursive-call contract at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:840-849 and src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:905-910, and the assertions target the published evidence carrier rather than unrelated pipeline internals.

1. LOCKED DESIGN DECISIONS.

N/A — no locked design document, schema surface, or substrate target is altered. The v2 reference in the comment at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:827-829 explains the regression contract but does not introduce or diverge from a locked design decision.

1. TRACKED vs UNTRACKED DEBT.

N/A — no TODO, scaffold, temporary bridge, compatibility shim, or new debt marker is introduced. The “Phase-1 broadening Slice 4” comments at src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:826 and src/v3/compiler/tests/integration/m2_substrate_inhabitance_test.rs:895 are provenance for the tests, not a new transitional representation.

3. Verdict

APPROVE. The diff strengthens regression coverage for multi-argument descent evidence without changing substrate or implementation shape. I do not see a diff-anchored invariant violation.