Skip to content

contract(trace-moe-gpu-sub-stages-v1): v1.3.0 → v1.4.0 — falsifier hygiene#1526

Merged
noahgift merged 1 commit intomainfrom
contract/moe-sub-falsifier-status-v1.4.0
May 6, 2026
Merged

contract(trace-moe-gpu-sub-stages-v1): v1.3.0 → v1.4.0 — falsifier hygiene#1526
noahgift merged 1 commit intomainfrom
contract/moe-sub-falsifier-status-v1.4.0

Conversation

@noahgift
Copy link
Copy Markdown
Contributor

@noahgift noahgift commented May 6, 2026

Summary

YAML-only amendment that realigns FALSIFY-MOE-SUB-001/002 falsifier test: invocation strings with their live bindings in code, and promotes their statuses to match the implementation_stages they actually discharge.

This is the same drift class M71 (#1511, pv lint --strict-test-binding PV-VER-002) closes mechanically going forward — trace-moe-gpu-sub-stages-v1 was authored before that gate landed and needs manual realignment.

Drift caught

Falsifier v1.3.0 cited test Live binding
SUB-001 cargo test -p aprender-serve --lib falsify_moe_sub_001_new_stages_parse (singular, doesn't exist) cargo test -p aprender-serve --lib falsify_moe_sub_001 (5-test suite — round_trip × 2 + canonical_order + parse_list × 2)
SUB-002 cargo test -p apr-cli --test falsify_moe_sub_002_byte_identity (wrong crate, wrong test name) cargo test -p aprender-serve --features cuda --test qwen3_moe_gpu_per_stage_diff falsify_moe_sub_002 -- --include-ignored (heavy harness from M80 #1524, cuda+ignore-gated)

Status promotions

Falsifier v1.3.0 v1.4.0 Justification
SUB-001 PROPOSED DISCHARGED 5 lib tests pass in <1s; bound to M-MOE-SUB-1 (M60 #1499 + M64 v1.1.0 wireup)
SUB-002 PROPOSED ALGORITHM_LEVEL_DISCHARGED Heavy harness exists + --ignored --list confirms 1 test discoverable; full DISCHARGED still blocks on operator-dispatched --include-ignored run on lambda-vector RTX 4090 against cached 17.3 GB Qwen3-Coder GGUF
SUB-003 PROPOSED unchanged Still requires LIVE bisection on RTX 4090 (same precondition as M-GPU-MOE-1.4)
SUB-004 PROPOSED unchanged Still requires M-GPU-MOE-1.4 fix PR to land citing a specific stage

Verification

$ pv validate contracts/trace-moe-gpu-sub-stages-v1.yaml
0 error(s), 0 warning(s)
Contract is valid.

$ cargo test -p aprender-serve --lib falsify_moe_sub_001
test result: ok. 5 passed; 0 failed; 0 ignored; 0 measured; 15274 filtered out

$ cargo test -p aprender-serve --features cuda --test qwen3_moe_gpu_per_stage_diff falsify_moe_sub_002 -- --ignored --list
falsify_moe_sub_002_cpu_gpu_traced_per_stage_diff: test
1 test, 0 benchmarks

Production hot paths

Byte-unchanged. Additive-purity invariant pinned in v1.1.0 still holds — this PR touches only the YAML metadata + falsifier text + status fields.

Test plan

  • pv validate contracts/trace-moe-gpu-sub-stages-v1.yaml → 0/0
  • SUB-001 binding: 5/5 lib tests pass
  • SUB-002 binding: heavy #[ignore]+cuda-gated test listed
  • No code touched

🤖 Generated with Claude Code

…giene amendment

Realigns FALSIFY-MOE-SUB-001/002 `test:` invocation strings with their
live test bindings in code, and promotes statuses to match the
implementation_stages they discharge.

Drift caught: at v1.3.0 the contract cited test names that didn't
match the actual code:
- SUB-001 cited `falsify_moe_sub_001_new_stages_parse` (singular)
  but live binding is a 5-test suite under prefix `falsify_moe_sub_001_*`
  (round_trip × 2 + canonical_order + parse_list × 2) in
  aprender-serve --lib.
- SUB-002 cited `cargo test -p apr-cli --test falsify_moe_sub_002_byte_identity`
  but live binding is `falsify_moe_sub_002_cpu_gpu_traced_per_stage_diff`
  in `aprender-serve --features cuda --test qwen3_moe_gpu_per_stage_diff`
  (heavy harness from M80 PR #1524, `#![cfg(feature = "cuda")]` +
  `#[ignore]`-gated).

This is the same drift class M71 closed mechanically via PV-VER-002
(`pv lint --strict-test-binding`) for §50.4 cascade contracts —
trace-moe-gpu-sub-stages-v1 was authored before that gate landed and
needs manual realignment.

Status promotions:
- FALSIFY-MOE-SUB-001 PROPOSED → DISCHARGED (5 lib tests pass in <1s,
  verified live)
- FALSIFY-MOE-SUB-002 PROPOSED → ALGORITHM_LEVEL_DISCHARGED (heavy
  harness exists + listed; full DISCHARGED blocks on RTX 4090
  --include-ignored dispatch)
- FALSIFY-MOE-SUB-003/004 unchanged (LIVE bisection + fix PR pending)

YAML-only — production hot paths byte-unchanged. Contract version
v1.3.0 → v1.4.0; status stays ACTIVE_ALGORITHM_LEVEL.

`pv validate` 0/0.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@noahgift noahgift enabled auto-merge (squash) May 6, 2026 04:24
@noahgift noahgift merged commit 01732c5 into main May 6, 2026
11 checks passed
@noahgift noahgift deleted the contract/moe-sub-falsifier-status-v1.4.0 branch May 6, 2026 04:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant