Prep #24: Define fixed-fixture B-card operator A/B gate

[inaniloquentee]

Background

Prerequisite for #24, assuming the Phase 1-7 roadmap foundation in #6 is already available.

After #24 obtains a stable Blackwell baseline, candidate B-card operators should first be tested on fixed samples/tensors before being compared inside the full post-training loop. This issue defines the #24-specific fixture/export layer that bridges real baseline runs to operator-level candidate A/B tests while reusing the existing Phase 1-7 comparison and reporting mechanisms.

Scope

• Define the fixed-sample and fixed-tensor fixture set that [Experiment] Build Blackwell train-inference mismatch baseline for VIME post-training #24 should export from a baseline run.
• Include only experiment-specific tensors/metadata needed for candidate operator A/B: hidden states or logits when available, LM-head weights/shards, target IDs, active masks, selected positions, TP metadata, dtype/quantization policy, old logprobs, recomputed logprobs, and existing contract/report IDs.
• Add a small export/replay entry point or runbook step that materializes sanitized fixtures from a [Experiment] Build Blackwell train-inference mismatch baseline for VIME post-training #24 baseline artifact.
• Add a synthetic smoke fixture path so the export/replay wiring can be tested without B-card hardware.
• Define candidate operator A/B gates over existing Phase 4/5 comparison paths: selected-logprob max/mean/p99 error, active-token dlogp metrics, backward parity when applicable, throughput, memory, and fallback status.
• Define how the same fixture is reused for baseline native path, vLLM path, RL-Kernel/reference path, and candidate B-card operator path.

Non-Goals

• Do not implement B-card CUDA kernels here.
• Do not create new operator comparison or distributed contract harnesses; reuse Phase 4: Add single-card operator comparisons and drift attribution #30/Phase 5: Add distributed operator contracts and A0-A5 profiles #31 outputs.
• Do not create new audit/result-cube/report infrastructure; reuse Phase 3/7 outputs.
• Do not run the final [Experiment] Build Blackwell train-inference mismatch baseline for VIME post-training #24 baseline or candidate experiment here.
• Do not claim train-inference consistency improvement unless the later [Experiment] Build Blackwell train-inference mismatch baseline for VIME post-training #24 A/B run uses the same fixed workload and declared contract.

Acceptance Criteria

• [Experiment] Build Blackwell train-inference mismatch baseline for VIME post-training #24 has a documented fixed-fixture export format for baseline-to-candidate operator testing.
• A synthetic smoke fixture can exercise the export/replay path on CPU or a single-card developer machine.
• Real baseline fixtures can be replayed through the existing operator/distributed comparison paths without inventing a separate controller.
• Candidate operator A/B gates are defined before any B-card kernel replacement is evaluated.
• The issue clearly separates fixed-fixture operator validation from full post-training A/B conclusions.

Hardware Notes

• The fixture format, synthetic smoke path, and replay wiring can be developed on CPU or a single-card machine.
• Capturing real [Experiment] Build Blackwell train-inference mismatch baseline for VIME post-training #24 fixtures requires the Blackwell/B200/GB200 baseline environment.
• Testing a real candidate B-card CUDA operator requires the target B-card hardware and matching RL-Kernel/Megatron/vLLM build.

Dependencies

• Parent experiment: [Experiment] Build Blackwell train-inference mismatch baseline for VIME post-training #24.
• Roadmap foundation: RL-Kernel integration roadmap for vime #6 Phase 1-7.
• Reuses Phase 4/5 comparison contracts: Phase 4: Add single-card operator comparisons and drift attribution #30, Phase 5: Add distributed operator contracts and A0-A5 profiles #31.
• Reuses Phase 6/7 profiling/report outputs: Phase 6: Add profiling-guided fast-path expansion telemetry #35, Phase 7: Add RL-Kernel CI, benchmark matrix, and release rules #36.