[Essoz]

The earlier RSI pilot kept the trainer and distributed integration outside the agent's editable surface. This adds a reproducible two-GPU qualification workload for the implementation-contract evaluation in #165, and records what the current instrumenter/checker actually supports.

The probe uses a 5,504-parameter Transformer with DDP, gradient accumulation, unequal valid-token counts per rank, clipping, AdamW, a scheduler, and FP32/BF16 modes. An independent combined-batch gradient calculation distinguishes correct global normalization from a local-mean fault. Fresh processes save/resume complete per-rank checkpoints; a typed comparator checks all stored tensor bits, scalar/container types and fields without normalization. API and active subclass tracing have explicit update scopes and suppress bookkeeping.

Measured on Ring18's two A16 GPUs, Python 3.10.14 / Torch 2.2.2+cu121:

• 13 runs completed; all 9 complete checkpoint comparisons pass, covering two ranks and 274 tensor pairs per comparison. Plain/API/subclass states, plain↔subclass checkpoint interchange, and plain/subclass BF16 agree exactly, including complete checkpoint-file hashes.
• Healthy FP32 gradient max error 4.47035e-8; injected local-mean fault 0.0695832. Faulty ranks still agree with each other, so rank equality alone misses this error.
• Rank-local API inference yields 592/593 active rules. Fresh healthy runs have 2/2 alarms; fault runs 1/2, with no fault-only violated rule. These diagnostics did not distinguish this known fault.
• Subclass samples expose stale timestamps: every recorded post-update sample in full FP32/BF16 runs predates that update's AdamW.step. First-update before/after times match despite changed hashes for all 27 parameters per rank. No core timestamp fix or rule filtering is included.
• Total supervised workload wall time 211.29 s. This is qualification, not an optimization or representative overhead claim; short-loop timing and trace-size observations are documented with their limitations.

Validation: 28 focused tests pass locally and 28 on Ring18 with CUDA visible, configured hooks pass, explicit strict mypy passes over seven modules, and 188 exported artifact hashes verify. The CUDA-hidden configuration separately preserves 26 passes / 2 tracer-startup failures caused by metadata inspection initializing CUDA; the passing configuration does not fix that existing compatibility gap. All study processes have exited and GPUs are idle.

Scope: no FSDP, stochastic dropout, DataLoader cursor recovery, partial accumulation windows, changed world size, FP16 scaler/overflow or paired agent study. Optimizer state is verified independently but not registered with the model tracker; no cross-rank invariant checker is supplied. The reused supervisor's process-group RSS excludes separate torchrun sessions, and its size limit covers API traces only; these limitations are recorded rather than presented as full enforcement guarantees. Large traces/checkpoints remain outside Git.

Stack dependency: based on open #164 (codex/rsi-pilot-20261030) to use its integrated core and supervision changes. This PR contains only the distributed probe, regression tests, comparator and measurement report; merge/rebase onto main after #164 lands and recheck the composed result. It does not close #165 or establish agent benefit.

Full reproduction, source revisions, scope and artifact locations: measurement report.