Fix training, serving, and verification with measured B200 follow-through

[cfregly]

This follow-through repairs incorrect or incomplete training, serving, numerical verification, and profiler behavior, then retains measured B200 and two-B200 outcomes without requiring every pair to exceed a fixed speedup.

• Reuse serving engines and separate startup from steady-state work; fix request admission, monotonic latency, exact GPU visibility, and optional spillover placement. Keep throughput/latency tradeoffs explicit.
• Correct training accumulation, loss reporting, FSDP objectives, rotary buffer precision, and timing. Add an explicitly bounded two-rank AdamW-overlap option and lossless full-reference transport under the existing 1 GiB limit.
• Tighten full-output and independent arithmetic checks for persistent decode, KV projection quantization, Ozaki, and MoE; remove graph temporary/copy work without weakening correctness or memory checks.
• Repair coordinated NCCL profiling and clock ownership, preserve failed or incomplete attempts, and synchronize the generated lab documentation.

Validation: supported direct training mains, public benchmark paths, independent numerical checks, and targeted memory/profiler runs executed on one or two B200s. The public DDP opt-in passes exact full-input/output checks at 8a698d73f22cc80c50b1e72e57eb230509f3f760; its short three-update timing remains a 0.929x no-win. The separate full-epoch DDP loop screen measured 1.0236x, and other results retain their workload-specific limits. All 61 generated README entries match; focused regression checks pass. Final hosted CI is being run against this completed implementation.

Limitations: these are portable development runs on the reported virtualized host. Generic FSDP wrappers remain intentionally unsupported despite all eight direct mains executing. Projected-K initcheck reports remain unresolved: the bounded diagnostic completes, but ordinary Linear and empty-output controls still report reads. No production prefill or suppression was added. No claim is made that every example is faster or that every memory check passes.

Detailed source identities, retained evidence hashes, successful checks, failures, and measured tradeoffs are in the follow-through report.