Diagnostic wave coverage and limits

This common ledger was consolidated during closure. The research plan, counter-arm plan, request fixtures, phase order and hashes were retained before their respective requests; this is not a retrospectively frozen serving comparison.

Request	Status	Evidence / limit
Research general serving stacks and vLLM/SGLang	Complete	research-and-plan.md and pinned source registry; analogies, not imported speed claims.
Refine and execute reverse/forward cache experiment	Complete after final analyzer gate	Two images, fresh forward/reverse pairs, exact bodies and per-request counters.
Identify internal bottleneck	Measured structural constraint	Counter-only image observes independent state-row pressure and page availability. Not an end-to-end speedup.
Add instrumentation and debugging	Complete locally	Ten host-pool metrics, CPU reproducer, 37 applicable tests, pinned managed build. Disk io_uring runtime tests remain unqualified.
Repeat trace off/on	Complete, causal conclusion blocked	Three C1 and three C4 pairs; all 15 request pairs differ in generated token sequence. Equal length/seed does not establish equal work.
Preserve monitoring across swaps	Live verified locally	Generation-specific Prometheus discovery; ten new series for each counter incarnation. Existing dashboard verification repeated after restoration.
Trace exact kernel limiter	Not collected	No new Nsight critical-path or hardware-counter capture. Scheduler samples are overlapping rank intervals.
Improve allocator or eviction policy	Next experiment	No policy change tested. Fixed-byte row/page repartition and protected workloads must precede gain claims.
Prove parity with incumbent	Not measured in this wave	Different checkpoints/stacks; no new matched incumbent performance matrix or semantic-quality qualification.
Reusable skill/plugin	Skill refined	Existing kernel-tuning skill covers pool counters, decoded-work matching and retained rollback. No new plugin needed.
Retain reports and easy conversion	Retained	Pinned source, recipe, patch/build, corpus, responses, counters, restoration, derived summary and bounded next test. No promotion or upstream submission.

All natural-output cache successes mean visible transport completion, not semantic correctness. Cache C1 means serial HTTP requests, while the serve permits C4. Host cache is configured 32 GiB; reported resident bytes and row/page capacities are separate measurements. Both GPUs retain 300 W limits.