[rwkeyes]

Problem and result

A month of performance work lives in 55 notes, a history table and a dozen scripts, and nothing in the repository says
which levers are already closed. This adds docs/PERF-NEXT.md with that record, and reports the first experiment it
proposed.

Result of that experiment: both PLE read-path knobs are closed. A 4 M-row cache and --ple-inflight 1024 are flat;
a 16 M-row cache (1.44 GB) costs 8–17 % prefill. The 94 MB default and inflight 256 stand, and the ~330 ms fixed
per-request cost is not PLE row-read starvation.

Related discussion

none — documentation and evidence.

Approach

1. Kept levers, each with the number that earned it (cache routing −22.1 % TTFT, short-answer drafting +11.3 %,
   thinking suppression −58 % wall, parking off −25…−30 %, NVMe move, calibration 394→276 s, plus upstream's).
2. Closed levers, each with the number that killed it: --batch 2/4, --kv-grow, --suffix-draft/--lookup-chain,
   prefill 16k/32k, --prompt-cache-tail, --conversation-cache-mib, spec-min-p 0.7, --mtp-max-t, 4-bit KV,
   PLE in RAM, expert-cache size, spec_window, DFlash2 on SYCL, and now the PLE row cache and inflight.
3. The remaining cost shape: prefill fits 336 ms fixed + 0.67 ms/token over the 18 fresh legs of the 0.2.13 battery;
   the engine is at parity with stock, so what is left is upstream's or in that fixed term.
4. The untried list, ranked — batch_after 1–2 (measured −21 % median TTFT, default still 4), the fixed-cost
   attribution, the SYCL residency router (route_prior.cu is CUDA-only; 12,288 of 24,576 experts are resident here),
   prefill-chunk intervals, pool affinity on the 7700X, SYCL dispatch overhead, DFlash2 on SYCL.
5. The sweep, six card windows of bench_prefill + longprobe on the 0.2.13 tree, one knob per window plus a
   closing default repeat for drift, production restored between windows.

Validation

Passed:

• Six arms, all with data: default, 4 M rows, 16 M rows, inflight 1024, both, default again. Drift band from the
  two default arms: 0.2 % on fresh 4.2k, 3.6 % on 8.8k, ≤1.2 % elsewhere.
• Verdicts against that band: cache 4 M and inflight 1024 inside it; 16 M rows −8 % (4.2k), −8…−12 % (8.8k),
  −17 % (4.4k follow-up), −10 % (30k), −9 % needle gain at most.
• Production: strata-b70 restored after every window with :8095 http=200 and engine 0.1.58 in the log, and
  checked after the last one — the sweep never left the box down.
• Links and rendering: no broken relative link in the new page, the index entry added, GitHub's Markdown API
  renders its four tables.

Failed:

none observed.

Not run:

• Nothing here re-measures the fixed cost itself — the sweep only excludes the PLE read path as its cause. That is
  the next experiment (item 2 in the new page), at 8–256-token prompts.
• No repeat-traffic test: no leg re-reads the same n-grams, so a bigger row cache for that shape stays untested and
  the page says so.

Limits

• One knob family, one box, one engine version. The ranking in the page is this deployment's, not a general claim.
• The event log, configs and per-leg JSON are shipped; the card-window launcher is on the operator's box as
  ~/bin/strata-card-window.sh, copied into the folder as the script that produced the run.