perf(dflash): add opt-in exact prefill bands

[cheese-cakee]

What this draft does

This is a hardware-blocked candidate for the August 10 DeepSeek V4 exact-prefill objective. It does not claim the 1.5x target.

• Adds an opt-in DFLASH_DS4_EXACT_PREFILL_BANDS=1 path for exact prefill widths 2 through 4.
• Keeps exact attention token-by-token and uses the existing compressor-safe boundary splitter.
• Skips intermediate terminal HC collapse, output norm, vocabulary projection, and host logits readback only for the new nonfinal exact width-2-through-4 bands.
• Keeps the final chunk, snapshot boundaries, and required capture paths producing logits.
• Keeps all existing q=1, dense, and sparse scheduler behavior unchanged.
• Keeps the default off. --chunk 1 remains the explicit escape hatch.
• Prevents exact prefill capture hooks from entering the intentionally approximate fused-verification path.

Why it is a draft

Neither lucebox3 nor lucebox5 was available. This branch has no HIP/gfx1151 production-model differential, no issue-#516 cached-graph burn-in, no sanitizer result, no performance profile, and no lucebox5 four-cell benchmark. It is not ready to merge or use for a performance claim.

The target benchmark contract is also incomplete. The default acceptance interpretation remains:

median heterogeneous-candidate throughput / median Strix-candidate throughput >= 1.5

with matched heterogeneous baseline/candidate and Strix baseline/candidate cells.

Local evidence

Base: 4fa8937c69aa68658a14ec3cab16e1da22886dd0

Head: ed8991707d25cfde956e2d599f4570514bc0146c

Passed:

• Tests were added before implementation; the first compile failed on the deliberately missing policy declarations.
• The modified backend and unit-test C++ objects compile successfully in WSL Ubuntu with GCC 13.3 and CUDA 12.6 headers, architecture 89.
• deepseek4_graph.cpp compiled successfully after the graph-hook change.
• git diff --check, staged diff checks, and git show --check pass.
• Policy coverage includes default-off behavior, widths 1/2/3/4, cap and fallback behavior, compressor-safe scheduling around ratio-4 boundaries, final/snapshot/execution logits policy, and fused-verify opt-out.

Not passed or not run:

• The full test_deepseek4_unit target did not finish. Earlier attempts failed while compiling vendored mmvq.cu with WSL service errors; the final single-job attempt timed out after five minutes. This is compile-only evidence, not a unit-test pass.
• No model run or exact differential exists yet.
• No HIP/gfx1151 qualification, repeated-request/reset test, snapshot/capture content comparison, sanitizer run, timing, or lucebox5 acceptance exists.

Required next validation

On lucebox3:

1. Build this head with the compatible, SHA-pinned model/toolchain.
2. Run the bounded issue-Bug: ggml_set_rows faults on HIP/ROCm after many reuses of a cached single-token graph (Strix Halo / gfx1151) #516 cached set-rows burn-in.
3. Compare exact width 1 against widths 2, 3, and 4 across prompt tails, compressor and SWA boundaries, first hash-routed and learned-router layers, repeated requests, cache reset, snapshots, and DSpark capture windows.
4. Compare routing IDs/weights, HC, raw/compressed KV, compressor/indexer state, logits at a declared tolerance, greedy/continuation tokens, capture rows, and final cache position.
5. Run the applicable sanitizer and record a profile.

Then run the matched four-cell acceptance on lucebox5 with identical model SHAs, inputs, settings, clocks, caches, warmups, and sample counts. Report medians, dispersion, both before/after changes, and the heterogeneous-after / Strix-after ratio.

Integration notes

This touches nearby backend/graph/docs code also changed by #548 and #554. Those changes need a semantic rebase and another review if they land first. Drafts #569/#575/#577 and lossy PFlash #589 are not acceptance evidence for this exact-prefill change.