[bong-water-water-bong]

engine ggml-org#681: a separate MTP draft head is rejected for qwen4exp on the HRX route. Two minimal fixes, plus what I could measure on the box today.

Fixes

1. src/llama-context.cpp: a qwen4exp/eagle3/dflash draft head has no tok_embd/output and needs its target context. The memory-fitting probe has no target, so the constructor threw "requires ctx_other to be set" before the real draft context (which sets ctx_other in common_speculative_init) was ever built. Warn instead of throw; real draft contexts still get ctx_other.
2. src/llama-sampler.cpp: the backend support probe rejected the top-k sampler candidate graph on HRX0 because it contains RESHAPE. Reshape/view/permute/transpose are metadata-only views resolved by the scheduler, so skip them in the probe instead of failing the sampler graph.

What this measured (strixhalo, fork tip + these commits)

Target: Qwen3.8-Flash-Next-UD-IQ3_XXS (76.3 GiB), -dev HRX0 -ngl 99 -c 32768 --parallel 1 -fa on --cache-reuse 256, house draft head mtp-Qwen3.8-Flash-Next-shared-Q4_K_M.gguf, --spec-type draft-mtp --spec-draft-n-max 2, greedy 512 tokens, 1 rep per cell (not the issue's median of 3).

arm	short (211 tok)	medium (1030 tok)	draft acceptance	GTT peak
no head	27.27 t/s	27.84 t/s	--	82.0 / 82.5 GiB
+ house head, n=2	41.76 t/s	41.25 t/s	315/390 = 80.8 % (short), 312/398 = 78.4 % (medium)	85.0 / 85.8 GiB
gain	+53.1 %	+48.2 %

Head-run clocks sampled per request: short 655-2857 MHz, medium 1064-2803 MHz (pp_dpm_sclk). The no-head arm's clock samples were not persisted before its ticket expired, so the rows are not clock-matched; the +48-53 % is far above the run-to-run spread seen in these cells.

Why this is not the issue's exact row

• -ctv q4_0 (the issue's protocol) is rejected by HRX during prefill: unsupported HRX node ... SET_ROWS ... q4_0 cache. The table above uses -ctk f16 -ctv f16.
• The 18,782-token prompt crashes the server in the CPU fallback: ggml-cpu/ops.cpp:5136: GGML_ASSERT(i1 >= 0 && i1 < ne1) failed inside ggml_compute_forward_set_rows_impl (GGML_OP_SET_ROWS). It reproduces without -md and on the unpatched fork tip 59d7f82 at the same prompt, so it is pre-existing and unrelated to this change (and unrelated to -md). It needs the SET_ROWS/CPU-fallback gap fixed before the long-context row can be measured.
• HRX also lacks TOP_K for the backend sampler (warning only, falls back).

Unproven

• Median-of-3, the issue's 18,782-token long row, and the -ctv q4_0 row: all blocked as above.
• Absolute numbers are on this box with a 76 GiB target and should be re-measured with the exact protocol once the SET_ROWS gap is fixed.

Assisted-by: DeepSeek Harness