[bong-water-water-bong]

Retires the known_fail("#450") guard: engine ggml-org#450's accepted regression is not a regression on AMD's current kernels, so the case states what ships instead of asserting the route is never selected.

Measurement (strixhalo, gfx1151)

Qwen3-4B-Q5_K_M, route on vs off (GGML_HRX_DISABLE_DISPATCH=common.rmsnorm_binary_symmetric_i4_k32), everything else identical, -dev HRX0 -ngl 99 -t 4 under ~/lb/thermal-run.sh:

	route on	route off
pp512	3033.38 ± 23.60	2991.08 ± 39.41
pp2048	2647.52 ± 95.31	2613.31 ± 117.61
tg64	65.65 ± 0.36	63.61 ± 0.11

Qwen3-4B-IQ4_XS, the second qualifying quant: pp512 2933 vs 2971, pp2048 2617 vs 2573, tg64 73.70 ± 0.15 vs 73.00 ± 0.46 — mixed, inside the noise.

Numerics on held-out natural text (Gutenberg Pride and Prejudice + Moby Dick, 1.97 M chars — not the repetitive text that made the first run meaningless):

• PPL = 16.3451 ± 0.74733 over 20 chunks at -c 512 (arm A, logits saved);
• cross-arm llama-perplexity KLD over 5 chunks: Mean 0.000000 ± 0.000000, maximum 0.000056, max Δp 0.005 %;
• three 64-token greedy completions at temperature 0: byte-identical between the arms (931 bytes each).

Device use was confirmed, not assumed: the runs show dev HRX0 / ngl 99 and reach pp512 ≈ 3000 tok/s on a 4B against 89.76 tok/s for the same model at -ngl 0.

What the case asserts now

symmetric_i4_consumer_qualification (re-added by #105 with its pre-#104 expectations) asserted the symmetric-i4 producer is never selected. It now asserts the truth: a qualifying graph — a Q5_K/IQ4_XS weight consuming the normalised value, directly or through a reshape — selects loom_libs:ggml_rmsnorm_binary_symmetric_i4_k32, and a graph with no such consumer does not.

Its q8-publication assertions are removed from this case on purpose: ggml_rmsnorm_binary_q8_1_x4_publish went with f2099e9b, so that is engine ggml-org#449's territory and stays with the guards there rather than failing a route case for a publication reason.

Gate (the fork has no CI)

Local host suite on this base (29dd8820, CPU only, no HRX device):

host: 66 case(s)
host: 47 passed, 13 failed, 6 known-fail, 0 unexpected-pass, 66 total

symmetric_i4_consumer_qualification is now [PASS]; the 6 remaining known-fail rows are ggml-org#449's; the 13 [FAIL] rows are the class A plan-shape, class B kernel-name and the qwen_router_top8_dispatch NEEDS-DEVICE rows, unchanged.

Related: engine ggml-org#450 (this ruling, with the numbers), ggml-org#445 (the classification), ggml-org#449 (publication).

🤖 Generated with Claude Code