hal0-quantize: validate & promote experimental ROCmFPX variant

[thinmintdev]

Context

The hal0-quantize agent-skill (installer/agent-skills/hal0-quantize/) already exposes an experimental build variant via ROCMFPX_VARIANT:

ROCMFPX_VARIANT	repo	branch	checkout/build
stable (default)	Hal0ai/Hal0_ROCmFPX	main	$HOME/ROCmFPX
experimental	charlie12345/ROCmFPX	experimental-rocmfpx-branch	$HOME/ROCmFPX-experimental

experimental builds in a separate checkout/build dir so it never clobbers the stable tree. The branch carries FP3/FP6/turbo3_0/turbo4_0 formats + MTP edge-case work not yet in the Hal0ai fork.

Deferred work (this issue)

The skill can build/quantize with the experimental tree today, but nothing downstream consumes experimental-only output yet.

• Build the experimental tree on the Strix Halo box and smoke-test llama-quantize against a real BF16 GGUF (ROCMFPX_VARIANT=experimental ARCH=strix scripts/rocmfpx-build.sh).
• Exercise the experimental-only presets (FP3 / FP6 / turbo3_0 / turbo4_0) end-to-end and confirm the quantize-rocmfpx-agent.sh preset mapping still holds on that branch (update references/presets.md if it diverges).
• A/B the experimental build vs stable (Hal0ai/Hal0_ROCmFPX@main = 5b39566) on our GGUFs for output-identity + perf, same as the prior charlie/main vs 7aa484a comparison.
• Decide whether to promote experimental formats into serving: would need a runner image build from the experimental branch (ghcr.io/hal0ai/hal0-rocmfpx), registry quant= entries, and slot wiring — the serving path in the hal0-rocmfpx-runner notes, out of scope for the quantize skill.
• If promoted, consider mirroring experimental-rocmfpx-branch into Hal0ai/Hal0_ROCmFPX so the default fork is the single source.

Notes

• The published ghcr.io/hal0ai/hal0-rocmfpx:server image only bundles llama-server/bench/cli (not llama-quantize), so quantization builds from source regardless of variant.
• No pinned commit — variants track their default branch; pin via ROCMFPX_BRANCH=<ref>.