hal0 v0.9.1

ROCmFPX llama.cpp runner support, plus a safer notes-aware self-update.

Highlights

• ROCmFPX runner support: GGUF quant-family detection, rocmfpx-rocm / vkfpx-moe seed profiles, and a build/quantize agent skill.
• hal0 update now shows cosign-verified release notes and asks before applying (staged prepare → commit).

Added

• Updater prepare / commit split. hal0 update downloads + cosign-verifies + extracts a release and shows its notes — with breaking/migration callouts — before activating anything; --yes skips the prompt for headless/cron. Adds POST /api/updates/prepare + /commit (#1075).
• Release notes in the update. The release build bundles RELEASE_NOTES.md + release.json into the cosign-verified tarball; a CHANGELOG section's ### Highlights / ### Breaking / ### Migrations subsections become the hal0 update callouts (#1078).
• ROCmFPX quant detection — the registry classifies the ROCmFPX / ROCmFP{3,4,6,8} quant family from a GGUF filename so FPX slots resolve their launch command (#1068).
• ROCmFPX seed profiles rocmfpx-rocm (ROCm0 dense) and vkfpx-moe (Vulkan0 MoE) for the custom ROCmFPX runner (#1069, renamed in #1076).
• vkfpx-dense seed profile — the Vulkan0 lane for DENSE ROCmFP4, for prefill-bound dense workloads (Vulkan wins prompt-processing); complements the decode-optimal ROCm0 rocmfpx-rocm.
• hal0-quantize agent skill — build the ROCmFPX toolchain and quantize a model to ROCmFP4/FPX (#1071).
• ROCmFPX bench tooling: server-ab (MTP / concurrency) aggregation in the benchmark SUMMARY, FPX sweep cells, run provenance (#1072).
• Continuous batching — per-slot parallel field. A slot can now set
  llama-server's --parallel / -np sequence-slot count so concurrent
  requests share the once-loaded weights instead of serializing through a
  single sequence and thrashing one prompt cache (the win the shared-slot
  architecture already earns but never harvested — every seed profile pins
  --parallel 1). None inherits the profile; a value >1 also emits
  --kv-unified so --ctx-size stays a SHARED pool (each request may use the
  full context) rather than being silently split to ctx/N per slot. Emitted as
  a slot override (beats the profile, loses to hand-authored extra_args);
  surfaced in the slot drawer with a shared-pool hint. The dead haloai
  workers field is deprecated (inert; a non-default value now logs at
  launch). MTP x batching runs but logs mtp.batched_speculation (unproven on
  gfx1151, bench-gated). Seed-profile defaults stay --parallel 1 pending the
  on-box -np sweep (server_ab.py --mode batch). See the
  concurrency-batching plan handoff.

Changed

• The plain rocm / vulkan seed profiles are reduced to basic flags
  (-ngl 999 -fa on --jinja); per-model KV/batch tuning now lives in the
  model's defaults.extra_args (#1076).
• Seed-profile intent labels normalised to terse structural tags
  (e.g. ROCmFPX · DENSE · MTP, VULKFPX · MOE · MTP, ROCm, Embeddings) —
  no served-model names, no filler.
• Slot units are re-rendered through the new code during an update's commit
  step, so a subsequent restart uses current argv (#1075).

Removed

• Legacy MTP toolbox seed profiles rocm-moe and rocm-dnse (superseded by the
  ROCmFPX profiles); rocmfpx-moe renamed to vkfpx-moe to indicate its Vulkan
  lane (#1076).

Migrations

• Slots pinned to a removed/renamed seed profile (rocm-moe, rocm-dnse, rocmfpx-moe) auto-fall-back to the backend's basic profile (rocm / vulkan) on launch — existing slots keep working with no operator action (#1076).