chore(release): bump version to 0.6.0

[timtoole02]

Version bump only: 0.5.4 → 0.6.0. No functional change in this PR.

Minor bump (not patch) — 81 commits since v0.5.4 add new model architectures, a new
multi-node execution lane, and new Metal/AVX-512 kernels.

What ships in 0.6.0

New architectures

• LFM2 / LFM2.5 hybrid short-conv + GQA bringup (feat(lfm2): implement the LFM2/LFM2.5 short-conv forward in the runnable lane #614), engine lane (lfm2: fix the reasoning split and fail closed on unrunnable rows #615), Metal (lfm2(metal): GPU-resident lane by default, and a 14x prefill #620)
• qwen3moe bringup (Add qwen3moe (Qwen3-30B-A3B) with streamed Q4_0/Q4_1 experts #612)
• gemma4 ghost-MoE on Metal (Add Gemma 4 Ghost-MoE Metal acceleration #602)
• qwen35 Q8 on Metal (runnable: admit Q8_0 weights on the qwen3.5 resident Metal lane #618) — thanks @heskew

Cluster fabric

• Whole-request placement across independent nodes, 0 hops/token (feat(fabric): place whole requests across independent Camelid nodes #616) — thanks @karan68
• Bearer auth for fabric nodes (feat(fabric): authenticate to nodes with a bearer token #627)

Performance

• Q4_K owner scale fold: AVX2 (perf(inference): carry the Q4_K owner's group scale in its widening madd #617, @karan68) and AVX-512 (perf(inference): carry the Q4_K owner group scale in its widening madd on AVX-512 #629) — folded inner needs
  no VNNI (avx512f+bw only)
• Q6_K owner maddubs inner (perf(inference): widen the Q6_K prefill owner's integer inner #613) — thanks @karan68
• Ghost-MoE CUDA pinned host expert tier + KV ring (Add catalog-managed Gemma 4 Ghost MoE on Windows CUDA #607)

Fixes

• Bonsai CPU lane on Windows — CPU block-dot restored, compile-time cfg! no longer
  stands in for a runtime GPU check (fix: give the CPU a consumer for every wire-only quant #628)
• Bonsai catalog arch split — 4B/8B are qwen3 (dense), only 27B is qwen35 (fix(catalog): label the Bonsai 4B/8B rows qwen3, matching their GGUF headers #624)
• Non-catalog SHA256 guard (fix(api): pin supported-row identity to the certified sha256 #625)
• Ornith support-level contradiction (plan: end the ornith support_level contradiction, arch-guarded #622)
• Q8 gate tests (qwen35: admit K-quants to the resident Metal lane, and make the plan disclose the lane that actually serves #621)

Docs

• qwen35 K-quant / Metal truth-sync (docs(qwen35): name the macOS Metal frontier the K-quant admission opened #626)
• README discoverability (Improve README search discoverability #604)

Not included

Open PRs #619 (fabric serve) and #623 (HF pull) are excluded — not on main at cut time.

Release checklist

• 4 files, 5 insertions, 5 deletions
• CI green on this PR (authoritative release-tree signal)
• Merge, then lightweight tag v0.6.0 on the merge commit
• release.yml publishes 9 signed assets