slots: assign+perf-tune chat=qwopus3.6-27b-v2 (vision) & agent=chadrock3.6-35b MoE (thinking off); retire gpu-rocmfp4* slots

[thinmintdev]

What to build

Point the renamed slots at the new ROCmFP4+MTP models and tune for performance.

chat ← qwopus3.6-27b-v2 (27B dense, ROCmFP4+MTP, vision via /mnt/ai-models/qwopus3.6-27b-v2/mmproj-F32.mmproj):

• wire the mmproj for vision; keep --flash-attn on --no-mmap + the MTP draft args; enable_thinking=true (reasoning available for Hermes).
• bump ctx 4096 → propose 32768 (full dense KV — memory-aware; record GTT).

agent ← chadrock3.6-35b-uncensored (35B-A3B MoE, MTP):

• enable_thinking=false by default; tune -b/-ub, ngl, ctx (MoE/hybrid KV can go larger); add MTP draft args if the MoE gguf supports --spec-type draft-mtp.

Retire/fold the now-redundant gpu-rocmfp4 / gpu-rocmfp4-moe slots (their tuned args move onto chat/agent).

Acceptance criteria

• chat serves qwopus-27b with working vision (image smoke)
• agent serves the MoE with thinking off by default
• perf recorded (tok/s, TTFT, GTT) and acceptable
• redundant gpu-rocmfp4* slots removed; CT105 verified

Blocked by

• slots: rename roles primary→chat/agent-hermes→agent