slots(bench): benchmark 35B-A3B MoE ROCmMTP candidates on the agent slot (Crown / Chadrock / ace-saber)

[thinmintdev]

What to build

Once the agent (MoE) slot is set up (#634), benchmark the recently-added 35B-A3B MoE ROCmMTP candidates sequentially on the same tuned agent-slot config:

1. Crown (locate exact registry id)
2. Chadrock — chadrock3.6-35b-uncensored
3. ace/saber — chadrock-35b-ace-saber (Qwen3.6-35B-A3B-NSC-ACE-SABER-MTP)

Record per model on a fixed prompt set: gen tok/s, TTFT, prompt-eval rate, GTT/VRAM, short quality impression. Recommend which to use primarily for the agent role. For the top 2, run thinking ON vs OFF and record the delta (tok/s + quality).

Acceptance criteria

• All 3 benched sequentially on the identical agent-slot config (same ctx/-b/-ub/MTP/ngl)
• Results table recorded to a doc (tok/s, TTFT, GTT, quality notes)
• Primary recommendation with rationale
• Top-2 thinking ON vs OFF comparison
• Any load failures noted

Blocked by

• slots: assign+perf-tune chat=qwopus3.6-27b-v2 (vision) & agent=chadrock3.6-35b MoE (thinking off); retire gpu-rocmfp4* slots #634 (agent slot setup + perf tune)

[thinmintdev]

Resolved: Crown = qwen3.6-35b-crown-halo (registry id). The three 35B-A3B MoE bench candidates are therefore:

1. qwen3.6-35b-crown-halo (Crown)
2. chadrock3.6-35b-uncensored (Chadrock)
3. chadrock-35b-ace-saber (Ace-Saber)