Add DeepSeek-V4-Flash-0731 guide + card (Config-I 2.88bpw GGUF)

[TheTom]

Adds an offlabel operating guide + card for DeepSeek-V4-Flash-0731 as a Config-I hybrid GGUF quant (284B / 21B-active MoE at 2.88 bpw, 95 GiB, fits one 128GB Spark or Mac). GGUF-focused; MLX guide is backlog.

Status: EARLY. One day of quant-release burn-in, single tester. Serving/config findings are solid (each cost real debugging time); behavioral findings are observational and labeled as such.

What's in it

• Quality: Wikitext PPL 15.64 (ctx 512) / 12.86 (ctx 2048), CUDA GB10.
• Two silent backend traps (both fixed in the fork build, both cost hours to find):
  • Metal fused rotate-act kernel corrupts batched prefill (PPL = vocab size) while decode looks fine; TQ_NO_ROTATE=1 routes around it.
  • CUDA fused-TQ3_1S kernel diverges ~2%/layer on real weight distributions and compounds to token soup; synthetic op-tests pass. Found by node-level activation diffing.
• Sampling floor: official temp 1.0 / top_p 0.95; greedy loops on this family.
• Two-stream trap: the answer often lands in reasoning_content while content is empty or degenerate; content-only harnesses score correct turns as failures.
• Repetition-loop failure mode on long symbolic math, with a confirmed --dry-multiplier 0.8 / --repeat-penalty 1.1 fix.

DwarfStar (antirez/ds4) as a second stack

A/B on the same 0731 weights, different low-bit quant + MLA-native engine, same M5 Metal:

• ~5x decode (29 to 35 vs ~6 tg t/s) and an order of magnitude on prefill.
• At the same sampling, no loop and terse output, which exonerates the base model and attributes the loop + budget-eating verbosity to our stack (Config-I quant and/or the fork's llama.cpp serving path). This A/B can't split quant-from-engine; that's flagged backlog.

Lint (scripts/check_guides.py) passes: valid frontmatter, required keys, no em/en-dashes.