akumaburn/Swift-Qwen3.8-27b-heretic

docs: link long-context drafter variant

docs: link W4A16 drafter; warn that >262k context kills the engine

docs: drafter selection + vLLM patch requirements for DFlash2

docs: correct DFlash2 acceptance (48-prompt harness supersedes 4-prompt sample)

docs: add served top-512 KL across all variants

Add variants table, unified prefill methodology, spec-depth vs concurrency guidance

Swift-Qwen3.8-27b-heretic: Heretic abliteration of UkisAI/Swift-Qwen3.8-27b (BF16, 12 shards + MTP)

initial commit