v0.3 — merged bf16, generation E6: LoRA granite_4b_dense_e6 (r=256, 1 epoch / 1,343 steps, eval_loss 0.1408) merged into the v0.2 bucket base; pinpoints unchanged; requires shrew-server >= 0.3.14 (36-value section_type contract)

README: GGUF Versions row — full context per slot rule

README: shrew-server pin v0.2.0 -> v0.2.1

README: add GGUF variant to Versions table

README: measured decode guidance (presence_penalty 0.3 first-pass default, 0.6 retry rescue; enforcement retry-only)

README: link shrew-server reference pipeline; add --served-model-name to vLLM examples

README: correct penalty-mechanics wording (frequency accumulates; repetition is flat per seen token)

README: technical-manual voice; add text-modality spec and penalty-retry guidance

e4 generation: LM-only LoRA r=256 merged, bucket-routed tiling (2026-08-16)

initial commit