Add v3 TriAttention sidecar (blended 1M corpus, R̄=0.610)

Calibrated 2026-04-28 on MI300X with the blended corpus recipe (Hermes + Aureth + Opus_reason + Qwen_reason + Claude_opus + prompts_chat). md5 40c7a828 9163 bytes header + dense per-(layer, head, band) FP64 reduction over 1M tokens.

Pairs with qwen3.6-27b.mq4 + qwen3.6-27b-dflash-mq4.hfq for long-context CASK eviction. Multi-turn smoke (5 turns @ default cask_budget=512+beta=128) on 7900 XTX: turns 1-2 clean, turn 3 soft repetition, turn 4 clean, turn 5 honest 'cannot recall' (eviction by-design drops specific facts). No hallucination, no attractor — degrades gracefully, which is the dense-model R̄=0.610 ceiling.

Usage: `hipfire config set-model qwen3.6:27b cask_sidecar ~/.hipfire/models/qwen3.6-27b.mq4.triattn.blended_v3.bin` and `cask=true`. Or set `cask_budget` higher if you need longer recall horizons before eviction kicks in.

qwen3.6-27b.mq4.triattn.blended_v3.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6cb8026841830cfeb82d2709453aa753f65b5596bfb9cc9c085c808fda6ad22
+size 2359324