v0.3.1: promote beta — DFlash repairs, Gemma/Maple, image and vision support

.agents/skills/hipfire-kernel-atlas/SKILL.md

@@ -93,7 +93,7 @@ mkdir -p .codeinsight+research/kernel-atlas/tasks
 Paths and model files must exist on the machine; swap tags/files from
 [`registry/models.json`](../../../registry/models.json).
 
-Collect AR smoke with ISA + dispatch (illustrative):
+Collect AR smoke with ISA + dispatch (illustrative; small model for speed):
 
 ```bash
 python3 scripts/kernel_atlas.py collect-ar \
@@ -114,20 +114,22 @@ python3 scripts/kernel_atlas.py collect-ar \
   --output .codeinsight+research/kernel-atlas/runs/atlas.jsonl
 ```
 
-DFlash collection (prompts under `benchmarks/prompts/` when present):
+DFlash collection (acceptance fixture = Qwen3.8-27B MQ4XT + measured draft;
+prompts under `benchmarks/prompts/` when present):
 
 ```bash
 python3 scripts/kernel_atlas.py collect-dflash \
-  --target ~/.hipfire/models/qwen3.5-27b.mq4 \
-  --draft ~/.hipfire/models/qwen35-27b-dflash-mq4.hfq \
+  --target ~/.hipfire/models/qwen3.8-27b.mq4-xt \
+  --draft ~/qcal/ladder-v2/drafts/qwen3.8-27b-dflash.mq4v2.hfq \
   --prompt-file benchmarks/prompts/merge_sort_thinking_off.txt \
-  --workload qwen3.5-27b-dflash-merge-sort \
+  --workload qwen3.8-27b-mq4-xt-dflash-merge-sort \
   --max-tokens 256 \
   --ctx 2048 \
   --kv-mode q8 \
   --output .codeinsight+research/kernel-atlas/runs/atlas-dflash.jsonl
 ```
 
+
 Suggest / task / eval:
 
 ```bash

.agents/skills/hipfire-tester/SKILL.md

@@ -39,13 +39,30 @@ earned).
 ## Quick start (bring-up only)
 
 ```bash
+# Fast smoke (small model — speed only, not acceptance):
 hipfire diag
-hipfire pull qwen3.5:4b          # tag from registry/models.json; other tags OK if VRAM fits
+hipfire pull qwen3.5:4b          # tag from registry/models.json; other small tags OK if VRAM fits
 hipfire run qwen3.5:4b "Explain WMMA in one paragraph."
 ```
 
+**Acceptance / validation fixture** (dense claims, DFlash, promotion-shaped
+reports): registry tag `qwen3.8:27b-mq4-xt` → on-disk
+`~/.hipfire/models/qwen3.8-27b.mq4-xt`, with measured draft
+`~/qcal/ladder-v2/drafts/qwen3.8-27b-dflash.mq4v2.hfq` (see
+[`AGENTS.md`](../../../AGENTS.md) §5 pin). Do not treat a small-model smoke
+pass as acceptance evidence.
+
+```bash
+hipfire pull qwen3.8:27b-mq4-xt
+# optional measured-draft override:
+# export HIPFIRE_DFLASH_DRAFT=~/qcal/ladder-v2/drafts/qwen3.8-27b-dflash.mq4v2.hfq
+hipfire config set dflash_mode auto   # pull ≠ enable
+hipfire run qwen3.8:27b-mq4-xt "Explain WMMA in one paragraph."
+```
+
 If `diag` or first run fails, chain to `hipfire-diag` / `hipfire-autoheal`.
 
+
 ## Claim → harness (summary)
 
 Full map: [`docs/VALIDATION.md`](../../../docs/VALIDATION.md). **Branch by

.agents/skills/hipfire-tester/guide.md

@@ -72,10 +72,14 @@ hipfire pull <tag>    # exact tag from registry; check min_vram_gb vs free VRAM
 ```
 
 Do not hard-code a “standard matrix” of sizes. Pick one primary tag that fits
-the card and the claim (e.g. `qwen3.5:4b` for dense MQ4 smoke; an `lfm2.5:*`
+the card and the claim (e.g. `qwen3.5:4b` for **fast dense MQ4 smoke only**;
+acceptance / dense validation uses `qwen3.8:27b-mq4-xt` →
+`~/.hipfire/models/qwen3.8-27b.mq4-xt` + draft
+`~/qcal/ladder-v2/drafts/qwen3.8-27b-dflash.mq4v2.hfq`; an `lfm2.5:*`
 tag only for LFM routes). Confirm the on-disk file under `~/.hipfire/models/`
 (or the path you pass to harnesses).
 
+
 DFlash: pulling a draft does **not** enable speculation. Set
 `hipfire config set dflash_mode auto` (or per-model overlay) and confirm logs
 show the paired draft. Config authority: [`docs/CONFIG.md`](../../../docs/CONFIG.md).

.gitattributes

@@ -0,0 +1,2 @@
+# Preserve byte-identical benchmark prompts whose whitespace is part of the fixture.
+benchmarks/prompts/qwen38_issue693_longcode_20676.txt whitespace=-trailing-space