steps456 README: note the 2026-09-11 code/exp/pipeline fix and the step-5 frame rule (shipped shards only, missing frames stay pending)

steps456_20260910/README_steps456.md

@@ -25,6 +25,12 @@ Eligible items of every step = multiple-choice items kept after dedup that survi
 by any other step whose rows exist in `results/pool/` (yes/no items are skipped). Steps 5 and 6 need the frame
 cache of the item's video; items without frames are skipped and stay pending.
 
+Frames for step 5: only the shards listed in section 3 (the 16 main `frames_*.tar` plus the
+`supplement_20260910/frames/frames_sup_*.tar` shards, all in the same public repo) are needed; both sets must be
+unpacked into `frames/pool_frames/`. Do not fetch or decode frames from anywhere else: an item whose video has no
+directory there is skipped by the client and stays pending, and the `--plan-only` line after step 5 reports
+those as remaining. Report that remaining count to us; we run those items here.
+
 ## 2. Hardware and software
 
 Same as the main package: Python 3.12, `pip install -r code/requirements.txt` (vLLM 0.11.0, torch 2.8 CUDA 12.8),
@@ -32,6 +38,12 @@ models `Qwen/Qwen3-VL-8B-Instruct` (rev `0c351dd01ed87e9c1b53cbc748cba10e6187ff3
 `Qwen/Qwen3-VL-2B-Instruct` (rev `89644892e4d85e24eaac8bacfd4f463576704203`) from the Hub. One vLLM server per GPU.
 Disk: frames 36 GB + 14 GB (tar) unpacked to ~50 GB, results 1 GB, models 21 GB.
 
+**Fix 2026-09-11:** `code/exp/pipeline/` (chance.py, scoring.py, stage_d_blind.py) was missing from the first
+upload; `pool_steps.eligible` imports `chance._yes_no_options` (the project's yes/no rule) from it and failed with
+`ModuleNotFoundError: No module named 'chance'`. Re-download `steps456_20260910/code/` (or just that folder) and
+copy it over `pool/code/` again; the three files are verbatim copies of our pipeline, so the screening rule is
+unchanged. No other file of the package changed.
+
 ## 3. Get the data
 
 ```bash