source-command-megaminx-rescue

name	source-command-megaminx-rescue
description	Run the top-N long-pid sym-ensemble rescue + merge cycle on the current best megaminx submission.

source-command-megaminx-rescue

Use this skill when the user asks to run the migrated source command megaminx-rescue.

Command Template

/megaminx-rescue

Automate the canonical "rescue the long-tail" cycle that we run after each new best submission. Extracts the top-N longest paths from the current best CSV, runs sym-ensemble + qshort + beam 524k on those pids, merges the rescued paths into the base, verifies the merged CSV, and prepares it for submission.

Usage

• /megaminx-rescue 50 — rescue top 50 long-pids of current best submission with K=4 sym (default).
• /megaminx-rescue 80 K=8 — rescue top 80 with K=8 (diminishing returns past K=4).
• /megaminx-rescue 50 base=path/to/sub.csv K=4 — explicit base CSV.

Defaults

• Base submission: latest in megaminx/submissions/ matching merge_*.csv (highest path count = lowest score).
• K (sym-ensemble): 4.
• Beam: 524288.
• Max-steps: 150.
• Teacher: megaminx/models/m05_bellman_warm/epoch_0499.pt.
• Student (qshort): megaminx/models/m23_v2_sym_aware/epoch_0499.pt (sym-aware, pair with --sym-ensemble).
• Output prefix: merge_plus_sym{K}_top{N}.
• Rotations: megaminx/data/rotations.npy (default; auto-loaded by --sym-ensemble).

Execution

1. Identify base submission and extract top-N pids

PYTHONUTF8=1 .venv/Scripts/python.exe -c "
import csv
sub = '<base>'
rows = list(csv.DictReader(open(sub)))
rows.sort(key=lambda r: -len(r['path'].split('.')))
print(','.join(r['initial_state_id'] for r in rows[:<N>]))
"

Report the pid list, length range of top-N, and sum of path lengths.

2. Launch the rescue solve (background)

nohup .venv/Scripts/python.exe megaminx/scripts/03_solve.py \
  --checkpoint megaminx/models/m05_bellman_warm/epoch_0499.pt \
  --qshort-student megaminx/models/m23_v2_sym_aware/epoch_0499.pt \
  --qshort-alpha 2 \
  --out megaminx/submissions/<output>.csv \
  --pids <pid_list> \
  --beams 524288 --max-steps 150 --bf16 \
  --sym-ensemble <K> \
  > megaminx/submissions/<output>.log 2>&1 &

3. Monitor with tight filter

# expected wall: K=2 ≈ 137s/pid, K=4 ≈ 280s/pid, K=8 ≈ 560s/pid
tail -F megaminx/submissions/<output>.log | grep -E --line-buffered \
  "attempts=20/|attempts=40/|attempts=60/|attempts=80/|Traceback|Error|verify:|stats:|wrote |OOM"

(Use Monitor tool with persistent: true.)

4. After completion: merge

.venv/Scripts/python.exe megaminx/scripts/16_merge_rescue.py \
  --base <base>.csv \
  --rescue megaminx/submissions/<output>.csv \
  --out megaminx/submissions/<merged>.csv

Report: replaced X/N (saved Y moves), new total, regressions count (must be 0).

5. Verify merged

The merge script verifies inline and prints verify: 1001/1001 valid, total <T>. If anything other than 1001/1001 valid: STOP, do not proceed to submit.

6. Tell user the result

Print: "ready to submit .csv (current_best - delta = new_total)". Suggest /megaminx-submit <merged>.csv "<description>" as next step.

DO NOT submit automatically. The user controls Kaggle slot usage.

Gotchas

• Pair m23_v2 (sym-aware) with --sym-ensemble, not the original m23 (distilled on forward-state V; lower recall on inverted/rotated states). See megaminx_gotchas.md "Q-shortlister student must match teacher" entry.
• K=4 is the sweet spot. K=8 saves ~80% as much per pid at 2× wall.
• Beam max-steps = 150 minimum (real path tail is ~99-103 in current submissions). Lower max-steps causes silent failures (model never converges).
• The merge takes per-pid min of base vs rescue. If rescue is uniformly worse than base for a pid, base is kept. 0 regressions is the expected outcome with K≥2.
• Rotation index 0 (identity) is always one of the K rotations selected.

Cost reference (4090 Laptop, beam 524k)

K	wall/pid	50 pids	80 pids
2	~140s	~2h	~3h
4	~280s	~4h	~6h
8	~560s	~8h	~12h

Pick K based on remaining session time and previous-pass exhaustion.