Stage ID is always 0 (NO_STAGE) in all parsed training data

[ScavieFae]

Bug

Every game in parsed-v2/games/ has stage=0 (NO_STAGE). 100% of training data — confirmed by sampling 200 random games.

Impact

The world model's stage embedding is useless. It receives the same stage=0 token for every game, so it can't distinguish between stages. This means:

• No platform awareness: Model sees players standing at y=27 (Battlefield platforms) and y=0 (FD ground) and has no signal for why. It learns a blurred average of all stage geometries.
• No stage-specific blast zones: Yoshi's Story blast zones are tiny (-175 to 173) vs Dreamland's are huge (-255 to 255). The model can't learn stage-specific KO thresholds.
• Randall noise: Yoshi's Story has a moving cloud platform (Randall). Without stage context, players standing on it look like they're hovering in mid-air.
• Platform tech patterns: BF/Dreamland tech chase patterns don't exist on FD, but the model mixes them all together.

Despite this, the model hits 97.6% action accuracy and 0.84 avg position error on teacher-forced evaluation — impressive, but position prediction in particular likely has a stage-confusion component.

Reproduction

from worldmodel.data.parse import load_game
from pathlib import Path
from collections import Counter
import random

games_dir = Path("~/claude-projects/nojohns-training/data/parsed-v2/games").expanduser()
all_games = [g for g in games_dir.iterdir() if not g.name.startswith('.')]
sample = random.sample(all_games, 200)
stage_counts = Counter()
for gf in sample:
    g = load_game(str(gf))
    stage_counts[g.stage] += 1

print(stage_counts)
# Counter({0: 200})  — 100% stage 0

Root cause

load_game() in worldmodel/data/parse.py:236 reads:

stage = root.field("stage")[0].as_py()

The data is genuinely 0 in the parquet files — the upstream Slippi parser that produced parsed-v2/ didn't extract the stage ID from the replay metadata. The stage field exists in the schema but was written as 0 for every game.

Fix

Need to either:

1. Fix the Slippi→parquet parser to extract game_start.stage from replay metadata, then re-parse all 22K+ games
2. Or backfill: read stage from raw .slp files and patch the existing parquet data

This is a re-parse of the full dataset, so it's a meaningful operation — but the fix itself should be straightforward once we find where stage extraction was dropped.

Discovery

Found during demo work — the viewer was showing "Stage 0" with fallback geometry for every game. Investigated expecting a viewer bug, turned out to be training data.