mtg/tools/card_stats.py

#!/usr/bin/env python3
"""Per-card statistics from Forge play-by-play logs.

Forge (without -q) logs every cast, zone change, and life change attributed to a
player. That lets us ask, per card:

  * how often was it CAST at all, when drawn into a game?
  * what turn did it typically come down?
  * did our deck WIN more when this card got cast?

    python3 tools/card_stats.py --deck decks/reaper-king-no-more decks/reaper-king-no-more/decks/<variant>/sim/<stamp>.log

The win-rate delta is the interesting column, but see the caveats it prints:
this is correlation over a small sample, not causation.

"Our seat" is matched by the deck's Forge metadata name (config.yaml, via
DeckContext) -- never a hardcoded deck name.
"""
import sys, os, re, json
from collections import defaultdict, Counter

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from cards import DB, load_decklist
from deckctx import DeckContext

OURS = "our deck"


def is_ours(name):
    """Match our seat even when -a <profile> renames it to 'NAME [Profile]'."""
    return name == OURS or name.startswith(OURS + " [")
CAST = re.compile(r"^Add To Stack: Ai\(\d+\)-(.+?) cast (.+?)(?: targeting .*)?$")
# In a 4-player pod Forge numbers EVERY player's turn sequentially, so "Turn 24"
# is round 6, not turn 24. Track the seat too so we can convert to rounds.
TURN = re.compile(r"^Turn: Turn (\d+) \(Ai\(\d+\)-(.+?)\)")
WON = re.compile(r"^Game Outcome: Ai\(\d+\)-(.+?) has won")
LOST = re.compile(r"^Game Outcome: Ai\(\d+\)-(.+?) has lost")
END = re.compile(r"^Game Result: Game (\d+) ended")


def parse_games(path):
    """Split the log into per-game records of what OUR deck cast, and who won."""
    games, cur = [], None
    turn = 0
    for line in open(path, errors="ignore"):
        line = line.rstrip("\n")
        if cur is None:
            cur = {"casts": [], "won": None, "turns": 0}
        m = TURN.match(line)
        if m:
            seq = int(m.group(1))
            n_players = cur.get("players") or 4
            turn = (seq + n_players - 1) // n_players     # -> round number
            cur["turns"] = max(cur["turns"], turn)
            continue
        m = CAST.match(line)
        if m and is_ours(m.group(1).strip()):
            cur["casts"].append((m.group(2).strip(), turn))
            continue
        m = WON.match(line)
        if m:
            cur["won"] = is_ours(m.group(1).strip())
            continue
        if END.match(line):
            games.append(cur)
            cur, turn = None, 0
    return games


def main():
    global OURS
    ctx = DeckContext.from_args()
    OURS = ctx.forge_metadata_name()
    args = sys.argv[1:]
    if not args:
        sys.exit(__doc__)
    log = args[0]
    games = parse_games(log)
    if not games:
        sys.exit("no games parsed -- was the log made without -q?")

    n = len(games)
    wins = sum(1 for g in games if g["won"])
    base = wins / n

    cast_games = defaultdict(int)     # games where card was cast >=1x
    cast_wins = defaultdict(int)
    turns = defaultdict(list)
    total_casts = Counter()

    for g in games:
        seen = set()
        for name, t in g["casts"]:
            total_casts[name] += 1
            if name not in seen:
                seen.add(name)
                cast_games[name] += 1
                if g["won"]:
                    cast_wins[name] += 1
            if t:
                turns[name].append(t)

    # only cards actually in our decklist
    db = DB(blacklist_path=ctx.blacklist_path)
    variant_dir = os.path.dirname(os.path.dirname(log))
    deck_file = os.path.join(variant_dir, "decklist.mtg")
    commander, deck = load_decklist(deck_file)
    deck_set = set(deck) | {commander}

    rows = []
    for name in deck_set:
        c = db.get(name)
        if c and c.is_land:
            continue
        g = cast_games.get(name, 0)
        wr = (cast_wins[name] / g) if g else None
        rows.append({
            "card": name,
            "games_cast": g,
            "cast_rate": g / n,
            "avg_turn": (sum(turns[name]) / len(turns[name])) if turns.get(name) else None,
            "win_rate_when_cast": wr,
            "delta": (wr - base) if wr is not None else None,
            "total_casts": total_casts.get(name, 0),
        })

    print("=" * 74)
    print(f"CARD STATISTICS — {os.path.basename(log)}")
    print("=" * 74)
    print(f"games: {n}   our wins: {wins} ({100*base:.1f}%)   deck: {deck_file}")
    print()

    never = sorted(r["card"] for r in rows if r["games_cast"] == 0)
    blocked = sorted(c for c in never if db.get(c) and db[c].ai_removed)
    playable = sorted(set(never) - set(blocked))

    print(f"NEVER CAST — AI-BLOCKED: {len(blocked)}   (simulator artifact, IGNORE)")
    print("-" * 74)
    print("  These carry `AI:RemoveDeck:All` in Forge's card scripts, so the engine")
    print("  is forbidden from ever playing them. Their absence says nothing about")
    print("  the card's value -- only that Forge cannot pilot it.")
    for c in blocked:
        print(f"    {c}")
    print()

    print(f"NEVER CAST — PLAYABLE: {len(playable)}   (real signal)")
    print("-" * 74)
    print("  Forge COULD have played these and didn't: too expensive, never drawn,")
    print("  or genuinely weak. These are worth investigating.")
    for c in playable:
        print(f"    {c}")
    print()

    played = [r for r in rows if r["games_cast"] > 0]
    played.sort(key=lambda r: -r["cast_rate"])
    print("MOST-CAST CARDS")
    print("-" * 74)
    print(f"  {'card':<34}{'cast in':>9}{'avg round':>11}{'win% when cast':>16}")
    for r in played[:20]:
        wt = f"{100*r['win_rate_when_cast']:.0f}%" if r["win_rate_when_cast"] is not None else "-"
        at = f"R{r['avg_turn']:.1f}" if r["avg_turn"] else "-"
        print(f"  {r['card'][:32]:<34}{r['games_cast']:>4}/{n:<4}{at:>11}{wt:>16}")
    print()

    print("CAVEATS — DO NOT CUT CARDS ON THIS ALONE")
    print("-" * 74)
    print(f"  * {n} games is a tiny sample. A card cast in 5 games has a win-rate")
    print("    swing of +/-40% from noise alone.")
    print("  * 'Never cast' conflates three very different things: a bad card, a")
    print("    card the AI cannot use, and a card that was simply never drawn.")
    print("  * Correlation only. Cards cast in LONG games look good because long")
    print("    games are ones we survived -- the card didn't cause the win.")
    print("  * Forge's AI is weak at combo; it under-casts combo pieces.")
    print()
    out = log.replace(".log", "-cards.json")
    json.dump({"log": log, "games": n, "base_win_rate": base,
               "cards": sorted(rows, key=lambda r: -r["cast_rate"])},
              open(out, "w"), indent=1)
    print(f"raw: {out}")


if __name__ == "__main__":
    main()