swap_analysis_cvbench/swap_analysis.py

#!/usr/bin/env python3
"""
Swap Analysis: Minimal Pair Probing for Spatial Representations

Creates minimal pairs by swapping obj1<->obj2 in spatial questions:
  Original: "Is A to the left or right of B?" -> left
  Swapped:  "Is B to the left or right of A?" -> right

Supported model types
---------------------
  Legacy (Qwen2.5-VL-3B scale experiments):
    molmo | nvila | qwen
  Synthetic-MCQ-trained NVILA:
    nvila_st    : NVILA trained with MCQ synthetic mix (80k-st / 400k-st / 800k-st)
  New large models:
    molmo_big   : Molmo2-8B
    qwen_big    : Qwen3-VL-32B-Instruct
    qwen_super  : Qwen3-VL-235B-A22B-Instruct
    big_trio    : Molmo2-8B + RoboRefer + Qwen3-VL-32B
  Merge-only (--merge required):
    molmo_all        : molmo (vanilla→2m) + molmo_big (molmo2)
    qwen_all         : qwen  (vanilla→2m) + qwen_big  (qwen3_32b)
    nvila_st_compare : nvila (vanilla→2m) + nvila_st  (80k-st / 400k-st / 800k-st)

Usage examples
--------------
  # Legacy model (Qwen2.5-VL-3B scale)
  python swap_analysis.py --model_type qwen

  # New large model (Qwen3-VL-32B)
  conda run -n qwen3 python swap_analysis.py --model_type qwen_big

  # Cross-family merge (combine qwen + qwen_big results)
  conda run -n qwen3 python swap_analysis.py --model_type qwen_all --merge

Analyses:
  1. Difference vectors: delta = feature(swapped) - feature(original)
  2. Within-category delta consistency (do all left->right swaps point same direction?)
  3. Sign-corrected group consistency (align opposite categories by flipping)
  4. Cross-group delta alignment (delta_vertical vs delta_distance) for perspective bias
  5. Delta-based 6x6 similarity heatmap (mean delta per category as representation)
  6. Prediction stats visualization (bar chart + cross-scale trajectory)
  7. Both-correct filtering for delta analysis
  8. PCA visualization of per-sample embeddings
  9. Scaling effects on all of the above

Fixes applied:
  Fix 1: "Answer with only one word." appended to all prompts
  Fix 2: Synonym handling (below/beneath->under, near/nearby->close, distant->far)
  Fix 4: Cross-group quads index matching via string normalization
  Fix 5: Within-category + sign-corrected delta consistency (replaces wrong group-level)
  Fix 6: Prediction stats bar chart + cross-scale line plot
  Fix 7: Delta-based 6x6 heatmap and trajectory
  Fix 8: Category validity check + both-correct delta filtering
"""

import ast
import os
import sys
import json
import argparse
import base64
import logging
import random
import re
from io import BytesIO
from collections import defaultdict
from typing import Dict, List, Tuple, Optional, Any
from abc import ABC, abstractmethod

import torch
import numpy as np
import pandas as pd
from PIL import Image
from tqdm import tqdm
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D  # noqa: F401
import seaborn as sns
from sklearn.metrics.pairwise import cosine_similarity
from sklearn.decomposition import PCA

logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

_HERE = os.path.dirname(os.path.abspath(__file__))

# ── Local HuggingFace cache helpers ──────────────────────────────────────────

HF_HUB_DIR = '/data/shared/Qwen/mydisk/huggingface/hub'


def resolve_local_path(model_path: str) -> str:
    """Return local snapshot path for a HF model ID if cached, else return the ID unchanged."""
    if os.path.isabs(model_path):
        return model_path
    cache_name    = 'models--' + model_path.replace('/', '--')
    snapshots_dir = os.path.join(HF_HUB_DIR, cache_name, 'snapshots')
    if os.path.isdir(snapshots_dir):
        snapshots = sorted(os.listdir(snapshots_dir))
        if snapshots:
            local_path = os.path.join(snapshots_dir, snapshots[-1])
            logger.info(f"Local cache found: {model_path}  →  {local_path}")
            return local_path
    logger.warning(
        f"Model not found in local cache: '{model_path}'\n"
        f"  Expected at: {snapshots_dir}\n"
        f"  Will fall back to online HuggingFace Hub download.\n"
        f"  To cache locally first: python -c \"from huggingface_hub import snapshot_download; "
        f"snapshot_download('{model_path}', cache_dir='{HF_HUB_DIR}')\""
    )
    return model_path


def _setup_file_logging(name: str, log_dir: str) -> str:
    """Attach a named FileHandler to the root logger.

    Writes to  {log_dir}/{name}.log  (append mode).
    Returns the log file path.
    """
    os.makedirs(log_dir, exist_ok=True)
    log_path = os.path.join(log_dir, f'{name}.log')
    fh = logging.FileHandler(log_path, mode='a', encoding='utf-8')
    fh.setLevel(logging.INFO)
    fh.setFormatter(logging.Formatter('%(asctime)s - %(levelname)s - %(message)s'))
    logging.getLogger().addHandler(fh)
    return log_path


def get_model_key(model_type: str, scale: str) -> str:
    """Return VLM key for (model_type, scale). E.g. nvila_80k, nvila_synthetic_80k-5pct."""
    return f"{model_type}_{scale}"


# ============================================================================
# Constants
# ============================================================================

CATEGORY_ORDER = ['left', 'right', 'above', 'below', 'far', 'close']

OPPOSITE_MAP = {
    'left': 'right', 'right': 'left',
    'above': 'below', 'below': 'above',
    'under': 'above',                   # short-mode vertical answer
    'far': 'close', 'close': 'far',
}

# Opposite map for short-answer mode (vertical uses 'above'/'under', not 'above'/'below')
SHORT_OPPOSITE_MAP = {
    'left': 'right', 'right': 'left',
    'above': 'below', 'below': 'above',
    'far': 'close', 'close': 'far',
}

GROUP_MAP = {
    'left': 'horizontal', 'right': 'horizontal',
    'above': 'vertical', 'below': 'vertical',
    'far': 'distance', 'close': 'distance',
}

GROUP_ORDER = ['horizontal', 'vertical', 'distance']

# Fix 5: Canonical categories for sign-corrected consistency
CANONICAL_CATEGORIES = {
    'horizontal': 'left',
    'vertical': 'above',
    'distance': 'far',
}

# Fix 2: Synonyms for answer matching
# 'below' is now primary; 'under'/'beneath' recognized as synonyms
SYNONYMS = {
    'below': ['under', 'beneath'],
    'close': ['near', 'nearby'],
    'far': ['distant'],
}

# ── MCQ question templates (option order alternated per pair for A/B bias control) ──
_Q_TAIL_MCQ = "Answer with a single letter A or B."
MCQ_TEMPLATES = {
    'horizontal': {
        'left_first':  "Is the {obj1} to the left or right of the {obj2}? (A) left  (B) right  " + _Q_TAIL_MCQ,
        'right_first': "Is the {obj1} to the left or right of the {obj2}? (A) right  (B) left  " + _Q_TAIL_MCQ,
    },
    'vertical': {
        'above_first': "Is the {obj1} above or below the {obj2}? (A) above  (B) below  " + _Q_TAIL_MCQ,
        'below_first': "Is the {obj1} above or below the {obj2}? (A) below  (B) above  " + _Q_TAIL_MCQ,
    },
    'distance': {
        'far_first':   "Compared to {ref}, is {subj} far or close from you? (A) far  (B) close  " + _Q_TAIL_MCQ,
        'close_first': "Compared to {ref}, is {subj} far or close from you? (A) close  (B) far  " + _Q_TAIL_MCQ,
    },
}
MCQ_LETTER = {
    'horizontal': {
        'left_first':  {'left': 'a', 'right': 'b'},
        'right_first': {'left': 'b', 'right': 'a'},
    },
    'vertical': {
        'above_first': {'above': 'a', 'below': 'b'},
        'below_first': {'above': 'b', 'below': 'a'},
    },
    'distance': {
        'far_first':   {'far': 'a', 'close': 'b'},
        'close_first': {'far': 'b', 'close': 'a'},
    },
}

SCALE_COLORS = {
    'vanilla':    '#1f77b4', '80k':      '#ff7f0e', '400k':     '#2ca02c',
    '800k':       '#d62728', '2m':       '#9467bd', 'roborefer':'#8c564b',
    # New large models
    'molmo2':     '#17becf',   # cyan
    'qwen3_32b':  '#bcbd22',   # yellow-green
    'qwen3_235b': '#e377c2',   # pink
    # Synthetic-mix NVILA at 80k scale (shades of teal, light→dark by mix ratio)
    '80k-5pct':   '#b2dfdb',   # very light teal
    '80k-10pct':  '#00b894',   # teal
    '80k-20pct':  '#00897b',   # darker teal
    '80k-30pct':  '#004d40',   # deep teal
    # Synthetic-mix NVILA at 400k scale
    '400k-5pct':  '#66bb6a',   # light green (near 400k's #2ca02c)
    # Synthetic-mix NVILA at 800k scale
    '800k-5pct':  '#ef9a9a',   # light red (near 800k's #d62728)
    # MCQ-synthetic-trained NVILA (darker shade of the matching base scale)
    '80k-st':     '#b85a00',   # dark orange  (darker than 80k  #ff7f0e)
    '400k-st':    '#1a6b1a',   # dark green   (darker than 400k #2ca02c)
    '800k-st':    '#911b1b',   # dark red     (darker than 800k #d62728)
}

# Canonical scale ordering used by accuracy/ylim plots (add new scales here to control x-axis)
SCALE_ORDER = [
    'vanilla', '80k', '80k-5pct', '80k-10pct', '80k-20pct', '80k-30pct', '80k-st',
    '400k', '400k-5pct', '400k-st', '800k', '800k-5pct', '800k-st', '2m', 'roborefer',
    'molmo2', 'qwen3_32b', 'qwen3_235b',
]

# Human-readable legend labels (only entries that differ from the key are needed)
SCALE_DISPLAY_NAMES = {
    '80k-5pct':  '80k 5%',
    '80k-10pct': '80k 10%',
    '80k-20pct': '80k 20%',
    '80k-30pct': '80k 30%',
    '400k-5pct': '400k 5%',
    '800k-5pct': '800k 5%',
    '80k-st':    '80k ST',
    '400k-st':   '400k ST',
    '800k-st':   '800k ST',
}
# Category colors aligned with group: horizontal=orange, vertical=green, distance=purple
CAT_COLORS = {
    'left':  '#ff7f0e', 'right':  '#ffbb78',  # horizontal → orange
    'above': '#2ca02c', 'below':  '#98df8a',  # vertical   → green
    'far':   '#9467bd', 'close':  '#c5b0d5',  # distance   → purple
}
GROUP_COLORS = {
    'horizontal': '#ff7f0e',
    'vertical':   '#2ca02c',
    'distance':   '#9467bd',
}

# Short-answer (non-MCQ) question templates
SHORT_TEMPLATES = {
    'horizontal': "Is the {obj1} to the left or right of the {obj2}? Answer with only one word.",
    'vertical':   "Is the {obj1} above or below the {obj2}? Answer with only one word.",
    'distance':   "Compared to {ref}, is {subj} far or close from you? Answer with only one word.",
}

MODEL_CONFIGS = {
    'molmo': {
        'vanilla': 'allenai/Molmo-7B-O-0924',
        '80k': '/data/shared/Qwen/molmo/outputs/data_scale_exp/data_scale_exp_80k/unshared',
        '400k': '/data/shared/Qwen/molmo/outputs/data_scale_exp/data_scale_exp_400k/unshared',
        '800k': '/data/shared/Qwen/molmo/outputs/data_scale_exp/data_scale_exp_800k/unshared',
        '2m': '/data/shared/Qwen/molmo/outputs/data_scale_exp/data_scale_exp_2m/unshared',
    },
    'nvila': {
        'vanilla': '/data/shared/Qwen/mydisk/NVILA-Lite-2B',
        '80k': '/data/shared/Qwen/mydisk/output/DATA/NVILA-Lite-2B-DATA_SCALE_EXP_80K-20251108_180221',
        '400k': '/data/shared/Qwen/mydisk/output/DATA/NVILA-Lite-2B-DATA_SCALE_EXP_400K-20251108_180221',
        '800k': '/data/shared/Qwen/mydisk/output/DATA/NVILA-Lite-2B-DATA_SCALE_EXP_800K-20251108_180221',
        '2m': '/data/shared/Qwen/mydisk/output/DATA/NVILA-Lite-2B-DATA_SCALE_EXP_2M-20260205_003632',
        # '80k': '/data/shared/Qwen/mydisk/output/SINGLE/NVILA-Lite-2B-SINGLE_REFSPATIAL_16M-20260217_035008/checkpoint-1250',
        # '400k': '/data/shared/Qwen/mydisk/output/SINGLE/NVILA-Lite-2B-SINGLE_REFSPATIAL_16M-20260217_035008/checkpoint-6250',
        # '800k': '/data/shared/Qwen/mydisk/output/SINGLE/NVILA-Lite-2B-SINGLE_REFSPATIAL_16M-20260217_035008/checkpoint-12500',
        # '2m': '/data/shared/Qwen/mydisk/output/SINGLE/NVILA-Lite-2B-SINGLE_REFSPATIAL_16M-20260217_035008/checkpoint-31250',        
        'roborefer': '/data/shared/Qwen/mydisk/RoboRefer_model',
        'roborefer_depth': '/data/shared/Qwen/mydisk/RoboRefer_depth_model',  # fill in actual path
    },
    'qwen': {
        'vanilla': 'Qwen/Qwen2.5-VL-3B-Instruct',
        '80k': '/data/shared/Qwen/mydisk/output/Qwen/data_scale_exp/Qwen2.5-VL-3B-Instruct-data_scale_exp_80k-20251114_120221',
        '400k': '/data/shared/Qwen/mydisk/output/Qwen/data_scale_exp/Qwen2.5-VL-3B-Instruct-data_scale_exp_400k-20251114_120221',
        '800k': '/data/shared/Qwen/mydisk/output/Qwen/data_scale_exp/Qwen2.5-VL-3B-Instruct-data_scale_exp_800k-20251114_120221',
        '2m': '/data/shared/Qwen/mydisk/output/Qwen/data_scale_exp/Qwen2.5-VL-3B-Instruct-data_scale_exp_2m-20260109_120517',
    },
    # NVILA trained with MCQ synthetic data mix (checkpoints at 80k/400k/800k steps)
    'nvila_st': {
        '80k-st':  '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_MCQ_5PCT_2M-20260302_030354/checkpoint-1250',
        '400k-st': '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_MCQ_5PCT_2M-20260302_030354/checkpoint-6250',
        '800k-st': '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_MCQ_5PCT_2M-20260302_030354/checkpoint-12500',
    },
    # NVILA trained with synthetic data mixed in at different ratios
    'nvila_synthetic': {
        '80k-5pct':  '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_5PCT_2M-20260226_023301/checkpoint-1250',
        '80k-10pct': '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_10PCT_80K-20260224_234537',
        '80k-20pct': '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_20PCT_80K-20260224_232347',
        '80k-30pct': '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_30PCT_80K-20260224_232347',
        '400k-5pct': '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_5PCT_2M-20260226_023301/checkpoint-6250',
        '800k-5pct': '/data/shared/Qwen/mydisk/output/SYNTHETIC/NVILA-Lite-2B-SYNTHETIC_MIX_5PCT_2M-20260226_023301/checkpoint-12500'
    },
}

# ── New large / cross-family models ──────────────────────────────────────────
# Each scale maps to (ExtractorClassName, HF-model-ID-or-absolute-path).
# resolve_local_path() converts HF IDs to local snapshot dirs when cached.
MODEL_CONFIGS_NEW = {
    'molmo_big': {
        'molmo2': ('Molmo2Extractor', 'allenai/Molmo2-8B'),
    },
    'qwen_big': {
        'qwen3_32b': ('Qwen3VLExtractor', 'Qwen/Qwen3-VL-32B-Instruct'),
    },
    'qwen_super': {
        'qwen3_235b': ('Qwen3VLExtractor', 'Qwen/Qwen3-VL-235B-A22B-Instruct'),
    },
    'big_trio': {
        'molmo2':    ('Molmo2Extractor',    'allenai/Molmo2-8B'),
        'roborefer': ('RoboReferExtractor', '/data/shared/Qwen/mydisk/RoboRefer_model'),
        'qwen3_32b': ('Qwen3VLExtractor',   'Qwen/Qwen3-VL-32B-Instruct'),
    },
}

# ── Merge-only: combine existing per-scale data from multiple source dirs ─────
MERGE_ONLY_CONFIGS = {
    'molmo_all': {
        'scale_order': ['vanilla', '80k', '400k', '800k', '2m', 'molmo2'],
        'scale_sources': {
            'vanilla': 'molmo', '80k': 'molmo', '400k': 'molmo',
            '800k': 'molmo', '2m': 'molmo', 'molmo2': 'molmo_big',
        },
        'required_dirs': ['molmo', 'molmo_big'],
    },
    'qwen_all': {
        'scale_order': ['vanilla', '80k', '400k', '800k', '2m', 'qwen3_32b'],
        'scale_sources': {
            'vanilla': 'qwen', '80k': 'qwen', '400k': 'qwen',
            '800k': 'qwen', '2m': 'qwen', 'qwen3_32b': 'qwen_big',
        },
        'required_dirs': ['qwen', 'qwen_big'],
    },
    # Compare NVILA baselines against MCQ-synthetic-trained checkpoints
    'nvila_st_compare': {
        'scale_order': ['vanilla', '80k', '80k-st', '400k', '400k-st', '800k', '800k-st', '2m'],
        'scale_sources': {
            'vanilla':  'nvila',
            '80k':      'nvila',    '80k-st':  'nvila_st',
            '400k':     'nvila',    '400k-st': 'nvila_st',
            '800k':     'nvila',    '800k-st': 'nvila_st',
            '2m':       'nvila',
        },
        'required_dirs': ['nvila', 'nvila_st'],
    },
    # Compare NVILA baselines against synthetic-mix checkpoints
    'nvila_synth_compare': {
        'scale_order': ['vanilla', '80k', '80k-5pct', '80k-10pct', '400k', '400k-5pct', '800k', '800k-5pct'],
        'scale_sources': {
            'vanilla':   'nvila', 
            '80k':       'nvila',
            '80k-5pct':  'nvila_synthetic', 
            '80k-10pct': 'nvila_synthetic',
            '400k':      'nvila',
            '400k-5pct': 'nvila_synthetic',
            '800k':      'nvila',
            '800k-5pct': 'nvila_synthetic'
        },
        'required_dirs': ['nvila', 'nvila_synthetic'],
    },
}

# Default scale run order for new runnable types
SCALE_ORDERS_NEW = {
    'molmo_big':  ['molmo2'],
    'qwen_big':   ['qwen3_32b'],
    'qwen_super': ['qwen3_235b'],
    'big_trio':   ['molmo2', 'roborefer', 'qwen3_32b'],
}

ALL_MODEL_TYPES = (
    list(MODEL_CONFIGS.keys()) +
    list(MODEL_CONFIGS_NEW.keys()) +
    list(MERGE_ONLY_CONFIGS.keys())
)


# ============================================================================
# Data Loading & Swap Pair Creation
# ============================================================================

# ── CV-Bench data paths ───────────────────────────────────────────────────────
CV_BENCH_2D_TSV = '/data/shared/Qwen/vlm-lens-mod/spatial/CV-Bench/CV-Bench-2D-Relation-binary.tsv'
CV_BENCH_3D_TSV = '/data/shared/Qwen/vlm-lens-mod/spatial/CV-Bench/CV-Bench-3D.tsv'

# Maps 2D relation strings to category keys
RELATION_TO_CATEGORY = {
    'above': 'above', 'below': 'below',
    'to the left of': 'left', 'to the right of': 'right',
    'left': 'left', 'right': 'right',
}

# Regex to extract obj1, obj2 from CV-Bench 2D question text
_CV2D_OBJ_PATTERN = re.compile(
    r'relative positions? of (.+?) and (.+?) in the image',
    re.IGNORECASE,
)
_LEADING_THE = re.compile(r'^the\s+', re.IGNORECASE)

OBJECT_PATTERNS = [
    re.compile(r'between\s+(.+?)\s+and\s+(.+?)\s+in', re.IGNORECASE),
    re.compile(r'of\s+(.+?)\s+and\s+(.+?)\s+in', re.IGNORECASE),
    re.compile(r'positions\s+of\s+(.+?)\s+and\s+(.+?)\s+interact', re.IGNORECASE),
    re.compile(r'How\s+are\s+(.+?)\s+and\s+(.+?)\s+positioned', re.IGNORECASE),
    re.compile(r'arrangement\s+of\s+(.+?)\s+and\s+(.+?)\s+in', re.IGNORECASE),
]


def extract_objects(question: str) -> Tuple[str, str]:
    for pattern in OBJECT_PATTERNS:
        m = pattern.search(question)
        if m:
            return m.group(1).strip(), m.group(2).strip()
    raise ValueError(f"Could not extract objects from: {question}")


def decode_base64_image(base64_str: str) -> Image.Image:
    image_data = base64.b64decode(base64_str)
    return Image.open(BytesIO(image_data)).convert('RGB')


# ============================================================================
# Answer Matching (Fix 2: synonym support)
# ============================================================================

def find_earliest_position(text: str, word: str) -> int:
    """Find earliest position of word or any of its synonyms in text."""
    positions = []
    pos = text.find(word)
    if pos != -1:
        positions.append(pos)
    for syn in SYNONYMS.get(word, []):
        pos = text.find(syn)
        if pos != -1:
            positions.append(pos)
    return min(positions) if positions else -1


def check_answer(generated_text: str, expected_category: str, mcq_map: dict = None) -> bool:
    if not generated_text or not generated_text.strip():
        return False
    text = generated_text.strip().lower()
    expected = expected_category.lower()
    opposite = OPPOSITE_MAP[expected]

    if mcq_map:
        exp_letter = mcq_map.get(expected)
        opp_letter = mcq_map.get(opposite)
        # Standalone letter response (e.g. "A", "A.", "A)", "B")
        if exp_letter and text in (exp_letter, exp_letter+'.', exp_letter+')', exp_letter+','):
            return True
        if opp_letter and text in (opp_letter, opp_letter+'.', opp_letter+')', opp_letter+','):
            return False
    else:
        exp_letter = opp_letter = None

    # MCQ inline pattern "(a)"/"(b)" — variant-aware
    mcq_exp = f'({exp_letter})' if exp_letter else None
    mcq_opp = f'({opp_letter})' if opp_letter else None

    def earliest_with_mcq(word, mcq_pat=None):
        positions = []
        pos = text.find(word)
        if pos != -1:
            positions.append(pos)
        for syn in SYNONYMS.get(word, []):
            pos = text.find(syn)
            if pos != -1:
                positions.append(pos)
        if mcq_pat:
            pos = text.find(mcq_pat)
            if pos != -1:
                positions.append(pos)
        return min(positions) if positions else -1

    pos_exp = earliest_with_mcq(expected, mcq_exp)
    pos_opp = earliest_with_mcq(opposite, mcq_opp)
    if pos_exp == -1:
        return False
    if pos_opp == -1:
        return True
    return pos_exp < pos_opp


# ============================================================================
# Swap Pair Loading (Fix 1: prompt suffix)
# ============================================================================

def load_swap_pairs(tsv_path: str, seed: int = 42, filter_unknown: bool = True,
                    question_type: str = 'mcq') -> List[dict]:
    """Load EmbSpatialBench TSV and create swap pairs for all samples.

    Args:
        filter_unknown: If True (default), skip far/close pairs where target_object
            is Unknown/empty, and remove Unknown/empty values from reference_object
            candidates before sampling. Pairs with no valid candidates are dropped.
            Use --no-filtering to disable.
        question_type: 'short_answer' (default) uses the original "Answer with only one word." format;
            'mcq' uses MCQ A/B templates with letter answers.
    """
    rng = random.Random(seed)
    df = pd.read_csv(tsv_path, sep='\t')

    pairs = []
    stats = defaultdict(lambda: {'total': 0, 'success': 0})

    def _valid_obj(v):
        return bool(v) and str(v).strip().lower() not in ('unknown', 'n/a', '')

    for _, row in df.iterrows():
        category = row['category']
        stats[category]['total'] += 1

        try:
            if category in ['left', 'right', 'above', 'under', 'below']:
                obj1, obj2 = extract_objects(row['question'])
                if category in ['left', 'right']:
                    grp = 'horizontal'
                else:
                    grp = 'vertical'

                if question_type == 'short_answer':
                    # Single-word format; normalize 'under' → 'below'
                    if category == 'under':
                        category = 'below'
                    tmpl = SHORT_TEMPLATES[grp]
                    pair = {
                        'index': row['index'],
                        'question_id': str(row['question_id']),
                        'image_base64': row['image'],
                        'original_question': tmpl.format(obj1=obj1, obj2=obj2),
                        'swapped_question': tmpl.format(obj1=obj2, obj2=obj1),
                        'original_answer': category,
                        'swapped_answer': SHORT_OPPOSITE_MAP[category],
                        'group': grp,
                        'category': category,
                        'obj1': obj1, 'obj2': obj2,
                        'mcq_map': None,
                    }
                else:
                    # MCQ format; normalize 'under' → 'below'
                    if category == 'under':
                        category = 'below'
                    variant = ('left_first'  if grp == 'horizontal' else 'above_first') \
                              if len(pairs) % 2 == 0 else \
                              ('right_first' if grp == 'horizontal' else 'below_first')
                    tmpl = MCQ_TEMPLATES[grp][variant]
                    mcq_map = MCQ_LETTER[grp][variant]
                    pair = {
                        'index': row['index'],
                        'question_id': str(row['question_id']),
                        'image_base64': row['image'],
                        'original_question': tmpl.format(obj1=obj1, obj2=obj2),
                        'swapped_question': tmpl.format(obj1=obj2, obj2=obj1),
                        'original_answer': category,
                        'swapped_answer': OPPOSITE_MAP[category],
                        'group': GROUP_MAP[category],
                        'category': category,
                        'obj1': obj1, 'obj2': obj2,
                        'mcq_map': mcq_map,
                    }

            elif category in ['far', 'close']:
                answer_key = row['answer']
                options = {k: row[k] for k in ['A', 'B', 'C', 'D']}
                target_object = options[answer_key]
                candidates = [v for k, v in options.items() if k != answer_key]

                if filter_unknown:
                    if not _valid_obj(target_object):
                        continue
                    candidates = [v for v in candidates if _valid_obj(v)]
                    if not candidates:
                        continue

                reference_object = rng.choice(candidates)

                if question_type == 'short_answer':
                    tmpl = SHORT_TEMPLATES['distance']
                    pair = {
                        'index': row['index'],
                        'question_id': str(row['question_id']),
                        'image_base64': row['image'],
                        'original_question': tmpl.format(ref=reference_object, subj=target_object),
                        'swapped_question': tmpl.format(ref=target_object, subj=reference_object),
                        'original_answer': category,
                        'swapped_answer': OPPOSITE_MAP[category],
                        'group': 'distance',
                        'category': category,
                        'target_object': target_object,
                        'reference_object': reference_object,
                        'mcq_map': None,
                    }
                else:
                    variant = 'far_first' if len(pairs) % 2 == 0 else 'close_first'
                    tmpl = MCQ_TEMPLATES['distance'][variant]
                    mcq_map = MCQ_LETTER['distance'][variant]
                    pair = {
                        'index': row['index'],
                        'question_id': str(row['question_id']),
                        'image_base64': row['image'],
                        'original_question': tmpl.format(ref=reference_object, subj=target_object),
                        'swapped_question': tmpl.format(ref=target_object, subj=reference_object),
                        'original_answer': category,
                        'swapped_answer': OPPOSITE_MAP[category],
                        'group': 'distance',
                        'category': category,
                        'target_object': target_object,
                        'reference_object': reference_object,
                        'mcq_map': mcq_map,
                    }
            else:
                continue

            pairs.append(pair)
            stats[category]['success'] += 1

        except Exception as e:
            logger.warning(f"Failed to create swap pair for index {row['index']}: {e}")
            continue

    logger.info("Swap pair creation stats:")
    for cat in CATEGORY_ORDER:
        s = stats[cat]
        logger.info(f"  {cat}: {s['success']}/{s['total']}")
    logger.info(f"  Total pairs: {len(pairs)}")

    return pairs


# ============================================================================
# CV-Bench Data Loading
# ============================================================================

def load_cv_bench_pairs(
    tsv_2d_path: str,
    tsv_3d_path: str,
    seed: int = 42,
    question_type: str = 'short_answer',
) -> List[dict]:
    """Load CV-Bench 2D/3D TSVs and create swap pairs.

    2D (Relation): horizontal (left/right) and vertical (above/below) categories.
      - Uses the same SHORT_TEMPLATES / MCQ_TEMPLATES as the original EmbSpatialBench loader.
    3D (Depth): distance (far/close) category.
      - Always uses SHORT_TEMPLATES['distance'] (short-answer format, mcq_map=None).
      - obj_A (red box) = subj, obj_B (blue box) = ref.
      - answer='A' → obj_A closer → category='close'; answer='B' → category='far'.
    """
    rng = random.Random(seed)
    pairs = []
    stats = defaultdict(lambda: {'total': 0, 'success': 0})

    # ── 2D pairs ──────────────────────────────────────────────────────────────
    df2d = pd.read_csv(tsv_2d_path, sep='\t')
    for idx, row in df2d.iterrows():
        actual_relation = (
            row['original_A'] if str(row['original_answer']).strip() == 'A'
            else row['original_B']
        )
        category = RELATION_TO_CATEGORY.get(str(actual_relation).strip().lower())
        if category is None:
            continue  # skip containment or unknown relations

        grp = GROUP_MAP[category]
        stats[category]['total'] += 1

        try:
            m = _CV2D_OBJ_PATTERN.search(str(row['question']))
            if not m:
                continue
            # Strip leading "the " to avoid "Is the the wall" duplication
            obj1 = _LEADING_THE.sub('', m.group(1).strip())
            obj2 = _LEADING_THE.sub('', m.group(2).strip())

            if question_type == 'short_answer':
                tmpl = SHORT_TEMPLATES[grp]
                pair = {
                    'index': idx,
                    'image_base64': row['image'],
                    'original_question': tmpl.format(obj1=obj1, obj2=obj2),
                    'swapped_question': tmpl.format(obj1=obj2, obj2=obj1),
                    'original_answer': category,
                    'swapped_answer': SHORT_OPPOSITE_MAP[category],
                    'group': grp,
                    'category': category,
                    'obj1': obj1, 'obj2': obj2,
                    'mcq_map': None,
                }
            else:  # mcq
                variant = (
                    ('left_first' if grp == 'horizontal' else 'above_first')
                    if len(pairs) % 2 == 0 else
                    ('right_first' if grp == 'horizontal' else 'below_first')
                )
                tmpl = MCQ_TEMPLATES[grp][variant]
                mcq_map = MCQ_LETTER[grp][variant]
                pair = {
                    'index': idx,
                    'image_base64': row['image'],
                    'original_question': tmpl.format(obj1=obj1, obj2=obj2),
                    'swapped_question': tmpl.format(obj1=obj2, obj2=obj1),
                    'original_answer': category,
                    'swapped_answer': OPPOSITE_MAP[category],
                    'group': grp,
                    'category': category,
                    'obj1': obj1, 'obj2': obj2,
                    'mcq_map': mcq_map,
                }

            pairs.append(pair)
            stats[category]['success'] += 1

        except Exception as e:
            logger.warning(f"[2D] Failed row {idx}: {e}")
            continue

    # ── 3D pairs (always short_answer using distance template) ────────────────
    df3d = pd.read_csv(tsv_3d_path, sep='\t')
    tmpl_dist = SHORT_TEMPLATES['distance']

    for idx, row in df3d.iterrows():
        obj_A = str(row['A']).strip()  # red box object
        obj_B = str(row['B']).strip()  # blue box object
        actual_answer = str(row['answer']).strip()

        # obj_A is the subject being asked about (subj), obj_B is the reference
        # answer='A' → obj_A is closer → subj is 'close' relative to ref
        # answer='B' → obj_B is closer → subj (obj_A) is 'far' relative to ref
        if actual_answer == 'A':
            category = 'close'
        elif actual_answer == 'B':
            category = 'far'
        else:
            continue

        stats[category]['total'] += 1

        try:
            pair = {
                'index': idx,
                'image_base64': row['image'],
                'original_question': tmpl_dist.format(ref=obj_B, subj=obj_A),
                'swapped_question': tmpl_dist.format(ref=obj_A, subj=obj_B),
                'original_answer': category,
                'swapped_answer': OPPOSITE_MAP[category],
                'group': 'distance',
                'category': category,
                'obj1': obj_A,   # subj in original question
                'obj2': obj_B,   # ref in original question
                'mcq_map': None,
                'bbox_str': row.get('bbox', None),
            }
            pairs.append(pair)
            stats[category]['success'] += 1

        except Exception as e:
            logger.warning(f"[3D] Failed row {idx}: {e}")
            continue

    logger.info("CV-Bench swap pair creation stats (before balancing):")
    for cat in CATEGORY_ORDER:
        s = stats[cat]
        if s['total'] > 0:
            logger.info(f"  {cat}: {s['success']}/{s['total']}")
    logger.info(f"  Total pairs (before balancing): {len(pairs)}")

    # ── Balance categories: downsample to minimum category count ─────────────
    grouped = defaultdict(list)
    for p in pairs:
        grouped[p['category']].append(p)

    cat_counts = {cat: len(grouped[cat]) for cat in CATEGORY_ORDER if grouped[cat]}
    if cat_counts:
        min_count = min(cat_counts.values())
        logger.info(f"  Balancing categories to {min_count} samples each "
                    f"(min category: {min(cat_counts, key=cat_counts.get)})")
        balanced = []
        for cat in CATEGORY_ORDER:
            samples = grouped[cat]
            if len(samples) > min_count:
                samples = rng.sample(samples, min_count)
            balanced.extend(samples)
        pairs = balanced

    logger.info("CV-Bench swap pair stats (after balancing):")
    for cat in CATEGORY_ORDER:
        n = sum(1 for p in pairs if p['category'] == cat)
        if n > 0:
            logger.info(f"  {cat}: {n}")
    logger.info(f"  Total pairs: {len(pairs)}")

    return pairs


# ============================================================================
# CV-Bench Bbox Helpers
# ============================================================================

def parse_cv_bench_bbox(bbox_str) -> Optional[List]:
    """Parse CV-Bench-3D bbox string: '[[x0,y0,x1,y1],[x0,y0,x1,y1]]' → list."""
    if bbox_str is None:
        return None
    try:
        if isinstance(bbox_str, float):
            return None
    except Exception:
        pass
    try:
        if pd.isna(bbox_str):
            return None
    except Exception:
        pass
    try:
        return ast.literal_eval(str(bbox_str))
    except Exception:
        return None


def get_cvbench_bbox_center_y(bbox_list: List, obj_idx: int) -> float:
    """Return vertical center of bbox.  obj_idx: 0=A(red), 1=B(blue).

    bbox format: [x0, y0, x1, y1] in absolute pixels.
    """
    bbox = bbox_list[obj_idx]
    return (bbox[1] + bbox[3]) / 2


# ============================================================================
# CV-Bench Cross-Group Quads
# ============================================================================

def create_cv_bench_cross_group_quads(
    depth_pairs: List[dict],
    threshold_ratio: float = 0.05,
) -> List[dict]:
    """For CV-Bench depth (distance) swap pairs, build cross-group quads by adding
    a vertical (above/below) question derived from bbox center-y positions.

    Both distance and vertical questions always use short-answer templates (mcq_map=None).
    """
    quads = []
    stats = {'total': 0, 'added': 0, 'ambiguous': 0, 'no_bbox': 0}

    vert_tmpl = SHORT_TEMPLATES['vertical']

    for pair in depth_pairs:
        stats['total'] += 1
        bbox_list = parse_cv_bench_bbox(pair.get('bbox_str'))
        if bbox_list is None or len(bbox_list) < 2:
            stats['no_bbox'] += 1
            continue

        center_y_A = get_cvbench_bbox_center_y(bbox_list, 0)  # obj1 = subj (red box)
        center_y_B = get_cvbench_bbox_center_y(bbox_list, 1)  # obj2 = ref  (blue box)

        # Estimate image height from bboxes; floor at 1024
        est_height = max(bbox_list[0][3], bbox_list[1][3]) * 1.1
        est_height = max(est_height, 1024)
        threshold = est_height * threshold_ratio

        y_diff = center_y_A - center_y_B
        if abs(y_diff) <= threshold:
            stats['ambiguous'] += 1
            continue

        obj_A = pair['obj1']  # subj in original distance question
        obj_B = pair['obj2']  # ref  in original distance question

        # Vertical label for obj_A relative to obj_B (smaller y = higher in image = 'above')
        vert_original_answer = 'above' if center_y_A < center_y_B else 'below'
        vert_swapped_answer = SHORT_OPPOSITE_MAP[vert_original_answer]

        quad = {
            'index': pair['index'],
            'image_base64': pair['image_base64'],
            'dist_original_q': pair['original_question'],
            'dist_swapped_q': pair['swapped_question'],
            'dist_original_answer': pair['original_answer'],
            'dist_swapped_answer': pair['swapped_answer'],
            'dist_mcq_map': None,
            'vert_original_q': vert_tmpl.format(obj1=obj_A, obj2=obj_B),
            'vert_swapped_q': vert_tmpl.format(obj1=obj_B, obj2=obj_A),
            'vert_original_answer': vert_original_answer,
            'vert_swapped_answer': vert_swapped_answer,
            'vert_mcq_map': None,
            'target_object': obj_A,
            'reference_object': obj_B,
            'target_bbox_y': center_y_A,
            'ref_bbox_y': center_y_B,
            'y_diff': y_diff,
            'data_source': 'cv_bench_3d',
        }
        quads.append(quad)
        stats['added'] += 1

    logger.info(f"CV-Bench cross-group quads: {stats['added']}/{stats['total']} "
                f"(ambiguous={stats['ambiguous']}, no_bbox={stats['no_bbox']})")
    return quads


# ============================================================================
# HF Bbox Cache (Fix 4: string-normalized keys)
# ============================================================================

def build_hf_bbox_cache(hf_dataset_name: str = 'FlagEval/EmbSpatial-Bench') -> Dict[str, dict]:
    """Load HF dataset and build bbox lookup cache keyed by string-normalized question_id."""
    from datasets import load_dataset
    logger.info(f"Loading HF dataset: {hf_dataset_name}")
    ds = load_dataset(hf_dataset_name, split='test')

    cache = {}
    for item in ds:
        # Fix 4: Normalize key to string for consistent matching
        qid = str(item['question_id'])
        cache[qid] = {
            'objects': item['objects'],
            'relation': item['relation'],
            'data_source': item['data_source'],
            'answer': item['answer'],
            'answer_options': item['answer_options'],
        }

    # Fix 4: Log sample keys for debugging
    sample_keys = list(cache.keys())[:5]
    logger.info(f"Built bbox cache: {len(cache)} entries (sample keys: {sample_keys})")
    return cache


def get_bbox_center_y(bbox: list) -> float:
    return bbox[1] + bbox[3] / 2


def create_cross_group_quads(
    swap_pairs: List[dict],
    hf_cache: Dict[str, dict],
    threshold_ratio: float = 0.05,
    question_type: str = 'mcq',
) -> List[dict]:
    """For far/close swap pairs, create additional vertical queries using bbox."""
    IMAGE_HEIGHTS = {'ai2thor': 300, 'mp3d': 480, 'scannet': 968}

    quads = []
    stats = {'total': 0, 'matched': 0, 'ambiguous': 0, 'no_bbox': 0}

    distance_pairs = [p for p in swap_pairs if p['group'] == 'distance']

    # Fix 4: Use question_id (e.g. 'mp3d_0') to match HF dataset, not integer index
    n_matched_keys = sum(1 for p in distance_pairs if p['question_id'] in hf_cache)
    logger.info(f"Matched {n_matched_keys}/{len(distance_pairs)} question_ids between TSV and HF dataset")

    for pair in distance_pairs:
        stats['total'] += 1
        qid = pair['question_id']

        if qid not in hf_cache:
            stats['no_bbox'] += 1
            continue

        hf_item = hf_cache[qid]
        names = hf_item['objects']['name']
        bboxes = hf_item['objects']['bbox']

        target = pair['target_object']
        reference = pair['reference_object']

        target_bbox_y, ref_bbox_y = None, None
        for name, bbox in zip(names, bboxes):
            if name == target:
                target_bbox_y = get_bbox_center_y(bbox)
            if name == reference:
                ref_bbox_y = get_bbox_center_y(bbox)

        if target_bbox_y is None or ref_bbox_y is None:
            stats['no_bbox'] += 1
            continue

        image_height = IMAGE_HEIGHTS.get(hf_item['data_source'], 480)
        threshold = image_height * threshold_ratio
        y_diff = target_bbox_y - ref_bbox_y

        if abs(y_diff) < threshold:
            stats['ambiguous'] += 1
            continue

        if target_bbox_y < ref_bbox_y:
            vert_original_answer = 'above'
        else:
            vert_original_answer = 'below'

        if question_type == 'short_answer':
            vert_tmpl = SHORT_TEMPLATES['vertical']
            vert_mcq_map = None
            vert_original_q = vert_tmpl.format(obj1=target, obj2=reference)
            vert_swapped_q = vert_tmpl.format(obj1=reference, obj2=target)
            vert_swapped_answer = SHORT_OPPOSITE_MAP[vert_original_answer]
        else:
            vert_variant = 'above_first' if len(quads) % 2 == 0 else 'below_first'
            vert_tmpl = MCQ_TEMPLATES['vertical'][vert_variant]
            vert_mcq_map = MCQ_LETTER['vertical'][vert_variant]
            vert_original_q = vert_tmpl.format(obj1=target, obj2=reference)
            vert_swapped_q = vert_tmpl.format(obj1=reference, obj2=target)
            vert_swapped_answer = OPPOSITE_MAP[vert_original_answer]

        quad = {
            'index': pair['index'],
            'image_base64': pair['image_base64'],
            'dist_original_q': pair['original_question'],
            'dist_swapped_q': pair['swapped_question'],
            'dist_original_answer': pair['original_answer'],
            'dist_swapped_answer': pair['swapped_answer'],
            'dist_mcq_map': pair['mcq_map'],
            'vert_original_q': vert_original_q,
            'vert_swapped_q': vert_swapped_q,
            'vert_original_answer': vert_original_answer,
            'vert_swapped_answer': vert_swapped_answer,
            'vert_mcq_map': vert_mcq_map,
            'target_object': target,
            'reference_object': reference,
            'target_bbox_y': target_bbox_y,
            'ref_bbox_y': ref_bbox_y,
            'y_diff': y_diff,
            'data_source': hf_item['data_source'],
        }
        quads.append(quad)
        stats['matched'] += 1

    logger.info(f"Cross-group quads: {stats['matched']}/{stats['total']} "
                f"(ambiguous={stats['ambiguous']}, no_bbox={stats['no_bbox']})")
    return quads


# ============================================================================
# Base Extractor
# ============================================================================

class BaseHiddenStateExtractor(ABC):
    def __init__(self, model_path: str, device: str = 'cuda', target_layers: List[int] = None):
        self.model_path = model_path
        self.device = device
        self.hidden_states = {}
        self.hooks = []
        self._load_model()
        num_layers = self._get_num_layers()
        if target_layers is None:
            self.target_layers = list(range(num_layers))
            logger.info(f"Model has {num_layers} layers. Extracting ALL.")
        else:
            self.target_layers = target_layers
        self._register_hooks()

    def _register_hooks(self):
        for layer_idx in self.target_layers:
            module = self._get_layer_module(layer_idx)
            if module is not None:
                hook = module.register_forward_hook(self._make_hook(layer_idx))
                self.hooks.append(hook)

    def _make_hook(self, layer_idx: int):
        def hook_fn(module, input, output):
            if isinstance(output, tuple):
                hidden = output[0]
            else:
                hidden = output
            if hidden.shape[1] > 1:  # prefill only
                last_token = hidden[:, -1, :].detach().cpu().float()
                self.hidden_states[layer_idx] = last_token.squeeze(0)
        return hook_fn

    @abstractmethod
    def _load_model(self): pass
    @abstractmethod
    def _get_num_layers(self) -> int: pass
    @abstractmethod
    def _get_layer_module(self, layer_idx: int): pass
    @abstractmethod
    def extract_and_predict(self, image: Image.Image, question: str) -> Tuple[Dict[int, torch.Tensor], str]: pass

    def cleanup(self):
        for hook in self.hooks:
            hook.remove()
        self.hooks = []
        if hasattr(self, 'model'):
            del self.model
        if hasattr(self, 'processor'):
            del self.processor
        torch.cuda.empty_cache()


# ============================================================================
# Molmo Extractor
# ============================================================================

class MolmoExtractor(BaseHiddenStateExtractor):
    def _load_model(self):
        config_path = os.path.join(self.model_path, "config.yaml")
        checkpoint_path = os.path.join(self.model_path, "model.pt")
        if os.path.exists(config_path) and os.path.exists(checkpoint_path):
            self._load_native_model()
            self.is_native = True
        else:
            self._load_hf_model()
            self.is_native = False

    def _load_native_model(self):
        from olmo.config import ModelConfig
        from olmo.model import Molmo as NativeMolmoModel
        from olmo.data.model_preprocessor import MultiModalPreprocessor
        from olmo.data.data_formatter import DataFormatter

        _original_load = torch.load
        def _unsafe_load_wrapper(*args, **kwargs):
            if 'weights_only' not in kwargs:
                kwargs['weights_only'] = False
            return _original_load(*args, **kwargs)
        torch.load = _unsafe_load_wrapper

        cfg = ModelConfig.load(
            os.path.join(self.model_path, "config.yaml"),
            key="model", validate_paths=False
        )
        cfg.init_device = "cpu"
        self.model = NativeMolmoModel(cfg)
        state_dict = torch.load(os.path.join(self.model_path, "model.pt"), map_location="cpu")
        self.model.load_state_dict(state_dict)
        self.model = self.model.to(self.device, dtype=torch.bfloat16).eval()
        self.tokenizer = cfg.get_tokenizer()

        v_cfg = cfg.vision_backbone
        h, w = cfg.llm_patches_per_crop()
        image_padding_mask = 2 if cfg.fix_image_padding else (1 if cfg.image_padding_embed else None)

        class SafeDataFormatter(DataFormatter):
            def get_system_prompt(self, style, for_inference, messages, rng=None):
                if style is None:
                    style = "User"
                return super().get_system_prompt(style, for_inference, messages, rng)

        self.formatter = SafeDataFormatter(
            prompt_templates=cfg.prompt_type, message_format=cfg.message_formatting,
            system_prompt=cfg.system_prompt_kind, always_start_with_space=cfg.always_start_with_space,
            default_inference_len=cfg.default_inference_len
        )
        self.preprocessor = MultiModalPreprocessor(
            tokenizer=self.tokenizer, normalize=str(v_cfg.image_model_type),
            crop_mode=cfg.crop_mode, max_crops=cfg.max_crops,
            overlap_margins=cfg.overlap_margins, resize=v_cfg.resize_mode,
            use_col_tokens=cfg.use_col_tokens, base_image_input_size=v_cfg.image_default_input_size,
            image_pooling_w=cfg.image_pooling_w, image_pooling_h=cfg.image_pooling_h,
            image_token_length_w=w, image_token_length_h=h,
            image_patch_size=v_cfg.image_patch_size, image_padding_mask=image_padding_mask,
            pad_value=cfg.pad_value, loss_token_weighting=cfg.multi_annotation_weighting,
        )
        logger.info(f"Loaded native Molmo from {self.model_path}")

    def _load_hf_model(self):
        from transformers import AutoModelForCausalLM, AutoProcessor
        self.model = AutoModelForCausalLM.from_pretrained(
            self.model_path, torch_dtype=torch.bfloat16,
            trust_remote_code=True, device_map=self.device
        ).eval()
        self.processor = AutoProcessor.from_pretrained(self.model_path, trust_remote_code=True)
        logger.info(f"Loaded HF Molmo from {self.model_path}")

    def _get_num_layers(self) -> int:
        if self.is_native:
            return len(self.model.transformer.blocks)
        if hasattr(self.model, 'model') and hasattr(self.model.model, 'transformer'):
            return len(self.model.model.transformer.blocks)
        return 32

    def _get_layer_module(self, layer_idx: int):
        if self.is_native:
            return self.model.transformer.blocks[layer_idx]
        return self.model.model.transformer.blocks[layer_idx]

    def extract_and_predict(self, image, question):
        self.hidden_states = {}
        if self.is_native:
            example = {"messages": [question], "image": image}
            messages, _ = self.formatter(example, is_training=False, for_inference=True, rng=np.random)
            batch = self.preprocessor(np.array(image), messages, is_training=False, require_image_features=True)
            if 'input_ids' not in batch and 'input_tokens' in batch:
                batch['input_ids'] = batch['input_tokens']

            def to_t(x):
                return torch.from_numpy(x) if isinstance(x, np.ndarray) else x

            input_ids = to_t(batch['input_ids']).unsqueeze(0).to(self.device).long()
            images_t = to_t(batch['images']).unsqueeze(0).to(self.device, dtype=torch.bfloat16)
            image_masks = to_t(batch['image_masks']).unsqueeze(0).to(self.device, dtype=torch.bfloat16)
            image_input_idx = to_t(batch['image_input_idx']).unsqueeze(0).to(self.device)

            with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
                gen = self.model.generate(
                    input_ids=input_ids, images=images_t,
                    image_masks=image_masks, image_input_idx=image_input_idx,
                    max_steps=20, beam_size=1,
                )
            generated_ids = gen.token_ids[0, 0]
            answer = self.tokenizer.decode(generated_ids.tolist()).strip()
            for eos in ['<|endoftext|>', '</s>', '<|end|>']:
                answer = answer.replace(eos, '').strip()
        else:
            from transformers import GenerationConfig
            inputs = self.processor.process(images=[image], text=question)
            processed = {}
            for k, v in inputs.items():
                v = v.to(self.device).unsqueeze(0)
                if v.dtype == torch.float32:
                    v = v.to(dtype=torch.bfloat16)
                processed[k] = v
            with torch.no_grad(), torch.autocast("cuda", dtype=torch.bfloat16):
                output = self.model.generate_from_batch(
                    processed,
                    GenerationConfig(max_new_tokens=20, stop_strings="<|endoftext|>"),
                    tokenizer=self.processor.tokenizer,
                )
            input_len = processed['input_ids'].shape[1]
            answer = self.processor.tokenizer.decode(output[0, input_len:], skip_special_tokens=True).strip()

        return self.hidden_states.copy(), answer


# ============================================================================
# NVILA Extractor
# ============================================================================

class NVILAExtractor(BaseHiddenStateExtractor):
    def _load_model(self):
        original_sys_path = sys.path.copy()
        sys.path = [p for p in sys.path if 'RoboRefer' not in p]
        modules_to_remove = [k for k in list(sys.modules.keys()) if 'llava' in k.lower()]
        removed = {m: sys.modules.pop(m) for m in modules_to_remove}
        try:
            import llava
            from llava.media import Image as LLaVAImage
            from llava import conversation as clib
        except Exception as err:
            sys.path = original_sys_path
            for m, mod in removed.items():
                sys.modules[m] = mod
            raise RuntimeError(f"Failed to import llava: {err}")
        sys.path = original_sys_path
        self.LLaVAImage = LLaVAImage
        self.clib = clib
        self.model = llava.load(self.model_path, model_base=None)
        self._find_llm_backbone()
        logger.info(f"Loaded NVILA from {self.model_path}")

    def _find_llm_backbone(self):
        candidates = []
        if hasattr(self.model, 'llm'):
            if hasattr(self.model.llm, 'model') and hasattr(self.model.llm.model, 'layers'):
                candidates.append(self.model.llm.model.layers)
            if hasattr(self.model.llm, 'layers'):
                candidates.append(self.model.llm.layers)
        if hasattr(self.model, 'model'):
            if hasattr(self.model.model, 'model') and hasattr(self.model.model.model, 'layers'):
                candidates.append(self.model.model.model.layers)
            if hasattr(self.model.model, 'layers'):
                candidates.append(self.model.model.layers)
        for name, module in self.model.named_modules():
            if name.endswith('.layers') and hasattr(module, '__len__') and len(module) > 0:
                candidates.append(module)
        if candidates:
            self.llm_backbone = candidates[0]
        else:
            raise ValueError("Could not locate transformer layers in NVILA model")

    def _get_num_layers(self) -> int:
        return len(self.llm_backbone) if hasattr(self, 'llm_backbone') else 24

    def _get_layer_module(self, layer_idx: int):
        return self.llm_backbone[layer_idx]

    def extract_and_predict(self, image, question):
        self.hidden_states = {}
        import tempfile
        with tempfile.NamedTemporaryFile(suffix='.png', delete=False) as f:
            temp_path = f.name
            image.save(temp_path)
        try:
            prompt = [self.LLaVAImage(temp_path), question]
            from transformers import GenerationConfig
            response = self.model.generate_content(
                prompt, generation_config=GenerationConfig(max_new_tokens=20, do_sample=False)
            )
        finally:
            os.unlink(temp_path)
        answer = str(response[0] if isinstance(response, list) else response).strip()
        return self.hidden_states.copy(), answer


class RoboReferExtractor(NVILAExtractor):
    ROBOREFER_PATH = '/data/shared/Qwen/RoboRefer'

    def _load_model(self):
        original_sys_path = sys.path.copy()
        if self.ROBOREFER_PATH not in sys.path:
            sys.path.insert(0, self.ROBOREFER_PATH)
        modules_to_remove = [k for k in list(sys.modules.keys()) if 'llava' in k.lower()]
        removed = {m: sys.modules.pop(m) for m in modules_to_remove}
        try:
            import llava
            from llava.media import Image as LLaVAImage
            from llava import conversation as clib
        except Exception as err:
            sys.path = original_sys_path
            for m, mod in removed.items():
                sys.modules[m] = mod
            raise RuntimeError(f"Failed to import RoboRefer llava: {err}")
        sys.path = original_sys_path
        self.LLaVAImage = LLaVAImage
        self.clib = clib
        self.model = llava.load(self.model_path, model_base=None)
        self._find_llm_backbone()
        logger.info(f"Loaded RoboRefer from {self.model_path}")


class RoboReferDepthExtractor(RoboReferExtractor):
    """RoboRefer with depth-image input instead of RGB.

    Usage: pass the depth PIL image to extract_and_predict() instead of the RGB image.
    In practice this means loading depth images in load_swap_pairs() / extract_swap_features()
    rather than changing anything here.  If the depth image is stored as a separate column in
    the dataset, add a 'depth_image_base64' key to the pair dict and decode it before calling
    run_single_query().

    TODO: confirm depth image path / format with the actual dataset layout.
    """
    # Inherits extract_and_predict() from NVILAExtractor (via RoboReferExtractor) unchanged.
    # The caller is responsible for passing the correct (depth) PIL Image.


# ============================================================================
# Qwen2.5-VL Extractor
# ============================================================================

class Qwen25VLExtractor(BaseHiddenStateExtractor):
    BASE_MODEL = "Qwen/Qwen2.5-VL-3B-Instruct"

    def _load_model(self):
        from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
        try:
            self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
                self.model_path, torch_dtype=torch.bfloat16, device_map=self.device
            )
        except ImportError:
            self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
                self.model_path, torch_dtype=torch.bfloat16
            ).to(self.device)
        self.model.eval()
        if self.model_path.startswith('/'):
            self.processor = AutoProcessor.from_pretrained(self.BASE_MODEL)
        else:
            self.processor = AutoProcessor.from_pretrained(self.model_path)
        logger.info(f"Loaded Qwen2.5-VL from {self.model_path}")

    def _get_num_layers(self) -> int:
        return len(self.model.model.layers)

    def _get_layer_module(self, layer_idx: int):
        return self.model.model.layers[layer_idx]

    def extract_and_predict(self, image, question):
        self.hidden_states = {}
        messages = [{"role": "user", "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": question}
        ]}]
        text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        from qwen_vl_utils import process_vision_info
        image_inputs, video_inputs = process_vision_info(messages)
        inputs = self.processor(
            text=[text], images=image_inputs, videos=video_inputs,
            padding=True, return_tensors="pt"
        ).to(self.device)
        with torch.no_grad():
            output_ids = self.model.generate(**inputs, max_new_tokens=20, do_sample=False)
        input_len = inputs['input_ids'].shape[1]
        answer = self.processor.tokenizer.decode(output_ids[0, input_len:], skip_special_tokens=True).strip()
        return self.hidden_states.copy(), answer


# ============================================================================
# New Extractors: Molmo2-8B and Qwen3-VL family
# ============================================================================

class Molmo2Extractor(BaseHiddenStateExtractor):
    """Extractor for allenai/Molmo2-8B (AutoModelForImageTextToText, messages-dict input)."""

    def _load_model(self):
        from transformers import AutoProcessor, AutoModelForImageTextToText
        self.processor = AutoProcessor.from_pretrained(self.model_path, trust_remote_code=True)
        self.model = AutoModelForImageTextToText.from_pretrained(
            self.model_path, trust_remote_code=True, torch_dtype='auto', device_map='auto',
        ).eval()
        self._find_llm_layers()
        logger.info(f"Loaded Molmo2 from {self.model_path}")

    def _find_llm_layers(self):
        candidates = [
            ['model', 'layers'],
            ['language_model', 'model', 'layers'],
            ['model', 'model', 'layers'],
        ]
        for path in candidates:
            obj = self.model
            for attr in path:
                obj = getattr(obj, attr, None)
                if obj is None:
                    break
            if obj is not None and hasattr(obj, '__len__') and len(obj) > 0:
                self.llm_layers = obj
                logger.info(f"Molmo2: layers at '{'.'.join(path)}', count={len(obj)}")
                return
        best, best_len = None, 0
        for name, module in self.model.named_modules():
            if name.endswith('.layers') and hasattr(module, '__len__') and len(module) > best_len:
                best, best_len = module, len(module)
                logger.info(f"Molmo2: layers via scan at '{name}', count={best_len}")
        if best is not None:
            self.llm_layers = best
            return
        raise ValueError("Could not find transformer layers in Molmo2 model")

    def _get_num_layers(self) -> int:
        return len(self.llm_layers)

    def _get_layer_module(self, layer_idx: int):
        return self.llm_layers[layer_idx]

    def extract_and_predict(self, image, question):
        self.hidden_states = {}
        messages = [{"role": "user", "content": [
            {"type": "image", "image": image},
            {"type": "text",  "text":  question},
        ]}]
        inputs = self.processor.apply_chat_template(
            messages, tokenize=True, add_generation_prompt=True,
            return_tensors="pt", return_dict=True,
        )
        inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
        with torch.inference_mode():
            generated_ids = self.model.generate(**inputs, max_new_tokens=20, do_sample=False)
        input_len = inputs['input_ids'].shape[1]
        answer = self.processor.tokenizer.decode(
            generated_ids[0, input_len:], skip_special_tokens=True).strip()
        return self.hidden_states.copy(), answer


class Qwen3VLExtractor(BaseHiddenStateExtractor):
    """Extractor for Qwen3-VL family (32B dense, 235B MoE).

    Key differences from Qwen25VLExtractor:
      - AutoModelForImageTextToText + trust_remote_code=True
      - process_vision_info requires image_patch_size=16
      - processor call requires do_resize=False
      - 32×32 px patches → different min/max_pixels
    """

    MIN_PIXELS = 256   * 32 * 32   # 262,144  (mp3d/scannet → natural res; ai2thor → ~256 tokens)
    MAX_PIXELS = 16384 * 32 * 32   # 16,777,216

    def _load_model(self):
        from transformers import AutoProcessor, AutoModelForImageTextToText
        self.processor = AutoProcessor.from_pretrained(self.model_path, trust_remote_code=True)
        self.model = AutoModelForImageTextToText.from_pretrained(
            self.model_path, trust_remote_code=True, torch_dtype='auto',
            device_map='auto', attn_implementation='flash_attention_2',
        ).eval()
        self._find_llm_layers()
        logger.info(f"Loaded Qwen3-VL from {self.model_path}")

    def _find_llm_layers(self):
        candidates = [
            ['model', 'language_model', 'model', 'layers'],  # Qwen3-VL expected
            ['language_model', 'model', 'layers'],
            ['model', 'model', 'layers'],
            ['model', 'layers'],
        ]
        for path in candidates:
            obj = self.model
            for attr in path:
                obj = getattr(obj, attr, None)
                if obj is None:
                    break
            if obj is not None and hasattr(obj, '__len__') and len(obj) > 0:
                self.llm_layers = obj
                logger.info(f"Qwen3-VL: layers at '{'.'.join(path)}', count={len(obj)}")
                return
        best, best_len = None, 0
        for name, module in self.model.named_modules():
            if name.endswith('.layers') and hasattr(module, '__len__') and len(module) > best_len:
                best, best_len = module, len(module)
                logger.info(f"Qwen3-VL: layers via scan at '{name}', count={best_len}")
        if best is not None:
            self.llm_layers = best
            return
        raise ValueError("Could not find transformer layers in Qwen3-VL model")

    def _get_num_layers(self) -> int:
        return len(self.llm_layers)

    def _get_layer_module(self, layer_idx: int):
        return self.llm_layers[layer_idx]

    def extract_and_predict(self, image, question):
        self.hidden_states = {}
        messages = [{"role": "user", "content": [
            {"type": "image", "image": image,
             "min_pixels": self.MIN_PIXELS, "max_pixels": self.MAX_PIXELS},
            {"type": "text", "text": question},
        ]}]
        text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        from qwen_vl_utils import process_vision_info
        images, videos, _ = process_vision_info(
            messages, image_patch_size=16, return_video_kwargs=True, return_video_metadata=True,
        )
        inputs = self.processor(
            text=text, images=images, videos=videos, do_resize=False, return_tensors="pt",
        ).to(self.model.device)
        with torch.no_grad():
            output_ids = self.model.generate(**inputs, max_new_tokens=20, do_sample=False)
        input_len = inputs['input_ids'].shape[1]
        answer = self.processor.tokenizer.decode(
            output_ids[0, input_len:], skip_special_tokens=True).strip()
        return self.hidden_states.copy(), answer


EXTRACTOR_CLASSES = {
    'MolmoExtractor':     MolmoExtractor,
    'NVILAExtractor':     NVILAExtractor,
    'RoboReferExtractor': RoboReferExtractor,
    'Qwen25VLExtractor':  Qwen25VLExtractor,
    'Molmo2Extractor':    Molmo2Extractor,
    'Qwen3VLExtractor':   Qwen3VLExtractor,
}


def get_extractor(model_type: str, model_path: str = None, scale: str = None, **kwargs):
    """Create an extractor for any model_type (legacy or new-large)."""
    # New large models: (ExtractorClass, path) tuples in MODEL_CONFIGS_NEW
    if model_type in MODEL_CONFIGS_NEW:
        cls_name, raw_path = MODEL_CONFIGS_NEW[model_type][scale]
        resolved = resolve_local_path(raw_path)
        logger.info(f"Creating {cls_name} for scale='{scale}' from {resolved}")
        return EXTRACTOR_CLASSES[cls_name](resolved, **kwargs)
    # Legacy models
    if model_type == 'nvila' and scale == 'roborefer':
        return RoboReferExtractor(model_path, **kwargs)
    if model_type == 'nvila' and scale == 'roborefer_depth':
        return RoboReferDepthExtractor(model_path, **kwargs)
    legacy = {
        'molmo': MolmoExtractor, 'nvila': NVILAExtractor, 'qwen': Qwen25VLExtractor,
        'nvila_synthetic': NVILAExtractor, 'nvila_st': NVILAExtractor,
    }
    return legacy[model_type](model_path, **kwargs)


# ============================================================================
# Feature Extraction Pipeline
# ============================================================================

def run_single_query(extractor, image, question):
    hidden_states, predicted = extractor.extract_and_predict(image, question)
    result = {}
    for layer_idx in extractor.target_layers:
        if layer_idx in hidden_states:
            state = hidden_states[layer_idx].numpy().flatten()
            if state.size > 0:
                result[layer_idx] = state
    return result, predicted


def extract_swap_features(
    extractor: BaseHiddenStateExtractor,
    swap_pairs: List[dict],
    max_samples_per_category: int = 0,
) -> List[dict]:
    """Extract features for all swap pairs."""
    rng = random.Random(42)

    if max_samples_per_category > 0:
        grouped = defaultdict(list)
        for p in swap_pairs:
            grouped[p['category']].append(p)
        limited = []
        for cat in CATEGORY_ORDER:
            samples = grouped[cat]
            if len(samples) > max_samples_per_category:
                samples = rng.sample(samples, max_samples_per_category)
            limited.extend(samples)
        swap_pairs = limited

    records = []
    for pair in tqdm(swap_pairs, desc="Swap pairs"):
        try:
            image = decode_base64_image(pair['image_base64'])
            hs_orig, pred_orig = run_single_query(extractor, image, pair['original_question'])
            hs_swap, pred_swap = run_single_query(extractor, image, pair['swapped_question'])

            is_correct_orig = check_answer(pred_orig, pair['original_answer'], pair['mcq_map'])
            is_correct_swap = check_answer(pred_swap, pair['swapped_answer'], pair['mcq_map'])

            delta = {}
            for layer_idx in extractor.target_layers:
                if layer_idx in hs_orig and layer_idx in hs_swap:
                    delta[layer_idx] = hs_swap[layer_idx] - hs_orig[layer_idx]

            record = {
                'index': pair['index'],
                'group': pair['group'],
                'category': pair['category'],
                'original_answer': pair['original_answer'],
                'swapped_answer': pair['swapped_answer'],
                'pred_orig': pred_orig,
                'pred_swap': pred_swap,
                'is_correct_orig': is_correct_orig,
                'is_correct_swap': is_correct_swap,
                'hs_orig': hs_orig,
                'hs_swap': hs_swap,
                'delta': delta,
            }
            records.append(record)

            mark_o = "O" if is_correct_orig else "X"
            mark_s = "O" if is_correct_swap else "X"
            logger.info(f"  #{pair['index']:<6} {pair['category']:<6} "
                        f"orig[{mark_o}]=\"{pred_orig[:40]}\" swap[{mark_s}]=\"{pred_swap[:40]}\""
                        + (f"  [{len(records)}/{len(swap_pairs)}]" if len(records) % 50 == 0 else ""))

        except Exception as e:
            logger.warning(f"Error on index {pair['index']}: {e}")
            continue

    logger.info(f"Extracted {len(records)} swap pair records")

    # Fix 8: Per-category accuracy logging
    for cat in CATEGORY_ORDER:
        cat_recs = [r for r in records if r['category'] == cat]
        n = len(cat_recs)
        if n == 0:
            continue
        c_orig = sum(1 for r in cat_recs if r['is_correct_orig'])
        c_swap = sum(1 for r in cat_recs if r['is_correct_swap'])
        c_both = sum(1 for r in cat_recs if r['is_correct_orig'] and r['is_correct_swap'])
        logger.info(f"  {cat:>6s} (n={n}): acc_orig={c_orig/n:.1%}, acc_swap={c_swap/n:.1%}, "
                     f"acc_both={c_both/n:.1%}")

    return records


def extract_cross_group_features(
    extractor: BaseHiddenStateExtractor,
    quads: List[dict],
) -> List[dict]:
    """Extract features for cross-group quads (4 forward passes each)."""
    records = []
    for quad in tqdm(quads, desc="Cross-group quads"):
        try:
            image = decode_base64_image(quad['image_base64'])
            hs_d_orig, pred_d_orig = run_single_query(extractor, image, quad['dist_original_q'])
            hs_d_swap, pred_d_swap = run_single_query(extractor, image, quad['dist_swapped_q'])
            hs_v_orig, pred_v_orig = run_single_query(extractor, image, quad['vert_original_q'])
            hs_v_swap, pred_v_swap = run_single_query(extractor, image, quad['vert_swapped_q'])

            delta_dist, delta_vert = {}, {}
            for layer_idx in extractor.target_layers:
                if layer_idx in hs_d_orig and layer_idx in hs_d_swap:
                    delta_dist[layer_idx] = hs_d_swap[layer_idx] - hs_d_orig[layer_idx]
                if layer_idx in hs_v_orig and layer_idx in hs_v_swap:
                    delta_vert[layer_idx] = hs_v_swap[layer_idx] - hs_v_orig[layer_idx]

            record = {
                'index': quad['index'],
                'delta_dist': delta_dist,
                'delta_vert': delta_vert,
                'pred_d_orig': pred_d_orig, 'pred_d_swap': pred_d_swap,
                'pred_v_orig': pred_v_orig, 'pred_v_swap': pred_v_swap,
                'is_correct_d_orig': check_answer(pred_d_orig, quad['dist_original_answer'], quad['dist_mcq_map']),
                'is_correct_d_swap': check_answer(pred_d_swap, quad['dist_swapped_answer'], quad['dist_mcq_map']),
                'is_correct_v_orig': check_answer(pred_v_orig, quad['vert_original_answer'], quad['vert_mcq_map']),
                'is_correct_v_swap': check_answer(pred_v_swap, quad['vert_swapped_answer'], quad['vert_mcq_map']),
                'data_source': quad['data_source'],
            }
            records.append(record)

            tqdm.write(f"  #{quad['index']:<6} dist=[{pred_d_orig[:20]}/{pred_d_swap[:20]}] "
                       f"vert=[{pred_v_orig[:20]}/{pred_v_swap[:20]}]")

        except Exception as e:
            logger.warning(f"Error on cross-group index {quad['index']}: {e}")
            continue

    logger.info(f"Extracted {len(records)} cross-group quad records")
    return records


# ============================================================================
# Analysis Functions
# ============================================================================

# Fix 5: Within-category + sign-corrected delta consistency

def compute_delta_consistency(records: List[dict], target_layers: List[int]):
    """Compute TWO types of delta consistency.

    Returns:
        within_cat_results: {(category, layer) -> {mean, std, n}}
        sign_corrected_results: {(group, layer) -> {mean, std, n}}
    """
    within_cat_results = {}
    sign_corrected_results = {}

    for group in GROUP_ORDER:
        canonical = CANONICAL_CATEGORIES[group]
        opposite = OPPOSITE_MAP[canonical]
        group_recs = [r for r in records if r['group'] == group]

        for layer in target_layers:
            # (a) Within-category consistency
            for cat in [canonical, opposite]:
                cat_deltas = [r['delta'][layer] for r in group_recs
                              if r['category'] == cat and layer in r['delta']]
                if len(cat_deltas) >= 2:
                    arr = np.array(cat_deltas)
                    sim = cosine_similarity(arr)
                    upper = sim[np.triu_indices(len(cat_deltas), k=1)]
                    within_cat_results[(cat, layer)] = {
                        'mean': float(np.mean(upper)),
                        'std': float(np.std(upper)),
                        'n': len(cat_deltas),
                    }

            # (b) Sign-corrected group consistency
            all_deltas = []
            for r in group_recs:
                if layer not in r['delta']:
                    continue
                d = r['delta'][layer]
                if r['category'] == opposite:
                    d = -d  # flip to align with canonical direction
                all_deltas.append(d)

            if len(all_deltas) >= 2:
                arr = np.array(all_deltas)
                sim = cosine_similarity(arr)
                upper = sim[np.triu_indices(len(all_deltas), k=1)]
                sign_corrected_results[(group, layer)] = {
                    'mean': float(np.mean(upper)),
                    'std': float(np.std(upper)),
                    'n': len(all_deltas),
                }

    return within_cat_results, sign_corrected_results


# Fix 7: Delta-based similarity matrix

def compute_delta_similarity_matrix(records: List[dict], layer: int) -> Optional[pd.DataFrame]:
    """Compute 6x6 cosine similarity using mean delta per category."""
    cat_deltas = {}
    for cat in CATEGORY_ORDER:
        deltas = [r['delta'][layer] for r in records if r['category'] == cat and layer in r['delta']]
        if deltas:
            cat_deltas[cat] = np.mean(deltas, axis=0)

    available = [c for c in CATEGORY_ORDER if c in cat_deltas]
    if len(available) < 2:
        return None

    vectors = np.array([cat_deltas[c] for c in available])
    sim = cosine_similarity(vectors)
    return pd.DataFrame(sim, index=available, columns=available)


# ============================================================
# CHANGE 1: Add this function right after compute_delta_similarity_matrix()
#           (around line ~820, after the "Fix 7: Delta-based similarity matrix" block)
# ============================================================

def compute_delta_norm_per_category(records: List[dict], layer: int) -> Optional[pd.DataFrame]:
    """Compute mean delta vector norm per category at a given layer.

    Returns a single-column DataFrame with index = category label, column = 'norm'.
    Format matches what summarize_metrics_updated.py expects:
        delta_norm_{scale}_L{layer}_all_pairs.csv
            ,norm
            left,12.34
            right,11.89
            above,9.45
            below,9.12
            far,7.23
            close,7.58

    Returns None if no records have delta vectors at this layer.
    """
    rows = {}
    for cat in CATEGORY_ORDER:
        deltas = [r['delta'][layer] for r in records
                  if r['category'] == cat and layer in r['delta']]
        if deltas:
            norms = [float(np.linalg.norm(d)) for d in deltas]
            rows[cat] = float(np.mean(norms))

    if not rows:
        return None

    df = pd.DataFrame.from_dict(rows, orient='index', columns=['norm'])
    # Reorder to canonical category order (skip any that are missing)
    ordered = [c for c in CATEGORY_ORDER if c in df.index]
    return df.loc[ordered]


# Fix 8: Both-correct filtering

def filter_both_correct(records: List[dict]) -> List[dict]:
    """Filter to pairs where both orig and swap predictions are correct."""
    return [r for r in records if r['is_correct_orig'] and r['is_correct_swap']]


# Fix 8: Category validity check

def check_category_validity(records: List[dict], scale: str) -> Dict[str, dict]:
    """Check per-category accuracy and flag unreliable categories."""
    validity = {}
    for cat in CATEGORY_ORDER:
        cat_recs = [r for r in records if r['category'] == cat]
        n = len(cat_recs)
        if n == 0:
            validity[cat] = {'n': 0, 'acc_orig': 0, 'acc_swap': 0, 'reliable': False}
            continue
        acc_orig = sum(1 for r in cat_recs if r['is_correct_orig']) / n
        acc_swap = sum(1 for r in cat_recs if r['is_correct_swap']) / n
        reliable = acc_orig >= 0.5 and acc_swap >= 0.5
        validity[cat] = {
            'n': n, 'acc_orig': acc_orig, 'acc_swap': acc_swap,
            'reliable': reliable,
        }
        if not reliable:
            logger.warning(f"  [!] Category '{cat}' unreliable at scale={scale}: "
                           f"acc_orig={acc_orig:.1%}, acc_swap={acc_swap:.1%}")
    return validity


def compute_cross_group_alignment(quad_records: List[dict], target_layers: List[int]) -> dict:
    results = {}
    for layer in target_layers:
        per_sample = []
        delta_verts, delta_dists = [], []

        for rec in quad_records:
            if layer in rec['delta_vert'] and layer in rec['delta_dist']:
                dv = rec['delta_vert'][layer]
                dd = rec['delta_dist'][layer]
                norm_v, norm_d = np.linalg.norm(dv), np.linalg.norm(dd)
                if norm_v > 1e-10 and norm_d > 1e-10:
                    per_sample.append(float(np.dot(dv, dd) / (norm_v * norm_d)))
                    delta_verts.append(dv)
                    delta_dists.append(dd)

        if not per_sample:
            continue

        mean_dv = np.mean(delta_verts, axis=0)
        mean_dd = np.mean(delta_dists, axis=0)
        norm_mv, norm_md = np.linalg.norm(mean_dv), np.linalg.norm(mean_dd)
        mean_alignment = float(np.dot(mean_dv, mean_dd) / (norm_mv * norm_md + 1e-10))

        rng = np.random.RandomState(42)
        perm_alignments = []
        for _ in range(100):
            shuffled_dd = [delta_dists[i] for i in rng.permutation(len(delta_dists))]
            perm_cos = []
            for dv, dd in zip(delta_verts, shuffled_dd):
                nv, nd = np.linalg.norm(dv), np.linalg.norm(dd)
                if nv > 1e-10 and nd > 1e-10:
                    perm_cos.append(np.dot(dv, dd) / (nv * nd))
            perm_alignments.append(np.mean(perm_cos))

        results[layer] = {
            'per_sample_mean': float(np.mean(per_sample)),
            'per_sample_std': float(np.std(per_sample)),
            'mean_delta_alignment': mean_alignment,
            'permutation_mean': float(np.mean(perm_alignments)),
            'permutation_std': float(np.std(perm_alignments)),
            'n_samples': len(per_sample),
        }
    return results


def compute_prediction_stats(records: List[dict], scale: str) -> dict:
    stats = {'scale': scale}
    total_correct_orig, total_correct_swap, total_both, total_n = 0, 0, 0, 0

    for group in GROUP_ORDER:
        group_recs = [r for r in records if r['group'] == group]
        n = len(group_recs)
        c_orig = sum(1 for r in group_recs if r['is_correct_orig'])
        c_swap = sum(1 for r in group_recs if r['is_correct_swap'])
        c_both = sum(1 for r in group_recs if r['is_correct_orig'] and r['is_correct_swap'])
        stats[f'{group}_n'] = n
        stats[f'{group}_acc_orig'] = c_orig / n if n > 0 else 0
        stats[f'{group}_acc_swap'] = c_swap / n if n > 0 else 0
        stats[f'{group}_acc_both'] = c_both / n if n > 0 else 0
        total_correct_orig += c_orig
        total_correct_swap += c_swap
        total_both += c_both
        total_n += n

    stats['overall_acc_orig'] = total_correct_orig / total_n if total_n > 0 else 0
    stats['overall_acc_swap'] = total_correct_swap / total_n if total_n > 0 else 0
    stats['overall_acc_both'] = total_both / total_n if total_n > 0 else 0
    stats['overall_n'] = total_n
    return stats


# ============================================================================
# Saving & Loading
# ============================================================================

def get_representative_layers(all_layers, n=5):
    if len(all_layers) <= n:
        return list(all_layers)
    indices = np.linspace(0, len(all_layers) - 1, n, dtype=int)
    return [all_layers[i] for i in indices]


def save_scale_results(
    scale, swap_records, quad_records,
    within_cat_consistency, sign_corrected_consistency,
    cross_alignment, pred_stats, target_layers,
    category_validity, delta_heatmaps,
    output_dir, both_correct_tag="all_pairs",
    save_alignment=True,
    delta_norms=None,           # <-- NEW: {layer: pd.DataFrame | None}
):
    """Save all per-scale results to disk.

    Args:
        save_alignment: If False, skip writing cross_alignment_{scale}.json.
                        Set False during Phase A save; call save_cross_alignment()
                        separately after Phase B completes.
    """
    csv_dir = os.path.join(output_dir, 'csv')
    json_dir = os.path.join(output_dir, 'json')
    os.makedirs(csv_dir, exist_ok=True)
    os.makedirs(json_dir, exist_ok=True)

    # 1. Predictions CSV (tagged so all_pairs and both_correct don't overwrite each other)
    pred_rows = []
    for r in swap_records:
        pred_rows.append({
            'index': r['index'], 'group': r['group'], 'category': r['category'],
            'pred_orig': r['pred_orig'], 'pred_swap': r['pred_swap'],
            'is_correct_orig': r['is_correct_orig'], 'is_correct_swap': r['is_correct_swap'],
        })
    pd.DataFrame(pred_rows).to_csv(
        os.path.join(csv_dir, f'predictions_{scale}_{both_correct_tag}.csv'), index=False)

    # 2. Within-category consistency JSON
    wc_data = {}
    for (cat, layer), vals in within_cat_consistency.items():
        wc_data[f'{cat}_L{layer}'] = vals
    with open(os.path.join(json_dir, f'within_cat_consistency_{scale}_{both_correct_tag}.json'), 'w') as f:
        json.dump(wc_data, f, indent=2)

    # 3. Sign-corrected consistency JSON
    sc_data = {}
    for (group, layer), vals in sign_corrected_consistency.items():
        sc_data[f'{group}_L{layer}'] = vals
    with open(os.path.join(json_dir, f'sign_corrected_consistency_{scale}_{both_correct_tag}.json'), 'w') as f:
        json.dump(sc_data, f, indent=2)

    # 4. Cross-group alignment JSON (only when save_alignment=True, i.e. after Phase B)
    if save_alignment:
        alignment_data = {}
        for layer, vals in cross_alignment.items():
            alignment_data[f'L{layer}'] = vals
        with open(os.path.join(json_dir, f'cross_alignment_{scale}.json'), 'w') as f:
            json.dump(alignment_data, f, indent=2)

    # 5. Prediction stats JSON
    with open(os.path.join(json_dir, f'pred_stats_{scale}.json'), 'w') as f:
        json.dump(pred_stats, f, indent=2)

    # 6. Category validity JSON (Fix 8)
    with open(os.path.join(json_dir, f'category_validity_{scale}.json'), 'w') as f:
        json.dump(category_validity, f, indent=2)

    # 7. Delta heatmap CSVs (Fix 7)
    for layer, df in delta_heatmaps.items():
        if df is not None:
            df.to_csv(os.path.join(csv_dir, f'delta_similarity_{scale}_L{layer}_{both_correct_tag}.csv'))

    # 8. Delta norm CSVs (mean L2 norm of delta vectors per category per layer)
    if delta_norms:
        for layer, df in delta_norms.items():
            if df is not None:
                df.to_csv(os.path.join(csv_dir, f'delta_norm_{scale}_L{layer}_{both_correct_tag}.csv'))


    logger.info(f"Saved results for scale={scale} ({both_correct_tag}) to {output_dir}")


def save_vectors_npz(scale, swap_records, target_layers, output_dir):
    """Save swap-pair vectors with correctness metadata to NPZ (Phase A result).

    This enables post-hoc filtering (both_correct, all_with_validity) from saved data.
    Cross-group vectors are saved separately by save_cross_group_npz() after Phase B.
    """
    rep_layers = list(target_layers)  # save ALL layers (not just 5 representative)
    delta_data = {}
    for layer in rep_layers:
        groups_list, categories_list, vectors = [], [], []
        orig_vecs, swap_vecs, labels = [], [], []
        correct_orig_list, correct_swap_list, indices_list = [], [], []
        for r in swap_records:
            if layer in r['delta']:
                groups_list.append(r['group'])
                categories_list.append(r['category'])
                vectors.append(r['delta'][layer])
                correct_orig_list.append(r['is_correct_orig'])
                correct_swap_list.append(r['is_correct_swap'])
                indices_list.append(r['index'])
            if layer in r['hs_orig'] and layer in r['hs_swap']:
                orig_vecs.append(r['hs_orig'][layer])
                swap_vecs.append(r['hs_swap'][layer])
                labels.append(r['category'])
        if vectors:
            delta_data[f'delta_L{layer}'] = np.array(vectors)
            delta_data[f'groups_L{layer}'] = np.array(groups_list)
            delta_data[f'categories_L{layer}'] = np.array(categories_list)
            delta_data[f'is_correct_orig_L{layer}'] = np.array(correct_orig_list)
            delta_data[f'is_correct_swap_L{layer}'] = np.array(correct_swap_list)
            delta_data[f'indices_L{layer}'] = np.array(indices_list)
        if orig_vecs:
            delta_data[f'orig_L{layer}'] = np.array(orig_vecs)
            delta_data[f'swap_L{layer}'] = np.array(swap_vecs)
            delta_data[f'labels_L{layer}'] = np.array(labels)

    npz_dir = os.path.join(output_dir, 'npz')
    os.makedirs(npz_dir, exist_ok=True)
    np.savez_compressed(os.path.join(npz_dir, f'vectors_{scale}.npz'), **delta_data)
    logger.info(f"Saved vectors NPZ with correctness metadata for scale={scale}")


def save_cross_group_npz(scale, quad_records, target_layers, output_dir):
    """Save cross-group delta vectors to NPZ (Phase B result)."""
    if not quad_records:
        return
    rep_layers = list(target_layers)
    cg_data = {}
    for layer in rep_layers:
        dverts, ddists = [], []
        for rec in quad_records:
            if layer in rec['delta_vert'] and layer in rec['delta_dist']:
                dverts.append(rec['delta_vert'][layer])
                ddists.append(rec['delta_dist'][layer])
        if dverts:
            cg_data[f'delta_vert_L{layer}'] = np.array(dverts)
            cg_data[f'delta_dist_L{layer}'] = np.array(ddists)
    npz_dir = os.path.join(output_dir, 'npz')
    os.makedirs(npz_dir, exist_ok=True)
    np.savez_compressed(os.path.join(npz_dir, f'cross_group_vectors_{scale}.npz'), **cg_data)
    logger.info(f"Saved cross-group vectors NPZ for scale={scale}")


def save_cross_alignment(scale, cross_alignment, output_dir):
    """Save cross-group alignment data to JSON (Phase B result)."""
    json_dir = os.path.join(output_dir, 'json')
    os.makedirs(json_dir, exist_ok=True)
    alignment_data = {f'L{layer}': vals for layer, vals in cross_alignment.items()}
    with open(os.path.join(json_dir, f'cross_alignment_{scale}.json'), 'w') as f:
        json.dump(alignment_data, f, indent=2)
    logger.info(f"Saved cross-alignment JSON for scale={scale}")


def load_scale_consistency(output_dir, scale, tag='all_pairs'):
    """Load sign-corrected consistency."""
    path = os.path.join(output_dir, 'json', f'sign_corrected_consistency_{scale}_{tag}.json')
    if not os.path.exists(path):
        return {}
    with open(path) as f:
        raw = json.load(f)
    result = {}
    for key, vals in raw.items():
        parts = key.rsplit('_L', 1)
        if len(parts) == 2:
            result[(parts[0], int(parts[1]))] = vals
    return result


def load_within_cat_consistency(output_dir, scale, tag='all_pairs'):
    path = os.path.join(output_dir, 'json', f'within_cat_consistency_{scale}_{tag}.json')
    if not os.path.exists(path):
        return {}
    with open(path) as f:
        raw = json.load(f)
    result = {}
    for key, vals in raw.items():
        parts = key.rsplit('_L', 1)
        if len(parts) == 2:
            result[(parts[0], int(parts[1]))] = vals
    return result


def load_scale_alignment(output_dir, scale):
    path = os.path.join(output_dir, 'json', f'cross_alignment_{scale}.json')
    if not os.path.exists(path):
        return {}
    with open(path) as f:
        raw = json.load(f)
    result = {}
    for key, vals in raw.items():
        result[int(key.replace('L', ''))] = vals
    return result


def load_delta_heatmaps(output_dir, scale, tag='all_pairs'):
    import glob as glob_mod
    pattern = os.path.join(output_dir, 'csv', f'delta_similarity_{scale}_L*_{tag}.csv')
    files = glob_mod.glob(pattern)
    result = {}
    for fpath in files:
        basename = os.path.basename(fpath)
        # delta_similarity_{scale}_L{layer}_{tag}.csv
        part = basename.replace(f'delta_similarity_{scale}_L', '').replace(f'_{tag}.csv', '')
        try:
            layer = int(part)
        except ValueError:
            continue
        result[layer] = pd.read_csv(fpath, index_col=0)
    return result


# ============================================================================
# Visualization
# ============================================================================

def plot_within_cat_consistency_trajectory(within_cat, scale, model_type, save_path):
    """Plot within-category delta consistency across layers."""
    fig, ax = plt.subplots(figsize=(12, 6))
    cat_colors = CAT_COLORS
    for cat in CATEGORY_ORDER:
        layers, vals = [], []
        for (c, l), v in sorted(within_cat.items(), key=lambda x: x[0][1]):
            if c == cat:
                layers.append(l)
                vals.append(v['mean'])
        if layers:
            ax.plot(layers, vals, '-o', color=cat_colors[cat], label=cat, linewidth=2, markersize=3)
    ax.set_xlabel('Layer Index')
    ax.set_ylabel('Within-Category Consistency')
    ax.set_title(f'{model_type.upper()} ({scale}) - Within-Category Delta Consistency', fontweight='bold')
    ax.legend(fontsize=9)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


def plot_sign_corrected_consistency_trajectory(sign_corrected, scale, model_type, save_path):
    """Plot sign-corrected group consistency across layers."""
    fig, ax = plt.subplots(figsize=(12, 6))
    colors = GROUP_COLORS
    for group in GROUP_ORDER:
        layers, vals = [], []
        for (g, l), v in sorted(sign_corrected.items(), key=lambda x: x[0][1]):
            if g == group:
                layers.append(l)
                vals.append(v['mean'])
        if layers:
            ax.plot(layers, vals, '-o', color=colors[group], label=group, linewidth=2, markersize=3)
    ax.set_xlabel('Layer Index')
    ax.set_ylabel('Sign-Corrected Consistency')
    ax.set_title(f'{model_type.upper()} ({scale}) - Sign-Corrected Group Consistency', fontweight='bold')
    ax.legend(fontsize=11)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


def plot_cross_group_alignment_trajectory(cross_alignment, scale, model_type, save_path):
    fig, ax = plt.subplots(figsize=(12, 6))
    layers = sorted(cross_alignment.keys())
    actual = [cross_alignment[l]['per_sample_mean'] for l in layers]
    mean_delta = [cross_alignment[l]['mean_delta_alignment'] for l in layers]
    perm_mean = [cross_alignment[l]['permutation_mean'] for l in layers]
    perm_std = [cross_alignment[l]['permutation_std'] for l in layers]

    ax.plot(layers, actual, '-o', color='#d62728', label='cos(d_vert, d_dist) per-sample mean',
            linewidth=2.5, markersize=3)
    ax.plot(layers, mean_delta, '--s', color='#e377c2', label='cos(mean_d_vert, mean_d_dist)',
            linewidth=1.5, markersize=3)
    ax.plot(layers, perm_mean, ':', color='gray', label='permutation control', linewidth=1.5)
    ax.fill_between(layers,
                     [m - 2*s for m, s in zip(perm_mean, perm_std)],
                     [m + 2*s for m, s in zip(perm_mean, perm_std)],
                     alpha=0.2, color='gray')
    ax.set_xlabel('Layer Index')
    ax.set_ylabel('Cosine Alignment')
    ax.set_title(f'{model_type.upper()} ({scale}) - Cross-Group Alignment (Perspective Bias)', fontweight='bold')
    ax.legend(fontsize=9)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


# Fix 7: Delta heatmap visualization

def plot_delta_heatmap(sim_df, title, save_path):
    """Plot delta-based similarity heatmap."""
    plt.figure(figsize=(10, 8))
    available_order = [c for c in CATEGORY_ORDER if c in sim_df.index]
    sim_df_ordered = sim_df.loc[available_order, available_order]

    annot = sim_df_ordered.round(4).astype(str)
    sns.heatmap(sim_df_ordered, annot=annot, fmt='', cmap='RdBu_r',
                center=0, vmin=-1, vmax=1, square=True, linewidths=0.5,
                cbar_kws={'label': 'Cosine Similarity'})
    plt.title(title, fontsize=14, fontweight='bold')
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved delta heatmap: {save_path}")


# Fix 6: Prediction stats visualization

def plot_pred_stats_bars(all_pred_stats, model_type, save_path):
    """Bar chart: per-group accuracy (orig/swap/both) across scales."""
    fig, axes = plt.subplots(1, len(GROUP_ORDER), figsize=(7 * len(GROUP_ORDER), 6))
    if len(GROUP_ORDER) == 1:
        axes = [axes]

    available = [s for s in SCALE_ORDER if any(d['scale'] == s for d in all_pred_stats)]
    if not available:
        # Fallback: use whatever scales are present (preserves insertion order)
        seen = []
        for d in all_pred_stats:
            if d['scale'] not in seen:
                seen.append(d['scale'])
        available = seen

    for idx, group in enumerate(GROUP_ORDER):
        ax = axes[idx]
        x = np.arange(3)  # orig, swap, both
        width = 0.8 / len(available)
        for i, scale in enumerate(available):
            entry = next((d for d in all_pred_stats if d['scale'] == scale), None)
            if entry is None:
                continue
            vals = [entry.get(f'{group}_acc_orig', 0),
                    entry.get(f'{group}_acc_swap', 0),
                    entry.get(f'{group}_acc_both', 0)]
            offset = (i - len(available) / 2 + 0.5) * width
            color = SCALE_COLORS.get(scale, 'gray')
            ax.bar(x + offset, vals, width, label=scale, color=color)
        ax.set_xticks(x)
        ax.set_xticklabels(['orig', 'swap', 'both'])
        ax.set_ylabel('Accuracy')
        ax.set_title(group, fontweight='bold')
        ax.legend(fontsize=7)
        ax.set_ylim(0, 1.1)
        ax.axhline(y=0.5, color='gray', linestyle='--', alpha=0.5)
        ax.grid(True, alpha=0.3, axis='y')

    fig.suptitle(f'{model_type.upper()} - Prediction Accuracy by Group', fontsize=15, fontweight='bold', y=1.02)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


def plot_pred_stats_trajectory(all_pred_stats, model_type, save_path):
    """Line plot: acc_both trajectory across scales per group."""
    fig, ax = plt.subplots(figsize=(10, 6))
    available = [s for s in SCALE_ORDER if any(d['scale'] == s for d in all_pred_stats)]
    if not available:
        seen = []
        for d in all_pred_stats:
            if d['scale'] not in seen:
                seen.append(d['scale'])
        available = seen
    colors = GROUP_COLORS

    for group in GROUP_ORDER:
        x_vals, y_vals = [], []
        for i, scale in enumerate(available):
            entry = next((d for d in all_pred_stats if d['scale'] == scale), None)
            if entry:
                x_vals.append(i)
                y_vals.append(entry.get(f'{group}_acc_both', 0))
        if x_vals:
            ax.plot(x_vals, y_vals, '-o', color=colors[group], label=group, linewidth=2.5, markersize=6)

    ax.set_xticks(range(len(available)))
    ax.set_xticklabels(available)
    ax.set_xlabel('Scale')
    ax.set_ylabel('Accuracy (both correct)')
    ax.set_title(f'{model_type.upper()} - Both-Correct Accuracy Across Scales', fontweight='bold')
    ax.legend(fontsize=10)
    ax.set_ylim(0, 1.05)
    ax.axhline(y=0.5, color='gray', linestyle='--', alpha=0.5)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


def plot_pca_embeddings(vectors_npz_path, scale, model_type, save_dir, bc_only=False):
    data = np.load(vectors_npz_path, allow_pickle=True)
    layer_keys = [k for k in data.files if k.startswith('orig_L')]
    layers = sorted([int(k.replace('orig_L', '')) for k in layer_keys])

    cat_colors = CAT_COLORS

    for layer in layers:
        orig   = data.get(f'orig_L{layer}')
        swap   = data.get(f'swap_L{layer}')
        labels = data.get(f'labels_L{layer}')
        deltas = data.get(f'delta_L{layer}')
        cats   = data.get(f'categories_L{layer}')
        groups = data.get(f'groups_L{layer}')

        if bc_only and deltas is not None:
            co = data.get(f'is_correct_orig_L{layer}')
            cs = data.get(f'is_correct_swap_L{layer}')
            if co is not None and cs is not None:
                bc_mask = co.astype(bool) & cs.astype(bool)
                if orig is not None and len(orig) == len(bc_mask):
                    orig   = orig[bc_mask]
                    swap   = swap[bc_mask]
                    labels = labels[bc_mask] if labels is not None else None
                if len(deltas) == len(bc_mask):
                    deltas = deltas[bc_mask]
                    cats   = cats[bc_mask]   if cats   is not None else None
                    groups = groups[bc_mask] if groups is not None else None

        if orig is None or swap is None or len(orig) == 0:
            continue

        fig, axes = plt.subplots(1, 3, figsize=(24, 7))

        pca = PCA(n_components=2)
        all_vecs = np.vstack([orig, swap])
        all_pca = pca.fit_transform(all_vecs)
        orig_pca = all_pca[:len(orig)]
        swap_pca = all_pca[len(orig):]

        ax = axes[0]
        for cat in CATEGORY_ORDER:
            mask = np.array([str(l) == cat for l in labels])
            if mask.any():
                ax.scatter(orig_pca[mask, 0], orig_pca[mask, 1],
                           c=cat_colors.get(cat, 'gray'), label=f'{cat} (orig)',
                           alpha=0.5, s=15, marker='o')
                ax.scatter(swap_pca[mask, 0], swap_pca[mask, 1],
                           c=cat_colors.get(cat, 'gray'),
                           alpha=0.5, s=15, marker='x')
        ax.set_title('Embeddings by Category\n(o=orig, x=swap)', fontsize=11)
        ax.legend(fontsize=7, ncol=2)
        ax.grid(True, alpha=0.2)

        ax = axes[1]
        if deltas is not None and cats is not None:
            pca_d = PCA(n_components=2)
            delta_pca = pca_d.fit_transform(deltas)
            group_colors = GROUP_COLORS
            if groups is not None:
                for group in GROUP_ORDER:
                    mask = np.array([str(g) == group for g in groups])
                    if mask.any():
                        ax.scatter(delta_pca[mask, 0], delta_pca[mask, 1],
                                   c=group_colors.get(group, 'gray'), label=group, alpha=0.5, s=15)
            ax.set_title('Delta Vectors by Group', fontsize=11)
            ax.legend(fontsize=9)
            ax.grid(True, alpha=0.2)

        ax = axes[2]
        if deltas is not None and cats is not None:
            for cat in CATEGORY_ORDER:
                mask = np.array([str(c) == cat for c in cats])
                if mask.any():
                    ax.scatter(delta_pca[mask, 0], delta_pca[mask, 1],
                               c=cat_colors.get(cat, 'gray'), label=cat, alpha=0.5, s=15)
            ax.set_title('Delta Vectors by Category', fontsize=11)
            ax.legend(fontsize=8, ncol=2)
            ax.grid(True, alpha=0.2)

        fig.suptitle(f'{model_type.upper()} ({scale}) - Layer {layer} - PCA', fontweight='bold')
        plt.tight_layout()
        plt.savefig(os.path.join(save_dir, f'pca_{scale}_L{layer}.png'), dpi=200, bbox_inches='tight')
        plt.close()

    logger.info(f"Saved PCA plots to {save_dir}")


def plot_pca_3d(vectors_npz_path, scale, model_type, save_dir, bc_only=False):
    """Generate single-panel 3D PCA figure (Delta Vectors by Category) per layer.

    Figure style follows pca_new.py: larger fonts, single panel, PC3 via fig.text.
    """
    # ── Font / marker sizes (from pca_new.py) ─────────────────────────────────
    _TITLE_FS    = 22
    _AXIS_FS     = 18
    _TICK_FS     = 14
    _LEGEND_FS   = 16
    _SUPTITLE_FS = 24
    _SCATTER_S   = 30

    def _normalise_label(raw):
        return 'below' if str(raw) == 'under' else str(raw)

    def scatter3d(ax, xs, ys, zs, c, label, alpha=0.55, s=_SCATTER_S, marker='o'):
        ax.scatter(xs, ys, zs, c=c, label=label, alpha=alpha, s=s, marker=marker)

    data = np.load(vectors_npz_path, allow_pickle=True)
    layer_keys = [k for k in data.files if k.startswith('orig_L')]
    layers = sorted([int(k.replace('orig_L', '')) for k in layer_keys])

    if not layers:
        logger.info(f"  [pca_3d] No orig_L* keys found in {vectors_npz_path}")
        return

    os.makedirs(save_dir, exist_ok=True)

    for layer in layers:
        deltas = data.get(f'delta_L{layer}')
        cats   = data.get(f'categories_L{layer}')

        # ── Both-correct filtering ────────────────────────────────────────────
        if bc_only and deltas is not None:
            co = data.get(f'is_correct_orig_L{layer}')
            cs = data.get(f'is_correct_swap_L{layer}')
            if co is not None and cs is not None:
                bc_mask = co.astype(bool) & cs.astype(bool)
                if len(deltas) == len(bc_mask):
                    deltas = deltas[bc_mask]
                    cats   = cats[bc_mask] if cats is not None else None

        has_delta = (deltas is not None and len(deltas) >= 3)
        if not has_delta:
            logger.info(f"  [pca_3d] Layer {layer}: no delta vectors, skipping")
            continue

        # ── PCA on delta vectors ──────────────────────────────────────────────
        pca_d = PCA(n_components=3)
        delta_proj = pca_d.fit_transform(deltas)
        ev = pca_d.explained_variance_ratio_

        # ── Figure ────────────────────────────────────────────────────────────
        fig = plt.figure(figsize=(13, 10))
        ax  = fig.add_subplot(111, projection='3d')

        if cats is not None:
            for cat in CATEGORY_ORDER:
                mask = np.array([_normalise_label(c) == cat for c in cats])
                if not mask.any():
                    continue
                scatter3d(ax,
                          delta_proj[mask, 0],
                          delta_proj[mask, 1],
                          delta_proj[mask, 2],
                          c=CAT_COLORS.get(cat, 'gray'),
                          label=cat)

        ax.set_title('Delta Vectors by Category', fontsize=_TITLE_FS, pad=12)
        ax.set_xlabel(f'PC1 ({ev[0]:.1%})', fontsize=_AXIS_FS, labelpad=25)
        ax.set_ylabel(f'PC2 ({ev[1]:.1%})', fontsize=_AXIS_FS, labelpad=25)
        ax.set_zlabel('')
        ax.tick_params(axis='both', labelsize=_TICK_FS)
        ax.legend(fontsize=_LEGEND_FS, ncol=2, loc='upper right')

        # ── Draw to get accurate axes bbox, then place PC3 label ─────────────
        fig.canvas.draw()
        ax_pos = ax.get_position()

        pc3_x = ax_pos.x1 + 0.04
        fig.text(
            pc3_x,
            (ax_pos.y0 + ax_pos.y1) / 2,
            f'PC3 ({ev[2]:.1%})',
            fontsize=_AXIS_FS,
            va='center', ha='center',
            rotation=90,
        )

        ax_cx = (ax_pos.x0 + ax_pos.x1) / 2
        fig.suptitle(
            f'{model_type.upper()} ({scale}) — L{layer}',
            fontsize=_SUPTITLE_FS, fontweight='bold',
            x=ax_cx, y=1.01,
        )

        out_path = os.path.join(save_dir, f'pca_{scale}_L{layer}.png')
        plt.savefig(out_path, dpi=200, bbox_inches='tight', pad_inches=0.5)
        plt.close()

    logger.info(f"Saved 3D PCA plots to {save_dir}")


# Cross-scale plots

def plot_cross_scale_consistency(all_consistency, model_type, save_path, title_prefix='Sign-Corrected'):
    fig, axes = plt.subplots(1, 3, figsize=(21, 6))

    for idx, group in enumerate(GROUP_ORDER):
        ax = axes[idx]
        for scale in SCALE_ORDER:
            if scale not in all_consistency:
                continue
            consistency = all_consistency[scale]
            layers, vals = [], []
            for (g, l), v in sorted(consistency.items(), key=lambda x: x[0][1]):
                if g == group:
                    layers.append(l)
                    vals.append(v['mean'])
            if layers:
                ax.plot(layers, vals, '-', color=SCALE_COLORS.get(scale, 'gray'),
                        label=SCALE_DISPLAY_NAMES.get(scale, scale), linewidth=2)
        ax.set_xlabel('Layer Index')
        ax.set_ylabel('Consistency')
        ax.set_title(group, fontweight='bold')
        ax.legend(fontsize=9)
        ax.grid(True, alpha=0.3)

    fig.suptitle(f'{model_type.upper()} - {title_prefix} Consistency Across Scales',
                 fontsize=15, fontweight='bold', y=1.02)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


def plot_cross_scale_within_cat_consistency(all_within_cat, model_type, save_path):
    """Cross-scale within-category consistency."""
    fig, axes = plt.subplots(2, 3, figsize=(21, 12))

    for idx, cat in enumerate(CATEGORY_ORDER):
        ax = axes[idx // 3][idx % 3]
        for scale in SCALE_ORDER:
            if scale not in all_within_cat:
                continue
            wc = all_within_cat[scale]
            layers, vals = [], []
            for (c, l), v in sorted(wc.items(), key=lambda x: x[0][1]):
                if c == cat:
                    layers.append(l)
                    vals.append(v['mean'])
            if layers:
                ax.plot(layers, vals, '-', color=SCALE_COLORS.get(scale, 'gray'),
                        label=SCALE_DISPLAY_NAMES.get(scale, scale), linewidth=2)
        ax.set_xlabel('Layer Index')
        ax.set_ylabel('Consistency')
        ax.set_title(cat, fontweight='bold')
        ax.legend(fontsize=8)
        ax.grid(True, alpha=0.3)

    fig.suptitle(f'{model_type.upper()} - Within-Category Consistency Across Scales',
                 fontsize=15, fontweight='bold', y=1.02)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


def plot_cross_scale_alignment(all_alignment, model_type, save_path):
    fig, ax = plt.subplots(figsize=(12, 6))
    for scale in SCALE_ORDER:
        if scale not in all_alignment:
            continue
        alignment = all_alignment[scale]
        layers = sorted(alignment.keys())
        vals = [alignment[l]['per_sample_mean'] for l in layers]
        ax.plot(layers, vals, '-', color=SCALE_COLORS.get(scale, 'gray'),
                label=SCALE_DISPLAY_NAMES.get(scale, scale), linewidth=2)
    ax.set_xlabel('Layer Index')
    ax.set_ylabel('cos(d_vert, d_dist)')
    ax.set_title(f'{model_type.upper()} - Cross-Group Alignment Across Scales\n'
                 f'(High=entangled, Low=disentangled)', fontweight='bold')
    ax.legend(fontsize=10)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


# Fix 7: Delta-based trajectory (cross-layer, per-scale)

def plot_delta_trajectory(all_delta_heatmaps, model_type, save_path):
    """Cross-layer trajectory of delta-based similarities for key pairs."""
    pairs = [
        ('above', 'far', 'above-far'), ('below', 'close', 'below-close'),
        ('left', 'right', 'left-right'),
    ]
    fig, axes = plt.subplots(1, len(pairs), figsize=(7 * len(pairs), 6))
    if len(pairs) == 1:
        axes = [axes]

    for idx, (cat1, cat2, label) in enumerate(pairs):
        ax = axes[idx]
        for scale in SCALE_ORDER:
            if scale not in all_delta_heatmaps:
                continue
            hm = all_delta_heatmaps[scale]
            layers = sorted(hm.keys())
            vals = []
            valid_layers = []
            for l in layers:
                df = hm[l]
                if df is not None and cat1 in df.index and cat2 in df.columns:
                    valid_layers.append(l)
                    vals.append(df.loc[cat1, cat2])
            if valid_layers:
                ax.plot(valid_layers, vals, '-', color=SCALE_COLORS.get(scale, 'gray'),
                        label=SCALE_DISPLAY_NAMES.get(scale, scale), linewidth=2)
        ax.set_xlabel('Layer Index')
        ax.set_ylabel('Delta Cosine Similarity')
        ax.set_title(label, fontweight='bold')
        ax.legend(fontsize=9)
        ax.grid(True, alpha=0.3)
        ax.axhline(y=0, color='gray', linestyle='--', alpha=0.5)

    fig.suptitle(f'{model_type.upper()} - Delta-Based Similarity Trajectory',
                 fontsize=15, fontweight='bold', y=1.02)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


def plot_summary_barplot(all_consistency, all_alignment, model_type, save_path):
    available_scales = [s for s in SCALE_ORDER if s in all_consistency]
    if not available_scales:
        return

    sample_cons = all_consistency[available_scales[0]]
    max_layer = max(l for (_, l) in sample_cons.keys())

    fig, axes = plt.subplots(1, 2, figsize=(16, 6))

    ax = axes[0]
    x = np.arange(len(GROUP_ORDER))
    width = 0.8 / len(available_scales)
    for i, scale in enumerate(available_scales):
        cons = all_consistency[scale]
        vals = [cons.get((g, max_layer), {}).get('mean', 0) for g in GROUP_ORDER]
        offset = (i - len(available_scales) / 2 + 0.5) * width
        ax.bar(x + offset, vals, width,
               label=SCALE_DISPLAY_NAMES.get(scale, scale),
               color=SCALE_COLORS.get(scale, 'gray'))
    ax.set_xticks(x)
    ax.set_xticklabels(GROUP_ORDER)
    ax.set_ylabel('Sign-Corrected Consistency')
    ax.set_title(f'Consistency at Layer {max_layer}', fontweight='bold')
    ax.legend(fontsize=8)
    ax.grid(True, alpha=0.3, axis='y')

    ax = axes[1]
    available_align = [s for s in available_scales if s in all_alignment]
    if available_align:
        vals = [all_alignment[s].get(max_layer, {}).get('per_sample_mean', 0) for s in available_align]
        colors = [SCALE_COLORS.get(s, 'gray') for s in available_align]
        ax.bar(range(len(vals)), vals, color=colors)
        ax.set_xticks(range(len(vals)))
        ax.set_xticklabels([SCALE_DISPLAY_NAMES.get(s, s) for s in available_align])
        ax.set_ylabel('cos(d_vert, d_dist)')
        ax.set_title(f'Cross-Group Alignment at L{max_layer}\n(Lower=disentangled)', fontweight='bold')
        ax.grid(True, alpha=0.3, axis='y')

    fig.suptitle(f'{model_type.upper()} - Summary at Deepest Layer', fontsize=15, fontweight='bold', y=1.02)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {save_path}")


# ============================================================================
# Main Pipeline
# ============================================================================

def process_scale(args, scale, swap_pairs, quads):
    # Resolve model path from the correct config dict
    if args.model_type in MODEL_CONFIGS_NEW:
        cls_name, model_path = MODEL_CONFIGS_NEW[args.model_type][scale]
    else:
        model_path = MODEL_CONFIGS[args.model_type][scale]
        cls_name   = None

    logger.info(f"\n{'='*60}")
    logger.info(f"Processing {args.model_type} - {scale}"
                + (f"  [{cls_name}]" if cls_name else ""))
    logger.info(f"Model path: {model_path}")
    logger.info(f"{'='*60}")

    extractor = get_extractor(args.model_type, model_path, scale=scale, device=args.device)
    target_layers = extractor.target_layers

    vlm_key = get_model_key(args.model_type, scale)
    output_dir = os.path.join(args.output_dir, vlm_key)
    plots_dir = os.path.join(output_dir, 'plots')
    os.makedirs(plots_dir, exist_ok=True)

    # ── Phase A: Extract swap pair features ───────────────────────────────────
    logger.info("\n--- Phase A: Extracting swap pair features ---")
    swap_records = extract_swap_features(extractor, swap_pairs,
                                         max_samples_per_category=args.max_samples_per_category)

    # ── Phase C_A: Analysis of swap-pair data ─────────────────────────────────
    logger.info("\n--- Phase C_A: Analysis (swap pairs) ---")

    # Fix 8: Category validity check
    category_validity = check_category_validity(swap_records, scale)
    unreliable_cats = [c for c, v in category_validity.items() if not v['reliable']]
    if unreliable_cats:
        logger.warning(f"  Unreliable categories: {unreliable_cats}")

    # Fix 5: Two types of consistency (all pairs)
    within_cat_all, sign_corrected_all = compute_delta_consistency(swap_records, target_layers)

    # Fix 8: Both-correct filtered consistency
    both_correct_records = filter_both_correct(swap_records)
    logger.info(f"  Both-correct pairs: {len(both_correct_records)}/{len(swap_records)}")
    within_cat_bc, sign_corrected_bc = compute_delta_consistency(both_correct_records, target_layers)

    pred_stats = compute_prediction_stats(swap_records, scale)

    # Fix 7: Delta-based heatmaps (for all layers)
    delta_heatmaps_all = {}
    delta_heatmaps_bc  = {}
    delta_norms_all    = {}
    delta_norms_bc     = {}
    for layer in target_layers:
        delta_heatmaps_all[layer] = compute_delta_similarity_matrix(swap_records, layer)
        delta_norms_all[layer]    = compute_delta_norm_per_category(swap_records, layer)
        if both_correct_records:
            delta_heatmaps_bc[layer] = compute_delta_similarity_matrix(both_correct_records, layer)
            delta_norms_bc[layer]    = compute_delta_norm_per_category(both_correct_records, layer)


    # Log Phase A key results
    max_layer = max(target_layers)
    for group in GROUP_ORDER:
        key = (group, max_layer)
        if key in sign_corrected_all:
            logger.info(f"  Sign-corrected [{group}, L{max_layer}]: "
                        f"{sign_corrected_all[key]['mean']:.4f} +/- {sign_corrected_all[key]['std']:.4f}")
    logger.info(f"  Accuracy orig={pred_stats['overall_acc_orig']:.1%}, "
                f"swap={pred_stats['overall_acc_swap']:.1%}, "
                f"both={pred_stats['overall_acc_both']:.1%}")

    # ── Phase D_A: Save Phase A results ───────────────────────────────────────
    logger.info("\n--- Phase D_A: Saving Phase A results ---")

    save_vectors_npz(scale, swap_records, target_layers, output_dir)

    save_scale_results(
        scale, swap_records, [],
        within_cat_all, sign_corrected_all,
        {}, pred_stats, target_layers,
        category_validity, delta_heatmaps_all,
        output_dir, both_correct_tag='all_pairs',
        save_alignment=False,
        delta_norms=delta_norms_all,
    )
    if both_correct_records:
        save_scale_results(
            scale, both_correct_records, [],
            within_cat_bc, sign_corrected_bc,
            {}, pred_stats, target_layers,
            category_validity, delta_heatmaps_bc,
            output_dir, both_correct_tag='both_correct',
            save_alignment=False,
            delta_norms=delta_norms_bc,
        )

    # ── Phase E_A: Per-scale plots from Phase A data ───────────────────────────
    if args.phase1_only:
        logger.info("\n--- Phase E_A: Per-scale plots [SKIPPED: --phase1-only] ---")
    else:
        logger.info("\n--- Phase E_A: Per-scale plots (swap-pair data) ---")

        for condition, wc_data, sc_data in [
            ('all',          within_cat_all, sign_corrected_all),
            ('both_correct', within_cat_bc,  sign_corrected_bc),
        ]:
            if condition == 'both_correct' and not both_correct_records:
                continue

            cond_dir = os.path.join(plots_dir, condition)
            os.makedirs(cond_dir, exist_ok=True)

            wc_dir = os.path.join(cond_dir, 'within_cat_consistency')
            sc_dir = os.path.join(cond_dir, 'sign_corrected')
            os.makedirs(wc_dir, exist_ok=True)
            os.makedirs(sc_dir, exist_ok=True)

            plot_within_cat_consistency_trajectory(
                wc_data, scale, args.model_type,
                os.path.join(wc_dir, f'within_cat_consistency_{scale}.png'))

            plot_sign_corrected_consistency_trajectory(
                sc_data, scale, args.model_type,
                os.path.join(sc_dir, f'sign_corrected_consistency_{scale}.png'))

        # PCA (from full NPZ) — 2D and 3D, all-pairs and both-correct
        npz_path = os.path.join(output_dir, 'npz', f'vectors_{scale}.npz')
        if os.path.exists(npz_path):
            pca_dir = os.path.join(plots_dir, 'all', 'pca')
            pca_3d_dir = os.path.join(plots_dir, 'all', 'pca_3d')
            bc_pca_dir = os.path.join(plots_dir, 'both_correct', 'pca')
            bc_pca_3d_dir = os.path.join(plots_dir, 'both_correct', 'pca_3d')
            for d in (pca_dir, pca_3d_dir, bc_pca_dir, bc_pca_3d_dir):
                os.makedirs(d, exist_ok=True)
            plot_pca_embeddings(npz_path, scale, args.model_type, pca_dir)
            plot_pca_3d(npz_path, scale, args.model_type, pca_3d_dir)
            plot_pca_embeddings(npz_path, scale, args.model_type, bc_pca_dir, bc_only=True)
            plot_pca_3d(npz_path, scale, args.model_type, bc_pca_3d_dir, bc_only=True)

        if pred_stats:
            pred_plot_dir = os.path.join(plots_dir, 'all', 'pred_stats')
            os.makedirs(pred_plot_dir, exist_ok=True)
            plot_pred_stats_bars([pred_stats], args.model_type,
                                 os.path.join(pred_plot_dir, f'pred_stats_{scale}.png'))

        if pred_stats:
            acc_dir = os.path.join(output_dir, 'accuracy')
            logger.info(f"\n--- Accuracy Charts [{scale}] ---")
            run_accuracy_charts([pred_stats], {scale: category_validity}, args.model_type, acc_dir)

        logger.info(f"\n--- All-Layer Heatmaps [{scale}] ---")
        run_all_layer_heatmaps(output_dir, args.model_type, [scale])
        logger.info(f"\n--- All-Layer PCA [{scale}] ---")
        run_all_layer_pca(output_dir, args.model_type, [scale])

    # ── Phase B: Extract cross-group features ─────────────────────────────────
    skip_b = getattr(args, 'skip_phase_b', False)
    if skip_b or not quads:
        if skip_b:
            logger.info("\n--- Phase B: Cross-group extraction [SKIPPED: --skip-phase-b] ---")
        quad_records = []
        cross_alignment = {}
    else:
        logger.info("\n--- Phase B: Extracting cross-group features ---")
        quad_records = extract_cross_group_features(extractor, quads)

        # ── Phase C_B: Cross-group analysis ───────────────────────────────────
        logger.info("\n--- Phase C_B: Analysis (cross-group) ---")
        cross_alignment = compute_cross_group_alignment(quad_records, target_layers)

        if max_layer in cross_alignment:
            ca = cross_alignment[max_layer]
            logger.info(f"  Cross-group alignment L{max_layer}: "
                        f"{ca['per_sample_mean']:.4f} (perm={ca['permutation_mean']:.4f})")

        # ── Phase D_B: Save Phase B results ───────────────────────────────────
        logger.info("\n--- Phase D_B: Saving Phase B results ---")
        save_cross_group_npz(scale, quad_records, target_layers, output_dir)
        save_cross_alignment(scale, cross_alignment, output_dir)

        # ── Phase E_B: Cross-alignment plots ──────────────────────────────────
        if args.phase1_only:
            logger.info("\n--- Phase E_B: Cross-alignment plots [SKIPPED: --phase1-only] ---")
        else:
            logger.info("\n--- Phase E_B: Per-scale plots (cross-group data) ---")
            for condition in ['all', 'both_correct']:
                if condition == 'both_correct' and not both_correct_records:
                    continue
                ca_dir = os.path.join(plots_dir, condition, 'cross_alignment')
                os.makedirs(ca_dir, exist_ok=True)
                plot_cross_group_alignment_trajectory(
                    cross_alignment, scale, args.model_type,
                    os.path.join(ca_dir, f'cross_alignment_{scale}.png'))

    # Cleanup
    del swap_records, quad_records, both_correct_records
    extractor.cleanup()

    logger.info(f"\n  Scale {scale} complete.")


# ============================================================================
# Accuracy Chart (integrated from accuracy_chart.py)
# ============================================================================

def _acc_plot_group_bars(pred_stats, model_type, ax_list):
    available = [s for s in SCALE_ORDER if any(d['scale'] == s for d in pred_stats)]
    x = np.arange(3)
    width = 0.8 / max(len(available), 1)
    for idx, group in enumerate(GROUP_ORDER):
        ax = ax_list[idx]
        for i, scale in enumerate(available):
            entry = next((d for d in pred_stats if d['scale'] == scale), None)
            if entry is None:
                continue
            vals = [entry.get(f'{group}_acc_orig', 0),
                    entry.get(f'{group}_acc_swap', 0),
                    entry.get(f'{group}_acc_both', 0)]
            offset = (i - len(available) / 2 + 0.5) * width
            ax.bar(x + offset, vals, width, label=scale,
                   color=SCALE_COLORS.get(scale, 'gray'), alpha=0.85)
        ax.set_xticks(x)
        ax.set_xticklabels(['orig', 'swap', 'both'], fontsize=10)
        ax.set_ylabel('Accuracy', fontsize=9)
        ax.set_title(group.capitalize(), fontweight='bold', fontsize=11,
                     color=GROUP_COLORS.get(group, 'black'))
        ax.legend(fontsize=7, ncol=2)
        ax.set_ylim(0, 1.15)
        ax.axhline(y=0.5, color='gray', linestyle='--', alpha=0.5, linewidth=1)
        ax.grid(True, alpha=0.3, axis='y')


def _acc_plot_both_trajectory(pred_stats, model_type, ax):
    available = [s for s in SCALE_ORDER if any(d['scale'] == s for d in pred_stats)]
    x_ticks = range(len(available))
    for group in GROUP_ORDER:
        y_vals = [next((d for d in pred_stats if d['scale'] == s), {}).get(f'{group}_acc_both', 0)
                  for s in available]
        ax.plot(x_ticks, y_vals, '-o', color=GROUP_COLORS.get(group, 'gray'),
                label=group, linewidth=2.5, markersize=7)
    y_overall = [next((d for d in pred_stats if d['scale'] == s), {}).get('overall_acc_both', 0)
                 for s in available]
    ax.plot(x_ticks, y_overall, '--s', color='black', label='overall',
            linewidth=2, markersize=6, alpha=0.7)
    ax.set_xticks(list(x_ticks))
    ax.set_xticklabels(available, fontsize=9)
    ax.set_xlabel('Scale', fontsize=9)
    ax.set_ylabel('Accuracy (both correct)', fontsize=9)
    ax.set_title('Both-Correct Accuracy Trajectory', fontweight='bold', fontsize=11)
    ax.legend(fontsize=9)
    ax.set_ylim(0, 1.05)
    ax.axhline(y=0.5, color='gray', linestyle='--', alpha=0.5, linewidth=1)
    ax.grid(True, alpha=0.3)


def _acc_plot_overall_trajectory(pred_stats, model_type, ax):
    available = [s for s in SCALE_ORDER if any(d['scale'] == s for d in pred_stats)]
    x_ticks = range(len(available))
    for metric, label, ls in [
        ('overall_acc_orig', 'orig', '-o'),
        ('overall_acc_swap', 'swap', '-s'),
        ('overall_acc_both', 'both', '-^'),
    ]:
        y_vals = [next((d for d in pred_stats if d['scale'] == s), {}).get(metric, 0)
                  for s in available]
        ax.plot(x_ticks, y_vals, ls, label=label, linewidth=2.2, markersize=6)
    ax.set_xticks(list(x_ticks))
    ax.set_xticklabels(available, fontsize=9)
    ax.set_xlabel('Scale', fontsize=9)
    ax.set_ylabel('Overall Accuracy', fontsize=9)
    ax.set_title('Overall Accuracy Trajectory', fontweight='bold', fontsize=11)
    ax.legend(fontsize=9)
    ax.set_ylim(0, 1.05)
    ax.axhline(y=0.5, color='gray', linestyle='--', alpha=0.5, linewidth=1)
    ax.grid(True, alpha=0.3)


def _acc_plot_category_accuracy(cat_validity, model_type, ax_orig, ax_swap, pred_stats=None):
    available = [s for s in SCALE_ORDER if s in cat_validity]
    cats_with_overall = CATEGORY_ORDER + ['overall']
    x = np.arange(len(cats_with_overall))
    width = 0.8 / max(len(available), 1)
    overall_key = {'acc_orig': 'overall_acc_orig', 'acc_swap': 'overall_acc_swap'}
    for ax, metric, title in [
        (ax_orig, 'acc_orig', 'Per-Category Accuracy (orig)'),
        (ax_swap, 'acc_swap', 'Per-Category Accuracy (swap)'),
    ]:
        for i, scale in enumerate(available):
            cv = cat_validity[scale]
            vals = [cv.get(cat, {}).get(metric, 0) for cat in CATEGORY_ORDER]
            if pred_stats is not None:
                entry = next((d for d in pred_stats if d['scale'] == scale), None)
                vals.append(entry.get(overall_key[metric], 0) if entry else 0)
            else:
                vals.append(0)
            offset = (i - len(available) / 2 + 0.5) * width
            ax.bar(x + offset, vals, width, label=scale,
                   color=SCALE_COLORS.get(scale, 'gray'), alpha=0.85)
        for j, cat in enumerate(CATEGORY_ORDER):
            ax.axvspan(j - 0.45, j + 0.45, color=CAT_COLORS.get(cat, 'gray'), alpha=0.06, linewidth=0)
        ax.axvline(x=len(CATEGORY_ORDER) - 0.5, color='black', linewidth=1.2, linestyle=':', alpha=0.6)
        ax.set_xticks(x)
        ax.set_xticklabels(cats_with_overall, fontsize=9, rotation=15)
        ax.set_ylabel('Accuracy', fontsize=9)
        ax.set_title(title, fontweight='bold', fontsize=11)
        ax.legend(fontsize=7, ncol=2)
        ax.set_ylim(0, 1.15)
        ax.axhline(y=0.5, color='gray', linestyle='--', alpha=0.5, linewidth=1)
        ax.grid(True, alpha=0.3, axis='y')
        if available:
            last_cv = cat_validity[available[-1]]
            for j, cat in enumerate(CATEGORY_ORDER):
                if not last_cv.get(cat, {}).get('reliable', True):
                    ax.text(j, 1.08, '✗', ha='center', va='center',
                            fontsize=9, color='red', fontweight='bold')


def _acc_plot_category_per_scale(cat_validity, model_type, save_dir, pred_stats=None):
    cats_with_overall = CATEGORY_ORDER + ['overall']
    overall_key = {'acc_orig': 'overall_acc_orig', 'acc_swap': 'overall_acc_swap'}
    for scale in sorted(cat_validity.keys(),
                        key=lambda s: SCALE_ORDER.index(s) if s in SCALE_ORDER else 99):
        cv = cat_validity[scale]
        ps_entry = next((d for d in pred_stats if d['scale'] == scale), None) if pred_stats else None
        fig, axes = plt.subplots(1, 2, figsize=(16, 5))
        x = np.arange(len(cats_with_overall))
        width = 0.55
        for ax, metric, title in [
            (axes[0], 'acc_orig', f'acc_orig  ({scale})'),
            (axes[1], 'acc_swap', f'acc_swap  ({scale})'),
        ]:
            vals = [cv.get(cat, {}).get(metric, 0) for cat in CATEGORY_ORDER]
            overall_val = ps_entry.get(overall_key[metric], 0) if ps_entry else 0
            vals.append(overall_val)
            colors = [CAT_COLORS.get(cat, 'gray') for cat in CATEGORY_ORDER] + ['#333333']
            bars = ax.bar(x, vals, width, color=colors, alpha=0.85, edgecolor='white')
            ax.axvline(x=len(CATEGORY_ORDER) - 0.5, color='black',
                       linewidth=1.2, linestyle=':', alpha=0.6)
            ax.set_xticks(x)
            ax.set_xticklabels(cats_with_overall, fontsize=10)
            ax.set_ylabel('Accuracy', fontsize=10)
            ax.set_title(title, fontweight='bold', fontsize=12)
            ax.set_ylim(0, 1.15)
            ax.axhline(y=0.5, color='gray', linestyle='--', alpha=0.5)
            ax.grid(True, alpha=0.3, axis='y')
            for bar, cat in zip(bars, cats_with_overall):
                reliable = cv.get(cat, {}).get('reliable', True) if cat != 'overall' else True
                h = bar.get_height()
                ax.text(bar.get_x() + bar.get_width() / 2, h + 0.02,
                        f'{h:.2f}' + ('' if reliable else ' ✗'),
                        ha='center', va='bottom', fontsize=8,
                        color='red' if not reliable else 'black')
        fig.suptitle(f'{model_type.upper()} - Category Accuracy  ({scale})',
                     fontsize=13, fontweight='bold')
        plt.tight_layout()
        out = os.path.join(save_dir, f'category_accuracy_{scale}.png')
        plt.savefig(out, dpi=200, bbox_inches='tight')
        plt.close()
        logger.info(f"Saved: {out}")


def run_accuracy_charts(pred_stats, cat_validity, model_type, save_dir):
    """Generate all accuracy chart plots into save_dir."""
    os.makedirs(save_dir, exist_ok=True)

    # Group bars
    fig, axes = plt.subplots(1, 3, figsize=(21, 6))
    _acc_plot_group_bars(pred_stats, model_type, axes)
    fig.suptitle(f'{model_type.upper()} - Prediction Accuracy by Group',
                 fontsize=15, fontweight='bold')
    plt.tight_layout()
    plt.savefig(os.path.join(save_dir, 'accuracy_group_bars.png'), dpi=200, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {os.path.join(save_dir, 'accuracy_group_bars.png')}")

    # Trajectory
    fig, axes = plt.subplots(1, 2, figsize=(16, 6))
    _acc_plot_both_trajectory(pred_stats, model_type, axes[0])
    _acc_plot_overall_trajectory(pred_stats, model_type, axes[1])
    fig.suptitle(f'{model_type.upper()} - Accuracy Trajectory Across Scales',
                 fontsize=14, fontweight='bold')
    plt.tight_layout()
    plt.savefig(os.path.join(save_dir, 'accuracy_trajectory.png'), dpi=200, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {os.path.join(save_dir, 'accuracy_trajectory.png')}")

    if cat_validity:
        # Category bars (all scales overlay)
        fig, axes = plt.subplots(1, 2, figsize=(20, 6))
        _acc_plot_category_accuracy(cat_validity, model_type, axes[0], axes[1],
                                    pred_stats=pred_stats)
        fig.suptitle(f'{model_type.upper()} - Per-Category Accuracy Across Scales',
                     fontsize=14, fontweight='bold')
        plt.tight_layout()
        plt.savefig(os.path.join(save_dir, 'accuracy_category.png'), dpi=200, bbox_inches='tight')
        plt.close()
        logger.info(f"Saved: {os.path.join(save_dir, 'accuracy_category.png')}")

        # Per-scale category bars
        _acc_plot_category_per_scale(cat_validity, model_type, save_dir, pred_stats=pred_stats)

    # Combined accuracy_chart.png
    fig = plt.figure(figsize=(24, 14))
    ax_h = fig.add_subplot(3, 3, 1)
    ax_v = fig.add_subplot(3, 3, 2)
    ax_d = fig.add_subplot(3, 3, 3)
    _acc_plot_group_bars(pred_stats, model_type, [ax_h, ax_v, ax_d])
    ax_tb = fig.add_subplot(3, 3, 4)
    ax_to = fig.add_subplot(3, 3, 5)
    _acc_plot_both_trajectory(pred_stats, model_type, ax_tb)
    _acc_plot_overall_trajectory(pred_stats, model_type, ax_to)
    ax_note = fig.add_subplot(3, 3, 6)
    ax_note.axis('off')
    available_scales = [s for s in SCALE_ORDER if any(d['scale'] == s for d in pred_stats)]
    ax_note.text(0.1, 0.6,
                 f'Scales: {", ".join(available_scales)}\n\n✗ = unreliable category\n-- = 0.5 chance level',
                 transform=ax_note.transAxes, fontsize=11, va='top', family='monospace')
    if cat_validity:
        ax_co = fig.add_subplot(3, 2, 5)
        ax_cs = fig.add_subplot(3, 2, 6)
        _acc_plot_category_accuracy(cat_validity, model_type, ax_co, ax_cs, pred_stats=pred_stats)
    fig.suptitle(f'{model_type.upper()} — Accuracy Summary',
                 fontsize=17, fontweight='bold', y=1.01)
    plt.tight_layout()
    plt.savefig(os.path.join(save_dir, 'accuracy_chart.png'), dpi=200, bbox_inches='tight')
    plt.close()
    logger.info(f"Saved: {os.path.join(save_dir, 'accuracy_chart.png')}")


# ============================================================================
# Unify Consistency Y-axis (integrated from unify_consistency_ylim.py)
# ============================================================================

def _ylim_compute(all_vals, margin_ratio=0.08):
    if not all_vals:
        return -1, 1
    ymin, ymax = min(all_vals), max(all_vals)
    margin = (ymax - ymin) * margin_ratio
    return ymin - margin, ymax + margin


def _ylim_load_keyed_json(path):
    if not os.path.exists(path):
        return None
    with open(path) as f:
        raw = json.load(f)
    if not raw:
        return None
    result = {}
    for key, vals in raw.items():
        parts = key.rsplit('_L', 1)
        if len(parts) == 2:
            result[(parts[0], int(parts[1]))] = vals
    return result if result else None


def _ylim_load_alignment_json(path):
    if not os.path.exists(path):
        return None
    with open(path) as f:
        raw = json.load(f)
    if not raw:
        return None
    result = {int(k[1:]): v for k, v in raw.items() if k.startswith('L')}
    return result if result else None


def _ylim_plot_sign_corrected(data, scale, model_type, save_path, ylim):
    fig, ax = plt.subplots(figsize=(12, 6))
    for group in GROUP_ORDER:
        layers, vals = [], []
        for (g, l), v in sorted(data.items(), key=lambda x: x[0][1]):
            if g == group:
                layers.append(l)
                vals.append(v['mean'])
        if layers:
            ax.plot(layers, vals, '-o', color=GROUP_COLORS[group],
                    label=group, linewidth=2, markersize=3)
    ax.set_ylim(ylim)
    ax.set_xlabel('Layer Index')
    ax.set_ylabel('Sign-Corrected Consistency')
    ax.set_title(f'{model_type.upper()} ({scale}) - Sign-Corrected Group Consistency',
                 fontweight='bold')
    ax.legend(fontsize=11)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()


def _ylim_plot_within_cat(data, scale, model_type, save_path, ylim):
    fig, ax = plt.subplots(figsize=(12, 6))
    for cat in CATEGORY_ORDER:
        layers, vals = [], []
        for (c, l), v in sorted(data.items(), key=lambda x: x[0][1]):
            if c == cat:
                layers.append(l)
                vals.append(v['mean'])
        if layers:
            ax.plot(layers, vals, '-o', color=CAT_COLORS[cat],
                    label=cat, linewidth=2, markersize=3)
    ax.set_ylim(ylim)
    ax.set_xlabel('Layer Index')
    ax.set_ylabel('Within-Category Consistency')
    ax.set_title(f'{model_type.upper()} ({scale}) - Within-Category Delta Consistency',
                 fontweight='bold')
    ax.legend(fontsize=9)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()


def _ylim_plot_cross_alignment(data, scale, model_type, save_path, ylim):
    fig, ax = plt.subplots(figsize=(12, 6))
    layers = sorted(data.keys())
    ax.plot(layers, [data[l]['per_sample_mean'] for l in layers], '-o', color='#d62728',
            label='cos(d_vert, d_dist) per-sample mean', linewidth=2.5, markersize=3)
    ax.plot(layers, [data[l]['mean_delta_alignment'] for l in layers], '--s', color='#e377c2',
            label='cos(mean_d_vert, mean_d_dist)', linewidth=1.5, markersize=3)
    perm_mean = [data[l]['permutation_mean'] for l in layers]
    perm_std  = [data[l]['permutation_std']  for l in layers]
    ax.plot(layers, perm_mean, ':', color='gray', label='permutation control', linewidth=1.5)
    ax.fill_between(layers,
                    [m - 2*s for m, s in zip(perm_mean, perm_std)],
                    [m + 2*s for m, s in zip(perm_mean, perm_std)],
                    alpha=0.2, color='gray')
    ax.set_ylim(ylim)
    ax.set_xlabel('Layer Index')
    ax.set_ylabel('Cosine Alignment')
    ax.set_title(f'{model_type.upper()} ({scale}) - Cross-Group Alignment (Perspective Bias)',
                 fontweight='bold')
    ax.legend(fontsize=9)
    ax.grid(True, alpha=0.3)
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()


def _ylim_process_plot_type(scale_dir_map, plots_dir, conditions, model_type,
                             plot_name, json_pattern, loader, val_gatherer, plotter,
                             subfolder=None):
    """Re-plot one plot type across all conditions with a unified y-axis.

    scale_dir_map: {scale: path_to_saved_data/vlm_key}  (ordered dict)
    """
    logger.info(f"  [unify ylim] {plot_name}")
    for condition, condition_tag in conditions:
        cond_plot_dir = os.path.join(plots_dir, condition)
        if not os.path.isdir(cond_plot_dir):
            continue
        save_dir = os.path.join(cond_plot_dir, subfolder) if subfolder else cond_plot_dir
        os.makedirs(save_dir, exist_ok=True)
        all_data = {}
        for scale, scale_dir in scale_dir_map.items():
            path = os.path.join(scale_dir, 'json',
                                json_pattern.format(scale=scale, tag=condition_tag))
            loaded = loader(path)
            if loaded:
                all_data[scale] = loaded
        if not all_data:
            continue
        all_vals = val_gatherer(all_data)
        ylim = _ylim_compute(all_vals)
        for scale, data in all_data.items():
            save_path = os.path.join(save_dir, f'{plot_name}_{scale}.png')
            plotter(data, scale, model_type, save_path, ylim)
        logger.info(f"    {condition}: y=[{ylim[0]:.4f}, {ylim[1]:.4f}], {len(all_data)} scales")


def run_unify_ylim(scale_dir_map: dict, plots_dir: str, model_type: str):
    """Unify y-axis for sign_corrected, within_cat, and cross_alignment plots.

    scale_dir_map: {scale: path_to_saved_data/vlm_key}
    plots_dir: compare/{group_name}/plots/ (where unified plots are written)
    """
    conditions = [
        ('all',          'all_pairs'),
        ('both_correct', 'both_correct'),
    ]

    def gather_keyed(all_data):
        return [v['mean'] for data in all_data.values() for v in data.values()]

    def gather_alignment(all_data):
        vals = []
        for data in all_data.values():
            for v in data.values():
                vals += [v['per_sample_mean'], v['mean_delta_alignment'],
                         v['permutation_mean'] + 2 * v['permutation_std'],
                         v['permutation_mean'] - 2 * v['permutation_std']]
        return vals

    _ylim_process_plot_type(
        scale_dir_map, plots_dir, conditions, model_type,
        plot_name='sign_corrected_consistency',
        json_pattern='sign_corrected_consistency_{scale}_{tag}.json',
        loader=_ylim_load_keyed_json,
        val_gatherer=gather_keyed,
        plotter=_ylim_plot_sign_corrected,
        subfolder='sign_corrected',
    )
    _ylim_process_plot_type(
        scale_dir_map, plots_dir, conditions, model_type,
        plot_name='within_cat_consistency',
        json_pattern='within_cat_consistency_{scale}_{tag}.json',
        loader=_ylim_load_keyed_json,
        val_gatherer=gather_keyed,
        plotter=_ylim_plot_within_cat,
        subfolder='within_cat_consistency',
    )
    _ylim_process_plot_type(
        scale_dir_map, plots_dir, conditions, model_type,
        plot_name='cross_alignment',
        json_pattern='cross_alignment_{scale}.json',
        loader=_ylim_load_alignment_json,
        val_gatherer=gather_alignment,
        plotter=_ylim_plot_cross_alignment,
        subfolder='cross_alignment',
    )


def _has_phase_b_data(scale_dir: str, scale: str) -> bool:
    """Return True if cross_alignment_{scale}.json exists and is non-empty.

    Used in merge to auto-detect whether Phase B was run for a given scale.
    """
    path = os.path.join(scale_dir, 'json', f'cross_alignment_{scale}.json')
    if not os.path.exists(path):
        return False
    try:
        with open(path) as f:
            data = json.load(f)
        return bool(data)
    except Exception:
        return False


def _check_merge_only_sources(output_dir: str, model_type: str) -> bool:
    """Verify required source directories have data for a merge-only model_type.

    With the new per-scale directory layout, data lives at
    {output_dir}/{model_type}_{scale}/ instead of {output_dir}/{model_type}/.
    Returns True if all sources look healthy, False (with warnings) if not.
    """
    mc = MERGE_ONLY_CONFIGS[model_type]
    ok = True
    for req_model_type in mc['required_dirs']:
        # Look for any saved_data/{req_model_type}_{scale}/ directories
        if not os.path.isdir(output_dir):
            logger.warning(
                f"[{model_type}] output_dir not found: {output_dir}\n"
                f"  → Run inference first: python swap_analysis.py --model_type {req_model_type}"
            )
            ok = False
            continue
        matching = [
            d for d in os.listdir(output_dir)
            if d.startswith(f'{req_model_type}_')
            and os.path.isdir(os.path.join(output_dir, d))
        ]
        if not matching:
            logger.warning(
                f"[{model_type}] No '{req_model_type}_*' directories found in {output_dir}\n"
                f"  → Run inference first: python swap_analysis.py --model_type {req_model_type}"
            )
            ok = False
            continue
        any_data = False
        for d in matching:
            json_dir = os.path.join(output_dir, d, 'json')
            if os.path.isdir(json_dir) and any(
                f.startswith('pred_stats_') for f in os.listdir(json_dir)
            ):
                scale = d[len(req_model_type) + 1:]  # strip "{req_model_type}_" prefix
                logger.info(f"  [{req_model_type}/{scale}] found data in {d}/")
                any_data = True
        if not any_data:
            logger.warning(
                f"[{model_type}] No pred_stats JSON found in any '{req_model_type}_*' directory.\n"
                f"  → Inference may not have completed for '{req_model_type}'."
            )
            ok = False
    return ok


def _load_scale_data_multi(output_dir: str, model_type: str, scale: str, scale_sources: dict):
    """Load per-scale data for one scale, looking in the correct source directory.

    With the new per-scale layout, data lives at {output_dir}/{src_model_type}_{scale}/.
    Returns (sc, sc_bc, wc, wc_bc, align, pred_stat, cat_validity, dh, dh_bc).
    Any unavailable item is None / {}.
    """
    src_model_type = scale_sources.get(scale, model_type)
    src_dir = os.path.join(output_dir, get_model_key(src_model_type, scale))

    sc    = load_scale_consistency(src_dir, scale, 'all_pairs')
    sc_bc = load_scale_consistency(src_dir, scale, 'both_correct')
    wc    = load_within_cat_consistency(src_dir, scale, 'all_pairs')
    wc_bc = load_within_cat_consistency(src_dir, scale, 'both_correct')
    align = load_scale_alignment(src_dir, scale)

    pred_stat = None
    pred_path = os.path.join(src_dir, 'json', f'pred_stats_{scale}.json')
    if os.path.exists(pred_path):
        with open(pred_path) as f:
            pred_stat = json.load(f)

    cat_validity = None
    cv_path = os.path.join(src_dir, 'json', f'category_validity_{scale}.json')
    if os.path.exists(cv_path):
        with open(cv_path) as f:
            cat_validity = json.load(f)

    dh    = load_delta_heatmaps(src_dir, scale, 'all_pairs')
    dh_bc = load_delta_heatmaps(src_dir, scale, 'both_correct')

    return sc, sc_bc, wc, wc_bc, align, pred_stat, cat_validity, dh, dh_bc


# ---------------------------------------------------------------------------
# All-layer heatmap + PCA helpers (called from run_merge / run_merge_extended)
# ---------------------------------------------------------------------------

def _get_csv_layers(csv_dir: str, scale: str, tag: str) -> list:
    """Return sorted list of layer indices that have a delta_similarity CSV."""
    import glob as _glob
    pattern = os.path.join(csv_dir, f'delta_similarity_{scale}_L*_{tag}.csv')
    layers = []
    for fpath in _glob.glob(pattern):
        m = re.search(
            rf'delta_similarity_{re.escape(scale)}_L(\d+)_{re.escape(tag)}\.csv$',
            os.path.basename(fpath))
        if m:
            layers.append(int(m.group(1)))
    return sorted(layers)


def run_all_layer_heatmaps(model_dir: str, model_type: str, scales: list):
    """Generate delta-similarity heatmaps for ALL layers from pre-computed CSVs.

    Reads  {model_dir}/csv/delta_similarity_{scale}_L{n}_{tag}.csv
    Writes {model_dir}/plots/all/heatmap/heatmap_{scale}_L{n}.png          (all_pairs)
           {model_dir}/plots/both_correct/heatmap/heatmap_{scale}_L{n}.png (both_correct)

    Skips a scale if the NPZ is missing or any all_pairs CSV is absent
    (indicates inference was not fully completed for that scale).
    """
    TAG_TO_DIR = {
        'all_pairs':    os.path.join(model_dir, 'plots', 'all',          'heatmap'),
        'both_correct': os.path.join(model_dir, 'plots', 'both_correct', 'heatmap'),
    }

    for scale in scales:
        npz_path = os.path.join(model_dir, 'npz', f'vectors_{scale}.npz')
        csv_dir  = os.path.join(model_dir, 'csv')

        if not os.path.exists(npz_path):
            logger.warning(f'  [{model_type}/{scale}] NPZ not found, skipping heatmaps.')
            continue

        data = np.load(npz_path, allow_pickle=True)
        npz_layers = sorted(
            int(k.replace('orig_L', ''))
            for k in data.files if k.startswith('orig_L')
        )
        data.close()

        if not npz_layers:
            logger.warning(f'  [{model_type}/{scale}] No orig_L* keys in NPZ, skipping heatmaps.')
            continue

        csv_layers = _get_csv_layers(csv_dir, scale, 'all_pairs')
        missing = set(npz_layers) - set(csv_layers)
        if missing:
            logger.warning(
                f'  [{model_type}/{scale}] {len(missing)} NPZ layers lack CSVs '
                f'(e.g. L{sorted(missing)[:5]}). Skipping all-layer heatmaps.')
            continue

        for out_dir in TAG_TO_DIR.values():
            os.makedirs(out_dir, exist_ok=True)

        logger.info(f'  [{model_type}/{scale}] Generating heatmaps for {len(npz_layers)} layers...')
        saved = 0
        for layer in npz_layers:
            for tag, out_dir in TAG_TO_DIR.items():
                csv_path = os.path.join(csv_dir, f'delta_similarity_{scale}_L{layer}_{tag}.csv')
                if not os.path.exists(csv_path):
                    continue  # both_correct CSV may be absent for some layers
                df = pd.read_csv(csv_path, index_col=0)
                available = [c for c in CATEGORY_ORDER if c in df.index]
                if not available:
                    continue
                df = df.loc[available, available]
                title = (
                    f'{model_type.upper()} ({scale}) \u2014 Delta Heatmap L{layer} '
                    f'({"both-correct" if tag == "both_correct" else "all pairs"})'
                )
                out_path = os.path.join(out_dir, f'heatmap_{scale}_L{layer}.png')
                plot_delta_heatmap(df, title, out_path)
                saved += 1
        logger.info(f'  [{model_type}/{scale}] Saved {saved} heatmaps')


def run_all_layer_pca(model_dir: str, model_type: str, scales: list):
    """Generate 2D and 3D PCA plots for ALL layers from saved NPZ files.

    Writes {model_dir}/plots/all/pca/pca_{scale}_L{n}.png          (all pairs)
           {model_dir}/plots/all/pca_3d/pca_{scale}_L{n}.png
           {model_dir}/plots/both_correct/pca/pca_{scale}_L{n}.png  (both-correct only)
           {model_dir}/plots/both_correct/pca_3d/pca_{scale}_L{n}.png
    """
    for scale in scales:
        npz_path = os.path.join(model_dir, 'npz', f'vectors_{scale}.npz')
        if not os.path.exists(npz_path):
            logger.warning(f'  [{model_type}/{scale}] NPZ not found, skipping PCA.')
            continue

        # All-pairs PCA
        pca_2d_dir = os.path.join(model_dir, 'plots', 'all', 'pca')
        pca_3d_dir = os.path.join(model_dir, 'plots', 'all', 'pca_3d')
        os.makedirs(pca_2d_dir, exist_ok=True)
        os.makedirs(pca_3d_dir, exist_ok=True)
        logger.info(f'  [{model_type}/{scale}] Generating all-layer 2D PCA...')
        plot_pca_embeddings(npz_path, scale, model_type, pca_2d_dir)
        logger.info(f'  [{model_type}/{scale}] Generating all-layer 3D PCA...')
        plot_pca_3d(npz_path, scale, model_type, pca_3d_dir)

        # Both-correct PCA
        bc_pca_2d_dir = os.path.join(model_dir, 'plots', 'both_correct', 'pca')
        bc_pca_3d_dir = os.path.join(model_dir, 'plots', 'both_correct', 'pca_3d')
        os.makedirs(bc_pca_2d_dir, exist_ok=True)
        os.makedirs(bc_pca_3d_dir, exist_ok=True)
        logger.info(f'  [{model_type}/{scale}] Generating both-correct 2D PCA...')
        plot_pca_embeddings(npz_path, scale, model_type, bc_pca_2d_dir, bc_only=True)
        logger.info(f'  [{model_type}/{scale}] Generating both-correct 3D PCA...')
        plot_pca_3d(npz_path, scale, model_type, bc_pca_3d_dir, bc_only=True)


def run_merge(args):
    # Per-scale data lives in saved_data/{model_type}_{scale}/
    def _scale_dir(scale):
        return os.path.join(args.output_dir, get_model_key(args.model_type, scale))

    # Cross-scale (compare) output: {question_type}/compare/{group_name}/
    group_name = args.group_name or args.model_type
    if args.merge_output_dir:
        merge_out = args.merge_output_dir
    else:
        qt_root = os.path.dirname(args.output_dir.rstrip('/'))   # one level up from saved_data/
        merge_out = os.path.join(qt_root, 'compare', group_name)
    plots_dir = os.path.join(merge_out, 'plots')
    os.makedirs(plots_dir, exist_ok=True)

    scale_order = ['vanilla', '80k', '400k', '800k', '2m', 'roborefer',
                   'roborefer_depth', '10pct', '20pct', '30pct']
    available_scales = [s for s in scale_order if s in args.scales]

    # Load per-scale results
    all_sign_corrected = {}
    all_sign_corrected_bc = {}
    all_within_cat = {}
    all_within_cat_bc = {}
    all_alignment = {}
    all_pred_stats = []
    all_cat_validity = {}
    all_delta_heatmaps = {}
    all_delta_heatmaps_bc = {}

    for scale in available_scales:
        sd = _scale_dir(scale)
        sc = load_scale_consistency(sd, scale, 'all_pairs')
        if sc:
            all_sign_corrected[scale] = sc
        sc_bc = load_scale_consistency(sd, scale, 'both_correct')
        if sc_bc:
            all_sign_corrected_bc[scale] = sc_bc
        wc = load_within_cat_consistency(sd, scale, 'all_pairs')
        if wc:
            all_within_cat[scale] = wc
        wc_bc = load_within_cat_consistency(sd, scale, 'both_correct')
        if wc_bc:
            all_within_cat_bc[scale] = wc_bc
        align = load_scale_alignment(sd, scale)
        if align:
            all_alignment[scale] = align
        pred_path = os.path.join(sd, 'json', f'pred_stats_{scale}.json')
        if os.path.exists(pred_path):
            with open(pred_path) as f:
                all_pred_stats.append(json.load(f))
        cv_path = os.path.join(sd, 'json', f'category_validity_{scale}.json')
        if os.path.exists(cv_path):
            with open(cv_path) as f:
                all_cat_validity[scale] = json.load(f)
        dh = load_delta_heatmaps(sd, scale, 'all_pairs')
        if dh:
            all_delta_heatmaps[scale] = dh
        dh_bc = load_delta_heatmaps(sd, scale, 'both_correct')
        if dh_bc:
            all_delta_heatmaps_bc[scale] = dh_bc

        logger.info(f"  Loaded data for {scale}")

    # Generate cross-scale plots into condition subdirs
    for condition, sc_data, wc_data, dh_data, tag_label in [
        ('all', all_sign_corrected, all_within_cat, all_delta_heatmaps, 'all pairs'),
        ('both_correct', all_sign_corrected_bc, all_within_cat_bc, all_delta_heatmaps_bc, 'both-correct'),
    ]:
        cond_dir = os.path.join(plots_dir, condition)
        sc_dir   = os.path.join(cond_dir, 'sign_corrected')
        wc_dir   = os.path.join(cond_dir, 'within_cat_consistency')
        dt_dir   = os.path.join(cond_dir, 'delta_trajectory')
        os.makedirs(sc_dir, exist_ok=True)
        os.makedirs(wc_dir, exist_ok=True)
        os.makedirs(dt_dir, exist_ok=True)

        if len(sc_data) > 1:
            plot_cross_scale_consistency(
                sc_data, args.model_type,
                os.path.join(sc_dir, 'cross_scale_sign_corrected.png'),
                title_prefix=f'Sign-Corrected ({tag_label})')

        if len(wc_data) > 1:
            plot_cross_scale_within_cat_consistency(
                wc_data, args.model_type,
                os.path.join(wc_dir, 'cross_scale_within_cat.png'))

        if dh_data:
            plot_delta_trajectory(dh_data, args.model_type,
                                  os.path.join(dt_dir, 'delta_trajectory.png'))

    # Cross-scale alignment + pred stats + summary (shared across conditions)
    all_cond_dir   = os.path.join(plots_dir, 'all')
    ca_dir         = os.path.join(all_cond_dir, 'cross_alignment')
    pred_stats_dir = os.path.join(all_cond_dir, 'pred_stats')
    summary_dir    = os.path.join(all_cond_dir, 'summary')
    os.makedirs(ca_dir, exist_ok=True)
    os.makedirs(pred_stats_dir, exist_ok=True)
    os.makedirs(summary_dir, exist_ok=True)

    if len(all_alignment) > 1:
        plot_cross_scale_alignment(
            all_alignment, args.model_type,
            os.path.join(ca_dir, 'cross_scale_alignment.png'))

    # Prediction stats plots
    if all_pred_stats:
        plot_pred_stats_bars(all_pred_stats, args.model_type,
                             os.path.join(pred_stats_dir, 'pred_stats_bars.png'))
        plot_pred_stats_trajectory(all_pred_stats, args.model_type,
                                   os.path.join(pred_stats_dir, 'pred_stats_trajectory.png'))

    # Summary barplot
    if all_sign_corrected:
        plot_summary_barplot(
            all_sign_corrected, all_alignment, args.model_type,
            os.path.join(summary_dir, 'summary_barplot.png'))

    # Summary CSV
    summary_rows = []
    for scale in available_scales:
        ps = next((p for p in all_pred_stats if p.get('scale') == scale), None)
        if ps is None:
            continue
        row = dict(ps)
        if scale in all_alignment:
            max_layer = max(all_alignment[scale].keys())
            row['alignment_deepest'] = all_alignment[scale][max_layer]['per_sample_mean']
            row['alignment_perm'] = all_alignment[scale][max_layer]['permutation_mean']
        summary_rows.append(row)

    if summary_rows:
        csv_dir = os.path.join(merge_out, 'csv')
        os.makedirs(csv_dir, exist_ok=True)
        pd.DataFrame(summary_rows).to_csv(os.path.join(csv_dir, 'summary.csv'), index=False)

    # Accuracy charts (cross-scale)
    if all_pred_stats:
        acc_dir = os.path.join(plots_dir, 'accuracy')
        logger.info("\n--- Accuracy Charts ---")
        run_accuracy_charts(all_pred_stats, all_cat_validity, args.model_type, acc_dir)

    # Unify y-axis across scales for per-scale trajectory plots
    logger.info("\n--- Unifying Y-axis ---")
    scale_dir_map = {s: _scale_dir(s) for s in available_scales}
    run_unify_ylim(scale_dir_map, plots_dir, args.model_type)

    # All-layer heatmaps + PCA are now done per-scale in process_scale(); skip here.

    logger.info(f"\n=== Merge Complete ===\nResults in: {merge_out}")


def run_merge_extended(args):
    """Generate cross-scale plots for new / merge-only model_types.

    - Runnable types (molmo_big, qwen_big, qwen_super, big_trio):
        loads all data from results/{model_type}/ and saves plots there.
    - Merge-only types (molmo_all, qwen_all):
        loads per-scale data from the respective source directories,
        saves all cross-scale plots to results/{model_type}/.
    """
    is_merge_only = args.model_type in MERGE_ONLY_CONFIGS

    # ── Determine scale order and data source strategy ────────────────────────
    if is_merge_only:
        mc            = MERGE_ONLY_CONFIGS[args.model_type]
        scale_order   = mc['scale_order']
        scale_sources = mc['scale_sources']

        logger.info(f"\n=== MERGE-ONLY mode: {args.model_type} ===")
        logger.info("Checking required source directories...")
        sources_ok = _check_merge_only_sources(args.output_dir, args.model_type)
        if not sources_ok:
            logger.warning(
                f"\n[WARNING] One or more source directories are missing or incomplete.\n"
                f"  Cross-scale plots for '{args.model_type}' may be partial.\n"
                f"  Run the missing model types first (see warnings above), then retry merge."
            )
    else:
        scale_order   = SCALE_ORDERS_NEW.get(
            args.model_type, list(MODEL_CONFIGS_NEW[args.model_type]))
        scale_sources = None  # all data lives in results/{model_type}/

    available_scales = [s for s in scale_order if s in args.scales]
    logger.info(f"Merging scales (in order): {available_scales}")

    # ── Determine output dir (compare/{group_name}/) ──────────────────────────
    group_name = args.group_name or args.model_type
    if args.merge_output_dir:
        merge_out = args.merge_output_dir
    else:
        qt_root = os.path.dirname(args.output_dir.rstrip('/'))
        merge_out = os.path.join(qt_root, 'compare', group_name)
    plots_dir = os.path.join(merge_out, 'plots')
    os.makedirs(plots_dir, exist_ok=True)

    # ── Per-scale data directory resolver ─────────────────────────────────────
    def _scale_dir(scale):
        if is_merge_only:
            src_model_type = scale_sources[scale]
        else:
            src_model_type = args.model_type
        return os.path.join(args.output_dir, get_model_key(src_model_type, scale))

    # ── Load per-scale data ───────────────────────────────────────────────────
    all_sign_corrected    = {}
    all_sign_corrected_bc = {}
    all_within_cat        = {}
    all_within_cat_bc     = {}
    all_alignment         = {}
    all_pred_stats        = []
    all_cat_validity      = {}
    all_delta_heatmaps    = {}
    all_delta_heatmaps_bc = {}

    for scale in available_scales:
        sd = _scale_dir(scale)
        sc    = load_scale_consistency(sd, scale, 'all_pairs')
        sc_bc = load_scale_consistency(sd, scale, 'both_correct')
        wc    = load_within_cat_consistency(sd, scale, 'all_pairs')
        wc_bc = load_within_cat_consistency(sd, scale, 'both_correct')
        align = load_scale_alignment(sd, scale)

        pred_stat = None
        pred_path = os.path.join(sd, 'json', f'pred_stats_{scale}.json')
        if os.path.exists(pred_path):
            with open(pred_path) as f:
                pred_stat = json.load(f)

        cat_validity = None
        cv_path = os.path.join(sd, 'json', f'category_validity_{scale}.json')
        if os.path.exists(cv_path):
            with open(cv_path) as f:
                cat_validity = json.load(f)

        dh    = load_delta_heatmaps(sd, scale, 'all_pairs')
        dh_bc = load_delta_heatmaps(sd, scale, 'both_correct')

        if sc:
            all_sign_corrected[scale] = sc
        if sc_bc:
            all_sign_corrected_bc[scale] = sc_bc
        if wc:
            all_within_cat[scale] = wc
        if wc_bc:
            all_within_cat_bc[scale] = wc_bc
        if align:
            all_alignment[scale] = align
        if pred_stat is not None:
            all_pred_stats.append(pred_stat)
        if cat_validity is not None:
            all_cat_validity[scale] = cat_validity
        if dh:
            all_delta_heatmaps[scale] = dh
        if dh_bc:
            all_delta_heatmaps_bc[scale] = dh_bc

        logger.info(f"  Loaded data for '{scale}'"
                    + (f" (from '{scale_sources[scale]}')" if is_merge_only else ""))

    # ── Auto-detect Phase B data ──────────────────────────────────────────────
    has_phase_b = all(_has_phase_b_data(_scale_dir(s), s) for s in available_scales)
    if has_phase_b:
        logger.info("  [Phase B] Cross-alignment data found for all scales → will include cross-alignment plots")
    else:
        missing_b = [s for s in available_scales if not _has_phase_b_data(_scale_dir(s), s)]
        logger.info(f"  [Phase B] Cross-alignment data missing for: {missing_b} → skipping cross-alignment plots")

    # ── Summary CSV (Phase 1 data — always saved) ─────────────────────────────
    summary_rows = []
    for scale in available_scales:
        ps = next((p for p in all_pred_stats if p.get('scale') == scale), None)
        if ps is None:
            continue
        row = dict(ps)
        if scale in all_alignment:
            max_layer = max(all_alignment[scale].keys())
            row['alignment_deepest'] = all_alignment[scale][max_layer]['per_sample_mean']
            row['alignment_perm']    = all_alignment[scale][max_layer]['permutation_mean']
        summary_rows.append(row)
    if summary_rows:
        csv_dir = os.path.join(merge_out, 'csv')
        os.makedirs(csv_dir, exist_ok=True)
        pd.DataFrame(summary_rows).to_csv(os.path.join(csv_dir, 'summary.csv'), index=False)

    # ── Cross-scale plots (Phase 2 — skipped when --phase1-only) ─────────────
    if args.phase1_only:
        logger.info("\n--- Cross-scale plots [SKIPPED: --phase1-only] ---")
    else:
        for condition, sc_data, wc_data, dh_data, tag_label in [
            ('all',          all_sign_corrected,    all_within_cat,    all_delta_heatmaps,    'all pairs'),
            ('both_correct', all_sign_corrected_bc, all_within_cat_bc, all_delta_heatmaps_bc, 'both-correct'),
        ]:
            cond_dir = os.path.join(plots_dir, condition)
            sc_dir   = os.path.join(cond_dir, 'sign_corrected')
            wc_dir   = os.path.join(cond_dir, 'within_cat_consistency')
            dt_dir   = os.path.join(cond_dir, 'delta_trajectory')
            os.makedirs(sc_dir, exist_ok=True)
            os.makedirs(wc_dir, exist_ok=True)
            os.makedirs(dt_dir, exist_ok=True)

            if len(sc_data) > 1:
                plot_cross_scale_consistency(
                    sc_data, args.model_type,
                    os.path.join(sc_dir, 'cross_scale_sign_corrected.png'),
                    title_prefix=f'Sign-Corrected ({tag_label})')

            if len(wc_data) > 1:
                plot_cross_scale_within_cat_consistency(
                    wc_data, args.model_type,
                    os.path.join(wc_dir, 'cross_scale_within_cat.png'))

            if dh_data:
                plot_delta_trajectory(
                    dh_data, args.model_type,
                    os.path.join(dt_dir, 'delta_trajectory.png'))

        # ── Alignment and prediction stats ────────────────────────────────────
        all_cond_dir   = os.path.join(plots_dir, 'all')
        pred_stats_dir = os.path.join(all_cond_dir, 'pred_stats')
        summary_dir    = os.path.join(all_cond_dir, 'summary')
        os.makedirs(pred_stats_dir, exist_ok=True)
        os.makedirs(summary_dir, exist_ok=True)

        if has_phase_b and len(all_alignment) > 1:
            ca_dir = os.path.join(all_cond_dir, 'cross_alignment')
            os.makedirs(ca_dir, exist_ok=True)
            plot_cross_scale_alignment(
                all_alignment, args.model_type,
                os.path.join(ca_dir, 'cross_scale_alignment.png'))

        if all_pred_stats:
            plot_pred_stats_bars(
                all_pred_stats, args.model_type,
                os.path.join(pred_stats_dir, 'pred_stats_bars.png'))
            plot_pred_stats_trajectory(
                all_pred_stats, args.model_type,
                os.path.join(pred_stats_dir, 'pred_stats_trajectory.png'))

        if all_sign_corrected:
            plot_summary_barplot(
                all_sign_corrected, all_alignment, args.model_type,
                os.path.join(summary_dir, 'summary_barplot.png'))

        # ── Accuracy charts ───────────────────────────────────────────────────
        if all_pred_stats:
            acc_dir = os.path.join(plots_dir, 'accuracy')
            logger.info("\n--- Accuracy Charts ---")
            run_accuracy_charts(all_pred_stats, all_cat_validity, args.model_type, acc_dir)

        # ── Unify y-axis ──────────────────────────────────────────────────────
        logger.info("\n--- Unifying Y-axis ---")
        scale_dir_map = {s: _scale_dir(s) for s in available_scales}
        run_unify_ylim(scale_dir_map, plots_dir, args.model_type)

    # All-layer heatmaps + PCA are now done per-scale in process_scale(); skip here.

    logger.info(f"\n=== Merge Complete ===\nResults saved to: {merge_out}")


def main():
    # Default scales per legacy model_type (new types use their own defaults)
    _LEGACY_DEFAULT_SCALES = {
        'molmo':            ['vanilla', '80k', '400k', '800k', '2m'],
        'nvila':            ['vanilla', '80k', '400k', '800k', '2m'],
        'qwen':             ['vanilla', '80k', '400k', '800k', '2m'],
        'nvila_synthetic':  ['80k-5pct', '80k-10pct', '80k-20pct', '80k-30pct', '400k-5pct'],
        'nvila_st':         ['80k-st', '400k-st', '800k-st'],
    }

    parser = argparse.ArgumentParser(
        description='Swap Analysis — Spatial Representation Probing',
        formatter_class=argparse.RawDescriptionHelpFormatter,
    )
    parser.add_argument('--tsv_2d', type=str, default=CV_BENCH_2D_TSV,
                        help='Path to CV-Bench-2D-Relation-binary.tsv')
    parser.add_argument('--tsv_3d', type=str, default=CV_BENCH_3D_TSV,
                        help='Path to CV-Bench-3D.tsv')
    parser.add_argument('--model_type', type=str, required=True,
                        choices=ALL_MODEL_TYPES,
                        help=(
                            'Legacy:      molmo | nvila | qwen\n'
                            'Synthetic:   nvila_synthetic\n'
                            'New large:   molmo_big | qwen_big | qwen_super | big_trio\n'
                            'Merge-only (--merge required): molmo_all | qwen_all'
                        ))
    parser.add_argument('--scales', type=str, nargs='+', default=None,
                        help='Scales to process (default: all for the given model_type).')
    parser.add_argument('--output_dir', type=str, default=None,
                        help='Root for saved_data/. Defaults to '
                             '{script_dir}/{question_type}/saved_data.')
    parser.add_argument('--device', type=str, default='cuda')
    parser.add_argument('--seed', type=int, default=42)
    parser.add_argument('--merge', action='store_true',
                        help='Merge mode: generate cross-scale plots from saved per-scale data.')
    parser.add_argument('--merge-output-dir', type=str, default=None, dest='merge_output_dir',
                        help='(Deprecated) Override output dir for cross-scale plots. '
                             'Use --group-name instead.')
    parser.add_argument('--group-name', type=str, default=None, dest='group_name',
                        help='Folder name under compare/ for merged output. '
                             'Defaults to model_type.')
    parser.add_argument('--no-auto-roborefer', action='store_true', dest='no_auto_roborefer',
                        help='Disable automatic inclusion of roborefer scale for nvila.')
    parser.add_argument('--skip-cross-group', action='store_true')
    parser.add_argument('--max-samples-per-category', type=int, default=200,
                        dest='max_samples_per_category')
    parser.add_argument('--question-type', type=str, default='short_answer',
                        choices=['short_answer', 'mcq'], dest='question_type',
                        help='short_answer (default): "Answer with only one word." format; '
                             'mcq: MCQ A/B format with letter answers.')
    parser.add_argument('--phase1-only', action='store_true', dest='phase1_only',
                        help='Skip all plot generation (per-scale and cross-scale). '
                             'Data (npz/csv/json) is still saved.')
    parser.add_argument('--skip-phase-b', action='store_true', dest='skip_phase_b',
                        help='Skip Phase B (cross-group feature extraction). '
                             'Phase A inference + analysis + plots still run normally. '
                             'Merge auto-detects whether Phase B data is available.')

    args = parser.parse_args()

    # ── Compute output_dir and log_dir from question_type ────────────────────
    _HERE_UPDATED = os.path.dirname(os.path.abspath(__file__))
    if args.output_dir is None:
        args.output_dir = os.path.join(_HERE_UPDATED, args.question_type, 'saved_data')
    log_dir = os.path.join(_HERE_UPDATED, args.question_type, 'logs')

    # ── Validate: merge-only types require --merge ────────────────────────────
    if args.model_type in MERGE_ONLY_CONFIGS and not args.merge:
        parser.error(
            f"'{args.model_type}' is a merge-only type. Add --merge to run it.\n"
            f"  Example: python swap_analysis.py --model_type {args.model_type} --merge"
        )

    # ── Default scales ────────────────────────────────────────────────────────
    if args.scales is None:
        if args.model_type in MERGE_ONLY_CONFIGS:
            args.scales = MERGE_ONLY_CONFIGS[args.model_type]['scale_order']
        elif args.model_type in MODEL_CONFIGS_NEW:
            args.scales = list(MODEL_CONFIGS_NEW[args.model_type].keys())
        else:
            args.scales = _LEGACY_DEFAULT_SCALES.get(
                args.model_type, ['vanilla', '80k', '400k', '800k', '2m'])

    # Legacy nvila: auto-include roborefer
    if args.model_type == 'nvila' and 'roborefer' not in args.scales and not args.no_auto_roborefer:
        args.scales.append('roborefer')

    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    random.seed(args.seed)

    # ── Merge mode ───────────────────────────────────────────────────────────
    if args.merge:
        group_name = args.group_name or args.model_type
        log_path = _setup_file_logging(group_name, log_dir)
        logger.info(f"Logging to: {log_path}")
        logger.info("\n=== MERGE MODE ===")
        if args.model_type in MODEL_CONFIGS_NEW or args.model_type in MERGE_ONLY_CONFIGS:
            run_merge_extended(args)
        else:
            run_merge(args)
        return

    # ── Inference mode ────────────────────────────────────────────────────────
    logger.info("\n=== Loading & Creating CV-Bench Swap Pairs ===")
    swap_pairs = load_cv_bench_pairs(
        args.tsv_2d, args.tsv_3d,
        seed=args.seed,
        question_type=args.question_type,
    )

    quads = []
    if not args.skip_cross_group and not getattr(args, 'skip_phase_b', False):
        try:
            depth_pairs = [p for p in swap_pairs if p['group'] == 'distance']
            quads = create_cv_bench_cross_group_quads(depth_pairs)
        except Exception as e:
            logger.warning(f"Cross-group setup failed: {e}. Skipping.")
            quads = []

    # ── Resolve config for the chosen model_type ─────────────────────────────
    if args.model_type in MODEL_CONFIGS_NEW:
        model_configs = MODEL_CONFIGS_NEW[args.model_type]
    else:
        model_configs = MODEL_CONFIGS[args.model_type]

    for scale in args.scales:
        if scale not in model_configs:
            logger.warning(f"Scale '{scale}' not in config for '{args.model_type}', skipping.")
            continue

        # Per-scale log file
        vlm_key = get_model_key(args.model_type, scale)
        log_path = _setup_file_logging(vlm_key, log_dir)
        logger.info(f"Logging to: {log_path}")

        # Validate model path exists (skip HF IDs that start with org/ prefix)
        if args.model_type in MODEL_CONFIGS_NEW:
            _, raw_path = model_configs[scale]
        else:
            raw_path = model_configs[scale]
        if not os.path.isabs(raw_path) and not raw_path.startswith(('Qwen/', 'allenai/')):
            if not os.path.exists(raw_path):
                logger.warning(f"Model path not found: {raw_path}  (scale='{scale}'), skipping.")
                continue

        try:
            process_scale(args, scale, swap_pairs, quads)
        except Exception as e:
            logger.error(f"Failed {args.model_type} - {scale}: {e}")
            import traceback
            traceback.print_exc()
            continue

    logger.info(f"\n{'='*60}")
    logger.info("=== All scales complete ===")
    logger.info(f"Results: {args.output_dir}")
    logger.info(f"{'='*60}")


if __name__ == '__main__':
    main()