"""
================================================================================
PUMP.FUN MEMECOIN RESEARCH CORPUS — EXPANDED QUICKSTART & ANALYST SUITE
================================================================================

HOW TO DOWNLOAD THE DATASET & LOAD YOUR WORKSPACE:
--------------------------------------------------------------------------------
1. Using huggingface-cli (Recommended — downloads all files & shards in parallel):
   $ pip install huggingface_hub
   $ huggingface-cli download <your-hf-username>/<your-repo-name> --local-dir ./pumpfun_data

2. Using Git LFS:
   $ git lfs install
   $ git clone https://huggingface.co/datasets/<your-hf-username>/<your-repo-name> ./pumpfun_data

3. Programmatic zero-disk load in Python via DuckDB over HTTP:
   import duckdb
   df = duckdb.query('''
       SELECT * FROM 'https://huggingface.co/datasets/<username>/<repo>/resolve/main/tokens.parquet' 
       LIMIT 10
   ''').df()

EXPECTED DIRECTORY STRUCTURE (--data-dir ./pumpfun_data):
   pumpfun_data/
   ├── tokens.parquet              (798,430 rows | Master definitions & creator features)
   ├── snapshots.parquet           (26,934,769 rows | Pre-grad bonding curve time-series)
   ├── postgard_snapshots.parquet  (1,392,133 rows | Post-grad DEX liquidity snapshots)
   ├── postgard_outcomes.parquet   (5,669 rows | Post-grad outcome labels & performance)
   ├── wallet_stats.parquet        (1,016,374 rows | User wallet activity & volume profiles)
   ├── migrations.parquet          (5,701 rows | Raydium migration event logs)
   └── trades/                     (33,581,704 rows | Sharded execution ledger)
       ├── shard_01.parquet
       ├── ...
       └── shard_11.parquet

MANDATORY DATA-QUALITY FILTERS APPLIED IN THIS SUITE:
   1. System Program Exclusion: 
      user_wallet != 'BwWK17cbHxwWBKZkUYvzxLcNQ1YVyaFezduWbtm2de6s'
      (Excludes protocol system-level SOL transfer accounting records from trade logs).
   2. Suspect Concentration Row Exclusion: 
      WHERE NOT COALESCE(top10_pct_suspect, FALSE)
      (Excludes tokens with scraped top-10 concentration data anomalies).
   3. Trade-Derived Feature Reconstruction:
      Rebuilds time-windowed microstructure signals directly from trades/*.parquet 
      to bypass snapshot heartbeat-duplication artifacts (~90-95% carry-forward dupes).

USAGE:
   $ python quickstart.py --data-dir ./pumpfun_data
   $ python quickstart.py --data-dir /path/to/dataset --threads 8 --memory-limit 8GB
================================================================================
"""

import argparse
import os
import sys
import time
from typing import List, Tuple
import duckdb

# System Program Wallet Address (MUST be excluded from trade-level analysis)
SYSTEM_PROGRAM_WALLET = "BwWK17cbHxwWBKZkUYvzxLcNQ1YVyaFezduWbtm2de6s"


def print_banner(section_num: int, title: str) -> None:
    """Prints a standardized visual CLI section banner."""
    print("\n" + "=" * 80)
    print(f"=== SECTION {section_num}: {title.upper()} ===")
    print("=" * 80)


def configure_duckdb(threads: int = 4, memory_limit: str = "4GB") -> duckdb.DuckDBPyConnection:
    """
    Initializes an in-memory DuckDB connection optimized for high-throughput 
    Parquet scanning across multi-threaded CPU cores.
    """
    con = duckdb.connect(database=":memory:")
    con.execute(f"PRAGMA threads={threads};")
    con.execute(f"PRAGMA memory_limit='{memory_limit}';")
    con.execute("PRAGMA preserve_insertion_order=false;")
    return con


# ==============================================================================
# 1. CORPUS DISCOVERY & TABLE FOOTPRINT AUDIT
# ==============================================================================
def audit_corpus_structure(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Validates physical presence and DuckDB readability of all 7 dataset components.
    Executes schema checks and scans row/column footprints, including multi-file 
    shards in `trades/*.parquet`.
    """
    print_banner(1, "Corpus Structure & File Inspection")

    table_configs: List[Tuple[str, str]] = [
        ("tokens", f"{data_dir}/tokens.parquet"),
        ("snapshots", f"{data_dir}/snapshots.parquet"),
        ("postgard_snapshots", f"{data_dir}/postgard_snapshots.parquet"),
        ("postgard_outcomes", f"{data_dir}/postgard_outcomes.parquet"),
        ("wallet_stats", f"{data_dir}/wallet_stats.parquet"),
        ("migrations", f"{data_dir}/migrations.parquet"),
        ("trades (sharded 11x)", f"{data_dir}/trades/*.parquet"),
    ]

    total_records = 0
    print(f"{'Table Name':<25} | {'Row Count':>15} | {'Col Count':>10} | {'Status'}")
    print("-" * 65)

    for name, path_glob in table_configs:
        try:
            # Query column structure and row counts using parallel Parquet readers
            col_query = f"SELECT COUNT(COLUMN_NAME) FROM (DESCRIBE SELECT * FROM read_parquet('{path_glob}'))"
            col_count = con.execute(col_query).fetchone()[0]

            row_query = f"SELECT COUNT(*) FROM read_parquet('{path_glob}')"
            row_count = con.execute(row_query).fetchone()[0]
            
            total_records += row_count
            print(f"{name:<25} | {row_count:>15,} | {col_count:>10} | VALID")
        except Exception as err:
            print(f"{name:<25} | {'N/A':>15} | {'N/A':>10} | ERROR ({err})")

    print("-" * 65)
    print(f"{'TOTAL INTEGRATED RECORDS':<25} | {total_records:>15,} |")
    print("\n[INFO] Sharded read check on trades/*.parquet verified across all Parquet shards.")


# ==============================================================================
# 2. DATA QUALITY AUDIT: TOKENS & CREATOR DIAGNOSTICS
# ==============================================================================
def audit_tokens_quality(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Audits the master token definitions table (`tokens.parquet`):
    - Evaluates total token population and baseline graduation rate (~0.71%).
    - Quantifies suspect top-10 concentration rows that must be excluded.
    - Inspects structural missingness in initial_gini (requires >=3 holders to compute).
    """
    print_banner(2, "Tokens Table Quality & Graduation Base Rate")

    query = f"""
        SELECT 
            COUNT(*) AS total_tokens,
            COUNT(*) FILTER (WHERE COALESCE(top10_pct_suspect, FALSE)) AS n_suspect_rows,
            COUNT(*) FILTER (WHERE graduated_at IS NOT NULL) AS n_graduated,
            COUNT(*) FILTER (WHERE initial_gini IS NULL) AS n_null_gini,
            COUNT(*) FILTER (WHERE initial_gini IS NULL AND graduated_at IS NOT NULL) AS n_null_gini_graduated,
            ROUND(AVG(initial_holder_count), 2) AS avg_initial_holders
        FROM read_parquet('{data_dir}/tokens.parquet')
    """
    stats = con.execute(query).fetchdf()

    total = stats['total_tokens'][0]
    suspect = stats['n_suspect_rows'][0]
    graduated = stats['n_graduated'][0]
    null_gini = stats['n_null_gini'][0]
    null_gini_grad = stats['n_null_gini_graduated'][0]

    grad_rate = (graduated / total) * 100.0
    clean_total = total - suspect

    print(f"Total Tokens Evaluated:      {total:>10,}")
    print(f"Suspect Top10% Rows:        {suspect:>10,} (Excl. required for clean feature models)")
    print(f"Clean Tokens Population:    {clean_total:>10,}")
    print(f"Graduated Tokens:           {graduated:>10,} ({grad_rate:.2f}% base graduation rate)")
    print(f"Initial Gini NULLs:         {null_gini:>10,} ({100*null_gini/total:.1f}% overall — structural: <3 holders)")
    print(f"Initial Gini NULLs (Grad):  {null_gini_grad:>10,} (Only {null_gini_grad} graduated token missing Gini)")
    print(f"\n[DIAGNOSTIC] Initial Gini missingness is strictly structural (tokens with <3 holders at block zero).")


# ==============================================================================
# 3. DATA QUALITY AUDIT: TRADES & SYSTEM PROGRAM FILTERING
# ==============================================================================
def audit_trades_quality(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Audits the execution ledger across all sharded Parquet files (`trades/*.parquet`):
    - Detects and quantifies System Program wallet records (must be filtered out).
    - Checks for price and volume integrity anomalies (NULL/zero amounts).
    - Measures overall distinct active trader addresses.
    """
    print_banner(3, "Trades Ledger Quality & System Wallet Filtering")

    query = f"""
        SELECT 
            COUNT(*) AS total_trades,
            COUNT(*) FILTER (WHERE user_wallet = '{SYSTEM_PROGRAM_WALLET}') AS sysprog_trades,
            COUNT(*) FILTER (WHERE price_sol IS NULL OR price_sol <= 0) AS invalid_price_trades,
            COUNT(*) FILTER (WHERE sol_amount IS NULL OR sol_amount <= 0) AS invalid_sol_trades,
            COUNT(DISTINCT user_wallet) AS unique_active_traders,
            COUNT(DISTINCT mint) AS unique_traded_mints
        FROM read_parquet('{data_dir}/trades/*.parquet')
    """
    df = con.execute(query).fetchdf()

    n_total = df['total_trades'][0]
    n_sys = df['sysprog_trades'][0]
    n_bad_price = df['invalid_price_trades'][0]
    n_bad_sol = df['invalid_sol_trades'][0]
    n_traders = df['unique_active_traders'][0]

    sys_pct = 100.0 * n_sys / n_total

    print(f"Total Granular Trades:       {n_total:>12,}")
    print(f"System Program Wallet Trades:{n_sys:>12,} ({sys_pct:.2f}% of corpus — EXCLUSION MANDATORY)")
    print(f"Unique Active Wallets:       {n_traders:>12,}")
    print(f"Invalid Price Rows (<=0/NULL):{n_bad_price:>11,}")
    print(f"Invalid SOL Amount Rows:     {n_bad_sol:>12,}")
    print(f"\n[FILTERING NOTE] Filtering out '{SYSTEM_PROGRAM_WALLET}' removes protocol accounting trades.")


# ==============================================================================
# 4. METRIC A: GRADUATION RATE BY CREATOR EXPERIENCE TIER
# ==============================================================================
def metric_creator_experience(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Computes token graduation rate bucketed by developer track record (`creator_past_tokens`).
    Demonstrates that experienced creators achieve up to ~19x-25x higher graduation rates.
    """
    print_banner(4, "Metric A: Graduation Rate by Creator Experience Tier")

    query = f"""
        WITH tiered AS (
            SELECT
                mint,
                graduated_at IS NOT NULL AS is_graduated,
                CASE
                    WHEN COALESCE(creator_past_tokens, 0) = 0 THEN '01_first_time (0)'
                    WHEN creator_past_tokens <= 10 THEN '02_novice (1-10)'
                    WHEN creator_past_tokens <= 100 THEN '03_experienced (11-100)'
                    ELSE '04_serial_creator (100+)'
                END AS creator_tier
            FROM read_parquet('{data_dir}/tokens.parquet')
            WHERE NOT COALESCE(top10_pct_suspect, FALSE)
        )
        SELECT
            creator_tier AS "Creator Past Tokens Tier",
            COUNT(*) AS n_tokens,
            SUM(CASE WHEN is_graduated THEN 1 ELSE 0 END) AS n_graduated,
            ROUND(100.0 * SUM(CASE WHEN is_graduated THEN 1 ELSE 0 END) / COUNT(*), 3) AS graduation_rate_pct
        FROM tiered
        GROUP BY 1
        ORDER BY 1
    """
    res = con.execute(query).fetchdf()
    print(res.to_string(index=False))
    print("\n[KEY FINDING] Creators with previous project history display dramatically higher graduation odds.")


# ==============================================================================
# 5. METRIC B: HOLDER CONCENTRATION VS RUG RISK ANALYSIS
# ==============================================================================
def metric_concentration_vs_outcomes(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Analyzes holder Gini index (`initial_gini`) vs post-graduation rug risk (`postgard_outcomes`).
    Filters out suspect concentration records for statistical validity.
    """
    print_banner(5, "Metric B: Holder Concentration (Gini) vs Post-Graduation Rug Risk")

    query = f"""
        WITH clean_tokens AS (
            SELECT 
                mint,
                initial_gini,
                initial_top10_pct
            FROM read_parquet('{data_dir}/tokens.parquet')
            WHERE NOT COALESCE(top10_pct_suspect, FALSE)
              AND initial_gini IS NOT NULL
        ),
        outcomes AS (
            SELECT 
                t.mint,
                t.initial_gini,
                o.outcome_label,
                o.rug_detected,
                CASE WHEN o.rug_detected THEN 1 ELSE 0 END AS is_rug
            FROM clean_tokens t
            INNER JOIN read_parquet('{data_dir}/postgard_outcomes.parquet') o ON t.mint = o.mint
        ),
        bucketed AS (
            SELECT 
                NTILE(5) OVER (ORDER BY initial_gini) AS gini_quintile,
                initial_gini,
                is_rug
            FROM outcomes
        )
        SELECT 
            gini_quintile AS "Gini Quintile",
            ROUND(MIN(initial_gini), 3) AS min_gini,
            ROUND(MAX(initial_gini), 3) AS max_gini,
            COUNT(*) AS n_graduated_tokens,
            SUM(is_rug) AS n_rugs,
            ROUND(100.0 * SUM(is_rug) / COUNT(*), 2) AS rug_rate_pct
        FROM bucketed
        GROUP BY 1
        ORDER BY 1
    """
    res = con.execute(query).fetchdf()
    print(res.to_string(index=False))
    print("\n[KEY FINDING] High holder inequality (Gini > 0.75) strongly predicts post-graduation rugging.")


# ==============================================================================
# 6. MICROSTRUCTURE: CLEAN TRADE-GRID FEATURE RECONSTRUCTION
# ==============================================================================
def compute_trade_microstructure_sample(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Demonstrates building time-windowed trade flow metrics (T+3m, T+6m) directly
    from sharded `trades/*.parquet` files to bypass snapshot heartbeat duplicates.
    """
    print_banner(6, "Trade Microstructure Flow (Derived directly from trades/*.parquet)")

    query = f"""
        WITH filtered_trades AS (
            SELECT 
                mint,
                seconds_since_launch,
                is_buy,
                sol_amount
            FROM read_parquet('{data_dir}/trades/*.parquet')
            WHERE user_wallet != '{SYSTEM_PROGRAM_WALLET}'
              AND seconds_since_launch <= 360  -- First 6 minutes post-launch
        ),
        aggregated AS (
            SELECT 
                mint,
                COUNT(*) AS trades_first_6m,
                SUM(CASE WHEN seconds_since_launch <= 180 AND is_buy THEN sol_amount ELSE 0 END) AS buy_vol_3m,
                SUM(CASE WHEN seconds_since_launch <= 180 AND NOT is_buy THEN sol_amount ELSE 0 END) AS sell_vol_3m,
                SUM(CASE WHEN is_buy THEN sol_amount ELSE 0 END) AS buy_vol_6m,
                SUM(sol_amount) AS total_vol_6m
            FROM filtered_trades
            GROUP BY mint
        )
        SELECT 
            t.mint,
            t.symbol,
            a.trades_first_6m,
            ROUND(a.buy_vol_3m, 2) AS buy_vol_3m_sol,
            ROUND(a.sell_vol_3m, 2) AS sell_vol_3m_sol,
            ROUND(a.buy_vol_6m / NULLIF(a.total_vol_6m, 0), 3) AS buy_pressure_6m
        FROM aggregated a
        JOIN read_parquet('{data_dir}/tokens.parquet') t ON a.mint = t.mint
        WHERE NOT COALESCE(t.top10_pct_suspect, FALSE)
        ORDER BY a.total_vol_6m DESC
        LIMIT 5
    """
    res = con.execute(query).fetchdf()
    print(res.to_string(index=False))
    print("\n[METHODOLOGY] Constructing signals directly from trades avoids snapshot heartbeat duplication.")


# ==============================================================================
# 7. POST-GRADUATION DEX OUTCOME & LIQUIDITY ANALYSIS
# ==============================================================================
def analyze_postgrad_performance(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Analyzes DEX survival classifications (`postgard_outcomes.parquet`) and liquidity 
    decay trajectories (`postgard_snapshots.parquet`).
    """
    print_banner(7, "Post-Graduation Outcomes & DEX Liquidity Retention")

    # Outcome Label Breakdown
    print("--> Categorical Survival Breakdown across Graduated Population:")
    query_outcomes = f"""
        SELECT 
            outcome_label AS "Outcome Label",
            COUNT(*) AS n_tokens,
            ROUND(100.0 * COUNT(*) / SUM(COUNT(*)) OVER(), 2) AS pct_of_graduated,
            ROUND(AVG(price_change_24h_pct), 2) AS avg_24h_price_change_pct,
            SUM(CASE WHEN still_liquid_at_24h THEN 1 ELSE 0 END) AS liquid_at_24h_count
        FROM read_parquet('{data_dir}/postgard_outcomes.parquet')
        GROUP BY outcome_label
        ORDER BY n_tokens DESC
    """
    res_outcomes = con.execute(query_outcomes).fetchdf()
    print(res_outcomes.to_string(index=False))

    # Post-Grad Snapshots Liquidity Sample Check
    print("\n--> DEX Liquidity & Volume Time-Series Sample (postgard_snapshots.parquet):")
    query_snapshots = f"""
        SELECT 
            mint,
            seconds_since_graduation,
            ROUND(price_usd, 6) AS price_usd,
            ROUND(liquidity_usd, 2) AS liquidity_usd,
            ROUND(volume_1h, 2) AS vol_1h_usd,
            buy_pressure_1h
        FROM read_parquet('{data_dir}/postgard_snapshots.parquet')
        WHERE liquidity_usd IS NOT NULL
        ORDER BY snapshot_time DESC
        LIMIT 5
    """
    res_snapshots = con.execute(query_snapshots).fetchdf()
    print(res_snapshots.to_string(index=False))


# ==============================================================================
# 8. WALLET ANALYTICS & HIGH-FREQUENCY TRADER PROFILING
# ==============================================================================
def analyze_wallet_profiles(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Profiles trader behaviors and volume distributions across 1.02M unique addresses 
    logged in `wallet_stats.parquet`.
    """
    print_banner(8, "Wallet Stats & High-Frequency Trader Distribution")

    query = f"""
        WITH wallet_tiers AS (
            SELECT 
                wallet,
                tokens_traded,
                graduated_tokens_traded,
                total_trades,
                total_buy_volume_sol + total_sell_volume_sol AS total_volume_sol,
                CASE 
                    WHEN total_trades >= 1000 THEN '01_bot_or_infra (1000+ trades)'
                    WHEN total_trades >= 100 THEN '02_heavy_trader (100-999)'
                    WHEN total_trades >= 10 THEN '03_regular_trader (10-99)'
                    ELSE '04_casual_trader (1-9)'
                END AS trader_tier
            FROM read_parquet('{data_dir}/wallet_stats.parquet')
        )
        SELECT 
            trader_tier AS "Trader Activity Tier",
            COUNT(*) AS n_wallets,
            ROUND(100.0 * COUNT(*) / SUM(COUNT(*)) OVER(), 2) AS pct_wallets,
            ROUND(AVG(tokens_traded), 1) AS avg_tokens_traded,
            ROUND(SUM(total_volume_sol), 1) AS tier_total_volume_sol
        FROM wallet_tiers
        GROUP BY 1
        ORDER BY 1
    """
    res = con.execute(query).fetchdf()
    print(res.to_string(index=False))
    print("\n[INSIGHT] A small percentage of automated bot wallets drive a major share of execution frequency.")


# ==============================================================================
# 9. MIGRATION EVENT LOGS & RAYDIUM TIMING ANALYSIS
# ==============================================================================
def analyze_migrations_performance(con: duckdb.DuckDBPyConnection, data_dir: str) -> None:
    """
    Inspects Raydium migration timing patterns logged in `migrations.parquet`.
    Computes summary distribution statistics for bonding curve graduation speed.
    """
    print_banner(9, "Raydium Migration Event Timing & Velocity Distribution")

    query = f"""
        SELECT 
            COUNT(*) AS total_migrations,
            ROUND(AVG(seconds_to_graduation), 1) AS avg_seconds_to_grad,
            ROUND(MEDIAN(seconds_to_graduation), 1) AS median_seconds_to_grad,
            ROUND(QUANTILE_CONT(seconds_to_graduation, 0.10), 1) AS p10_seconds,
            ROUND(QUANTILE_CONT(seconds_to_graduation, 0.90), 1) AS p90_seconds,
            MIN(seconds_to_graduation) AS fastest_grad_seconds,
            MAX(seconds_to_graduation) AS slowest_grad_seconds
        FROM read_parquet('{data_dir}/migrations.parquet')
    """
    res = con.execute(query).fetchdf()
    print(res.to_string(index=False))
    print("\n[INSIGHT] Graduation speed ranges from hyper-fast sniped launches to multi-day gradual curves.")


# ==============================================================================
# 10. SUMMARY & PIPELINE DATA-QUALITY CHECKLIST
# ==============================================================================
def print_pipeline_summary() -> None:
    """Prints a checklist of required data quality rules before building models."""
    print_banner(10, "Summary & Data Quality Checklist for Machine Learning")

    checklist = [
        ("1. System Program Exclusion", "ALWAYS filter out user_wallet = 'BwWK17cbHxwWBKZkUYvzxLcNQ1YVyaFezduWbtm2de6s'"),
        ("2. Suspect Row Exclusion", "ALWAYS exclude rows where COALESCE(top10_pct_suspect, FALSE) is True"),
        ("3. Heartbeat Duplicate Prevention", "Rebuild time-series signals directly from trades/*.parquet shards"),
        ("4. Structural Missingness", "Treat initial_gini NULLs as structural (<3 holders), not random missingness"),
        ("5. Walk-Forward CV Splits", "Use time-based walk-forward splits to account for non-stationary base rates"),
    ]

    for item, rule in checklist:
        print(f"  [CHECK] {item:<32} -> {rule}")
    print("\nSee README.md for full model feature engineering guidelines and replication benchmarks.")


# ==============================================================================
# MAIN CLI DRIVER
# ==============================================================================
def main() -> None:
    parser = argparse.ArgumentParser(
        description="Expanded Quickstart & Quality Audit Script for Pump.fun Memecoin Dataset"
    )
    parser.add_argument(
        "--data-dir", 
        required=True, 
        help="Path to root directory containing parquet files and trades/ subfolder"
    )
    parser.add_argument(
        "--threads", 
        type=int, 
        default=4, 
        help="Number of CPU threads for DuckDB engine (default: 4)"
    )
    parser.add_argument(
        "--memory-limit", 
        type=str, 
        default="4GB", 
        help="DuckDB memory ceiling, e.g. '4GB', '8GB' (default: 4GB)"
    )

    args = parser.parse_args()
    data_dir = args.data_dir.rstrip("/")

    # Directory existence check
    if not os.path.exists(data_dir):
        print(f"[ERROR] Directory '{data_dir}' does not exist. Please check --data-dir path.")
        sys.exit(1)

    t_start = time.time()
    print(f"Initializing DuckDB Engine (Threads: {args.threads}, Memory Limit: {args.memory_limit})...")
    con = configure_duckdb(threads=args.threads, memory_limit=args.memory_limit)

    # Run analytical sections sequentially
    audit_corpus_structure(con, data_dir)
    audit_tokens_quality(con, data_dir)
    audit_trades_quality(con, data_dir)
    metric_creator_experience(con, data_dir)
    metric_concentration_vs_outcomes(con, data_dir)
    compute_trade_microstructure_sample(con, data_dir)
    analyze_postgrad_performance(con, data_dir)
    analyze_wallet_profiles(con, data_dir)
    analyze_migrations_performance(con, data_dir)
    print_pipeline_summary()

    t_elapsed = time.time() - t_start
    print("\n" + "=" * 80)
    print(f"SUCCESS: ALL 10 ANALYTICAL SECTIONS COMPLETED IN {t_elapsed:.2f} SECONDS.")
    print("=" * 80 + "\n")


if __name__ == "__main__":
    main()