src/checkpoint.rs

// SPF Smart Gateway - Transformer Checkpoint System
// Copyright 2026 Joseph Stone - All Rights Reserved
//
// Save/load transformer weights to/from binary format.
// Versioned checkpoints for rollback safety.
// Delta format for mesh weight sharing (send only changed weights).
//
// Storage: LMDB via agent_state.rs or flat binary files in DEPLOY.
// Binary format: [magic:4][version:4][num_tensors:4][tensor_headers][tensor_data]
//
// Depends on: tensor.rs, transformer.rs

use crate::tensor::Tensor;
use std::io::{self, Read, Write, Cursor};

// ============================================================================
// CHECKPOINT FORMAT
// ============================================================================

/// Magic bytes identifying SPF checkpoint files
const CHECKPOINT_MAGIC: &[u8; 4] = b"SPFC";

/// Current checkpoint format version
const CHECKPOINT_VERSION: u32 = 1;

/// Metadata for a saved checkpoint
#[derive(Debug, Clone)]
pub struct CheckpointMeta {
    /// Version of the checkpoint format
    pub format_version: u32,
    /// Number of tensor parameters
    pub num_tensors: u32,
    /// Total bytes of weight data
    pub total_bytes: u64,
    /// Checkpoint creation timestamp (RFC3339)
    pub timestamp: String,
    /// Model config identifier (e.g., "spf_writer_v1")
    pub model_id: String,
    /// Training step at which checkpoint was saved
    pub step: u64,
}

// ============================================================================
// SERIALIZE — Weights to bytes
// ============================================================================

/// Serialize a list of tensors to binary checkpoint format.
///
/// Format:
/// ```text
/// [SPFC]              4 bytes  magic
/// [version]           4 bytes  u32 BE
/// [num_tensors]       4 bytes  u32 BE
/// [model_id_len]      2 bytes  u16 BE
/// [model_id]          N bytes  UTF-8
/// [step]              8 bytes  u64 BE
/// For each tensor:
///   [ndim]            4 bytes  u32 BE
///   [shape_dims]      ndim × 4 bytes  u32 BE each
///   [data]            numel × 4 bytes  f32 LE (native float layout)
/// ```
pub fn serialize_weights(
    weights: &[&Tensor],
    model_id: &str,
    step: u64,
) -> Result<Vec<u8>, io::Error> {
    let mut buf: Vec<u8> = Vec::new();

    // Header
    buf.write_all(CHECKPOINT_MAGIC)?;
    buf.write_all(&CHECKPOINT_VERSION.to_be_bytes())?;
    buf.write_all(&(weights.len() as u32).to_be_bytes())?;

    // Model ID
    let id_bytes = model_id.as_bytes();
    if id_bytes.len() > u16::MAX as usize {
        return Err(io::Error::new(io::ErrorKind::InvalidInput, "Model ID too long"));
    }
    buf.write_all(&(id_bytes.len() as u16).to_be_bytes())?;
    buf.write_all(id_bytes)?;

    // Step
    buf.write_all(&step.to_be_bytes())?;

    // Tensors
    for tensor in weights {
        // Number of dimensions
        buf.write_all(&(tensor.ndim() as u32).to_be_bytes())?;

        // Shape
        for &dim in &tensor.shape {
            buf.write_all(&(dim as u32).to_be_bytes())?;
        }

        // Data as raw f32 bytes (little-endian, native on ARM)
        for &val in &tensor.data {
            buf.write_all(&val.to_le_bytes())?;
        }
    }

    Ok(buf)
}

/// Deserialize weights from binary checkpoint format.
/// Returns (tensors, metadata).
pub fn deserialize_weights(data: &[u8]) -> Result<(Vec<Tensor>, CheckpointMeta), io::Error> {
    let mut cursor = Cursor::new(data);

    // Magic
    let mut magic = [0u8; 4];
    cursor.read_exact(&mut magic)?;
    if &magic != CHECKPOINT_MAGIC {
        return Err(io::Error::new(
            io::ErrorKind::InvalidData,
            format!("Invalid checkpoint magic: {:?}", magic),
        ));
    }

    // Version
    let mut ver_buf = [0u8; 4];
    cursor.read_exact(&mut ver_buf)?;
    let version = u32::from_be_bytes(ver_buf);
    if version != CHECKPOINT_VERSION {
        return Err(io::Error::new(
            io::ErrorKind::InvalidData,
            format!("Unsupported checkpoint version: {} (expected {})", version, CHECKPOINT_VERSION),
        ));
    }

    // Num tensors
    let mut nt_buf = [0u8; 4];
    cursor.read_exact(&mut nt_buf)?;
    let num_tensors = u32::from_be_bytes(nt_buf);

    // Model ID
    let mut id_len_buf = [0u8; 2];
    cursor.read_exact(&mut id_len_buf)?;
    let id_len = u16::from_be_bytes(id_len_buf) as usize;
    let mut id_buf = vec![0u8; id_len];
    cursor.read_exact(&mut id_buf)?;
    let model_id = String::from_utf8(id_buf)
        .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;

    // Step
    let mut step_buf = [0u8; 8];
    cursor.read_exact(&mut step_buf)?;
    let step = u64::from_be_bytes(step_buf);

    // Read tensors
    let mut tensors = Vec::with_capacity(num_tensors as usize);
    let mut total_bytes: u64 = 0;

    for _ in 0..num_tensors {
        // ndim
        let mut ndim_buf = [0u8; 4];
        cursor.read_exact(&mut ndim_buf)?;
        let ndim = u32::from_be_bytes(ndim_buf) as usize;

        // Shape
        let mut shape = Vec::with_capacity(ndim);
        for _ in 0..ndim {
            let mut dim_buf = [0u8; 4];
            cursor.read_exact(&mut dim_buf)?;
            shape.push(u32::from_be_bytes(dim_buf) as usize);
        }

        // Data
        let numel: usize = shape.iter().product();
        let mut data = Vec::with_capacity(numel);
        for _ in 0..numel {
            let mut f_buf = [0u8; 4];
            cursor.read_exact(&mut f_buf)?;
            data.push(f32::from_le_bytes(f_buf));
        }

        total_bytes += (numel * 4) as u64;
        tensors.push(Tensor { data, shape });
    }

    let meta = CheckpointMeta {
        format_version: version,
        num_tensors,
        total_bytes,
        timestamp: String::new(), // Not stored in binary — caller fills from filesystem
        model_id,
        step,
    };

    Ok((tensors, meta))
}

// ============================================================================
// APPLY WEIGHTS — Load deserialized tensors into a model
// ============================================================================

/// Apply deserialized weights to a model's weight tensors.
/// Verifies shape compatibility before applying.
pub fn apply_weights(
    model_weights: &mut [&mut Tensor],
    checkpoint_weights: &[Tensor],
) -> Result<(), String> {
    if model_weights.len() != checkpoint_weights.len() {
        return Err(format!(
            "Weight count mismatch: model has {}, checkpoint has {}",
            model_weights.len(), checkpoint_weights.len()
        ));
    }

    for (i, (model_w, ckpt_w)) in model_weights.iter_mut().zip(checkpoint_weights.iter()).enumerate() {
        if model_w.shape != ckpt_w.shape {
            return Err(format!(
                "Shape mismatch at weight {}: model {:?}, checkpoint {:?}",
                i, model_w.shape, ckpt_w.shape
            ));
        }
        model_w.data.copy_from_slice(&ckpt_w.data);
    }

    Ok(())
}

// ============================================================================
// DELTA CHECKPOINTS — For mesh weight sharing
// ============================================================================

/// Compute weight delta: new_weights - old_weights
/// Only non-zero deltas are included (sparse representation)
/// Returns: (tensor_index, delta_tensor) pairs
pub fn compute_delta(
    old_weights: &[&Tensor],
    new_weights: &[&Tensor],
    threshold: f32,
) -> Vec<(usize, Tensor)> {
    let mut deltas = Vec::new();

    for (i, (old, new)) in old_weights.iter().zip(new_weights.iter()).enumerate() {
        if old.shape != new.shape {
            continue; // Shape mismatch — skip (shouldn't happen)
        }

        let diff: Vec<f32> = old.data.iter()
            .zip(&new.data)
            .map(|(&a, &b)| b - a)
            .collect();

        // Check if any element exceeds threshold
        let has_change = diff.iter().any(|&d| d.abs() > threshold);
        if has_change {
            deltas.push((i, Tensor {
                data: diff,
                shape: old.shape.clone(),
            }));
        }
    }

    deltas
}

/// Apply weight deltas to model weights
/// delta: (tensor_index, delta_values) pairs from compute_delta
pub fn apply_delta(
    weights: &mut [&mut Tensor],
    deltas: &[(usize, Tensor)],
) -> Result<(), String> {
    for (idx, delta) in deltas {
        if *idx >= weights.len() {
            return Err(format!("Delta index {} exceeds weight count {}", idx, weights.len()));
        }
        if weights[*idx].shape != delta.shape {
            return Err(format!(
                "Delta shape mismatch at {}: weight {:?}, delta {:?}",
                idx, weights[*idx].shape, delta.shape
            ));
        }
        for (w, &d) in weights[*idx].data.iter_mut().zip(&delta.data) {
            *w += d;
        }
    }
    Ok(())
}

// ============================================================================
// MESH STREAM HANDLER — WeightSync
// ============================================================================

/// Handle an incoming WeightSync mesh frame.
/// Receives transformer weight deltas from peer nodes for federated learning.
/// Validates checkpoint format (SPFC magic bytes), returns acknowledgment.
/// Zero silent drops.
///
/// Called from: mesh.rs stream_router() for StreamType::WeightSync (0x07)
pub fn handle_weight_sync(
    frame: &crate::framing::Frame,
    peer_key: &str,
    transformer: &Option<std::sync::Arc<std::sync::RwLock<crate::transformer_tools::TransformerState>>>,
) -> Option<crate::framing::Frame> {
    let payload_len = frame.payload.len();
    let valid_format = payload_len >= 4 && &frame.payload[..4] == CHECKPOINT_MAGIC;

    eprintln!("[SPF-WEIGHT-SYNC] Received from {}: {} bytes, valid_format={}",
        &peer_key[..8.min(peer_key.len())], payload_len, valid_format);

    // Apply weights if valid format and transformer loaded
    let mut applied = false;
    let mut apply_error: Option<String> = None;
    if valid_format {
        if let Some(ref t) = transformer {
            match deserialize_weights(&frame.payload) {
                Ok((checkpoint_weights, meta)) => {
                    match t.write() {
                        Ok(mut state) => {
                            let mut model_weights = state.model.weights_mut();
                            match apply_weights(&mut model_weights, &checkpoint_weights) {
                                Ok(()) => {
                                    applied = true;
                                    eprintln!("[SPF-WEIGHT-SYNC] Applied from {}: model={}, step={}",
                                        &peer_key[..8.min(peer_key.len())],
                                        meta.model_id, meta.step);
                                }
                                Err(e) => {
                                    eprintln!("[SPF-WEIGHT-SYNC] Apply failed: {}", e);
                                    apply_error = Some(e);
                                }
                            }
                        }
                        Err(e) => { apply_error = Some(format!("Lock: {}", e)); }
                    }
                }
                Err(e) => { apply_error = Some(format!("Deserialize: {}", e)); }
            }
        }
    }

    let ack = serde_json::json!({
        "type": "weight_sync_ack",
        "bytes_received": payload_len,
        "valid_format": valid_format,
        "applied": applied,
        "error": apply_error,
        "from": peer_key,
        "status": if applied { "applied" } else if valid_format { "accepted" } else { "rejected" }
    });
    Some(crate::framing::Frame::new(
        crate::framing::StreamType::WeightSync,
        ack.to_string().into_bytes(),
    ))
}

// ============================================================================
// TESTS
// ============================================================================

#[cfg(test)]
mod tests {
    use super::*;

    fn make_test_weights() -> Vec<Tensor> {
        vec![
            Tensor::from_data(vec![1.0, 2.0, 3.0, 4.0], vec![2, 2]).unwrap(),
            Tensor::from_data(vec![0.5, -0.5, 1.5], vec![3]).unwrap(),
            Tensor::randn(&[4, 8], 42),
        ]
    }

    #[test]
    fn test_serialize_deserialize_roundtrip() {
        let weights = make_test_weights();
        let refs: Vec<&Tensor> = weights.iter().collect();

        let bytes = serialize_weights(&refs, "test_model", 100).unwrap();
        let (loaded, meta) = deserialize_weights(&bytes).unwrap();

        assert_eq!(meta.format_version, CHECKPOINT_VERSION);
        assert_eq!(meta.num_tensors, 3);
        assert_eq!(meta.model_id, "test_model");
        assert_eq!(meta.step, 100);

        assert_eq!(loaded.len(), weights.len());
        for (orig, load) in weights.iter().zip(&loaded) {
            assert_eq!(orig.shape, load.shape);
            for (a, b) in orig.data.iter().zip(&load.data) {
                assert!((a - b).abs() < 1e-7, "Data mismatch: {} vs {}", a, b);
            }
        }
    }

    #[test]
    fn test_invalid_magic() {
        let data = b"XXXX\x00\x00\x00\x01\x00\x00\x00\x00";
        let result = deserialize_weights(data);
        assert!(result.is_err());
    }

    #[test]
    fn test_apply_weights() {
        let weights = make_test_weights();
        let refs: Vec<&Tensor> = weights.iter().collect();
        let bytes = serialize_weights(&refs, "test", 0).unwrap();
        let (loaded, _) = deserialize_weights(&bytes).unwrap();

        let mut target = vec![
            Tensor::zeros(&[2, 2]),
            Tensor::zeros(&[3]),
            Tensor::zeros(&[4, 8]),
        ];
        let mut target_refs: Vec<&mut Tensor> = target.iter_mut().collect();
        apply_weights(&mut target_refs, &loaded).unwrap();

        assert_eq!(target[0].data, weights[0].data);
        assert_eq!(target[1].data, weights[1].data);
    }

    #[test]
    fn test_apply_weights_shape_mismatch() {
        let loaded = vec![Tensor::zeros(&[3, 3])]; // wrong shape
        let mut target = vec![Tensor::zeros(&[2, 2])];
        let mut target_refs: Vec<&mut Tensor> = target.iter_mut().collect();
        assert!(apply_weights(&mut target_refs, &loaded).is_err());
    }

    #[test]
    fn test_compute_delta() {
        let old = vec![
            Tensor::from_data(vec![1.0, 2.0, 3.0], vec![3]).unwrap(),
            Tensor::from_data(vec![0.0, 0.0], vec![2]).unwrap(),
        ];
        let new = vec![
            Tensor::from_data(vec![1.1, 2.0, 3.2], vec![3]).unwrap(),
            Tensor::from_data(vec![0.0, 0.0], vec![2]).unwrap(), // no change
        ];
        let old_refs: Vec<&Tensor> = old.iter().collect();
        let new_refs: Vec<&Tensor> = new.iter().collect();

        let deltas = compute_delta(&old_refs, &new_refs, 0.05);
        assert_eq!(deltas.len(), 1); // Only first tensor changed
        assert_eq!(deltas[0].0, 0);  // Index 0
        assert!((deltas[0].1.data[0] - 0.1).abs() < 1e-5);
        assert!((deltas[0].1.data[2] - 0.2).abs() < 1e-5);
    }

    #[test]
    fn test_apply_delta() {
        let mut weights = vec![
            Tensor::from_data(vec![1.0, 2.0, 3.0], vec![3]).unwrap(),
            Tensor::from_data(vec![10.0, 20.0], vec![2]).unwrap(),
        ];
        let delta = vec![
            (0, Tensor::from_data(vec![0.1, 0.2, 0.3], vec![3]).unwrap()),
        ];

        let mut refs: Vec<&mut Tensor> = weights.iter_mut().collect();
        apply_delta(&mut refs, &delta).unwrap();

        assert!((weights[0].data[0] - 1.1).abs() < 1e-5);
        assert!((weights[0].data[1] - 2.2).abs() < 1e-5);
        assert!((weights[0].data[2] - 3.3).abs() < 1e-5);
        // Second tensor unchanged
        assert_eq!(weights[1].data[0], 10.0);
    }

    #[test]
    fn test_checkpoint_size() {
        let weights = make_test_weights();
        let refs: Vec<&Tensor> = weights.iter().collect();
        let bytes = serialize_weights(&refs, "test", 0).unwrap();

        // Header: 4(magic) + 4(ver) + 4(num) + 2(id_len) + 4(id) + 8(step) = 26
        // Tensor 0: 4(ndim) + 8(shape) + 16(data) = 28
        // Tensor 1: 4(ndim) + 4(shape) + 12(data) = 20
        // Tensor 2: 4(ndim) + 8(shape) + 128(data) = 140
        // Total: 26 + 28 + 20 + 140 = 214
        assert_eq!(bytes.len(), 214);
    }

    #[test]
    fn test_large_model_id() {
        let weights = vec![Tensor::zeros(&[1])];
        let refs: Vec<&Tensor> = weights.iter().collect();
        let long_id = "x".repeat(70000);
        // Should fail — model ID > u16::MAX
        assert!(serialize_weights(&refs, &long_id, 0).is_err());
    }
}