package_code_imports/v4p3_scripts/model_v4p3.py

from __future__ import annotations

import math
import sys
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any

import torch
import torch.nn as nn
import torch.nn.functional as F


SCRIPT_DIR = Path(__file__).resolve().parent
V3P5_SCRIPT_DIR = SCRIPT_DIR.parents[1] / "v3p5_static" / "scripts"
if str(V3P5_SCRIPT_DIR) not in sys.path:
    sys.path.insert(0, str(V3P5_SCRIPT_DIR))

from model_v3p5 import (  # noqa: E402
    DrugSetEncoder,
    FieldSpecificCategoricalHead,
    PatientStaticEncoder,
    ProportionalOddsOrdinalHead,
    build_service_attention_prior,
)


PWE_BIN_EDGES_DAYS = (7.0, 14.0, 30.0, 60.0, 90.0, 365.0, math.inf)
PWE_FINITE_LOWER_DAYS = (0.0, 7.0, 14.0, 30.0, 60.0, 90.0, 365.0)


@dataclass
class SCTMv4p3Config:
    cat_vocab_size: int
    drug_vocab_size: int
    n_cat_fields: int
    n_numeric_fields: int
    n_ordinal_fields: int
    cbe_dim: int
    max_drug_atoms: int
    n_events: int
    d_model: int = 192
    n_heads: int = 6
    n_layers: int = 4
    dropout: float = 0.05
    n_missing: int = 5
    n_drug_classes: int = 4
    n_service_states: int = 8
    n_active_states: int = 5
    n_pwe_causes: int = 3
    n_pwe_bins: int = 7
    n_eras: int = 3
    n_conditional_status: int = 4
    static_vocab_size: int = 0
    n_static_fields: int = 0
    year_min: int = 2010
    year_max: int = 2026
    missing_observed_id: int = 0
    missing_era_unavailable_id: int = 1
    missing_visit_missing_id: int = 3
    missing_structural_inactive_id: int = 2
    missing_no_clinical_id: int = 4
    existing_missing_visit_missing_id: int = 2
    existing_missing_pad_id: int = 3
    unknown_cat_value_id: int = 1
    contact_injection_init: float = 0.05
    two_year_days: float = 730.0
    conditional_specs: tuple[dict[str, Any], ...] = field(default_factory=tuple)
    color_cat_field_index: int = -1
    color_value_to_class: tuple[tuple[int, int], ...] = field(default_factory=tuple)


def era_tokens(visit_year: torch.Tensor) -> torch.Tensor:
    year = visit_year.long()
    return torch.where(year <= 2013, torch.zeros_like(year), torch.where(year <= 2018, torch.ones_like(year), torch.full_like(year, 2)))


def year_features(visit_year: torch.Tensor, year_min: int, year_max: int) -> torch.Tensor:
    year = visit_year.float().clamp(min=year_min, max=year_max)
    scale = max(1.0, float(year_max - year_min))
    phase = (year - year_min) / scale * (2.0 * math.pi)
    return torch.stack([torch.sin(phase), torch.cos(phase)], dim=-1)


def previous_interval_log(time_since_start_days: torch.Tensor) -> torch.Tensor:
    prev_gap = torch.zeros_like(time_since_start_days)
    prev_gap[:, 1:] = (time_since_start_days[:, 1:] - time_since_start_days[:, :-1]).clamp(min=0.0)
    return torch.log1p(prev_gap)


class AxisSlotTabEncoderV4p3(nn.Module):
    """v4.3 intra-visit encoder with conditional-field status and 3-era tokens."""

    def __init__(self, config: SCTMv4p3Config) -> None:
        super().__init__()
        self.config = config
        d = config.d_model
        self.cat_field_emb = nn.Embedding(config.n_cat_fields, d)
        self.cat_value_emb = nn.Embedding(config.cat_vocab_size, d, padding_idx=0)
        self.missing_emb = nn.Embedding(config.n_missing, d)
        self.conditional_status_emb = nn.Embedding(config.n_conditional_status, d)
        self.numeric_field_emb = nn.Embedding(max(1, config.n_numeric_fields), d)
        self.numeric_value_proj = nn.Linear(1, d)
        self.ordinal_field_emb = nn.Embedding(max(1, config.n_ordinal_fields), d)
        self.ordinal_cbe_proj = nn.Linear(config.cbe_dim, d)
        self.drug_encoder = DrugSetEncoder(config)
        self.service_emb = nn.Embedding(config.n_service_states, d)
        self.year_proj = nn.Linear(2, d)
        self.era_emb = nn.Embedding(config.n_eras, d)
        self.axis_slot_emb = nn.Parameter(torch.randn(1, 1, 7, d) * 0.02)
        self.axis_attn = nn.MultiheadAttention(d, config.n_heads, dropout=config.dropout, batch_first=True)
        self.axis_norm = nn.LayerNorm(d)
        self.axis_ff = nn.Sequential(nn.Linear(d, d * 2), nn.GELU(), nn.Dropout(config.dropout), nn.Linear(d * 2, d))
        self.out_norm = nn.LayerNorm(d)
        self.register_buffer("cat_field_ids", torch.arange(config.n_cat_fields), persistent=False)
        self.register_buffer("numeric_field_ids", torch.arange(max(1, config.n_numeric_fields)), persistent=False)
        self.register_buffer("ordinal_field_ids", torch.arange(max(1, config.n_ordinal_fields)), persistent=False)

    def conditional_status(self, cat_value_ids: torch.Tensor, missing_ids: torch.Tensor) -> torch.Tensor:
        """Return ACTIVE/STRUCTURAL_INACTIVE/PARENT_UNKNOWN/DATA_INCONSISTENT codes."""

        status = torch.zeros_like(cat_value_ids, dtype=torch.long)
        if not self.config.conditional_specs:
            return status
        missing = missing_ids.clamp(min=0, max=self.config.n_missing - 1)
        observed = missing.eq(self.config.missing_observed_id)
        unknown = self.config.unknown_cat_value_id
        for spec in self.config.conditional_specs:
            parent_idx = int(spec["parent_idx"])
            child_indices = [int(x) for x in spec["child_indices"]]
            active_ids = torch.tensor([int(x) for x in spec["active_value_ids"]], device=cat_value_ids.device, dtype=cat_value_ids.dtype)
            if parent_idx < 0 or parent_idx >= cat_value_ids.shape[-1] or not child_indices or active_ids.numel() == 0:
                continue
            parent_val = cat_value_ids[:, :, parent_idx]
            parent_observed = observed[:, :, parent_idx] & parent_val.ne(unknown)
            parent_active = (parent_val[:, :, None] == active_ids[None, None, :]).any(dim=-1) & parent_observed
            parent_unknown = ~parent_observed
            for child_idx in child_indices:
                if child_idx < 0 or child_idx >= cat_value_ids.shape[-1]:
                    continue
                child_val = cat_value_ids[:, :, child_idx]
                child_observed = observed[:, :, child_idx] & child_val.ne(unknown)
                code = torch.zeros_like(parent_val, dtype=torch.long)
                code = torch.where(parent_unknown, torch.full_like(code, 2), code)
                code = torch.where((~parent_unknown) & (~parent_active) & (~child_observed), torch.full_like(code, 1), code)
                code = torch.where((~parent_unknown) & (~parent_active) & child_observed, torch.full_like(code, 3), code)
                status[:, :, child_idx] = code
        return status

    def forward(
        self,
        cat_value_ids: torch.Tensor,
        missing_ids: torch.Tensor,
        numeric_values: torch.Tensor,
        numeric_mask: torch.Tensor,
        ordinal_cbe: torch.Tensor,
        ordinal_mask: torch.Tensor,
        drug_name_ids: torch.Tensor,
        drug_class_ids: torch.Tensor,
        drug_mask: torch.Tensor,
        service_state: torch.Tensor,
        visit_year: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        bsz, seq_len, _ = cat_value_ids.shape
        device = cat_value_ids.device
        cat_fields = self.cat_field_emb(self.cat_field_ids.to(device))[None, None, :, :]
        missing = missing_ids.clamp(min=0, max=self.config.n_missing - 1)
        conditional = self.conditional_status(cat_value_ids, missing)
        cat = (
            cat_fields
            + self.cat_value_emb(cat_value_ids.clamp(min=0, max=self.config.cat_vocab_size - 1))
            + self.missing_emb(missing)
            + self.conditional_status_emb(conditional.clamp(0, self.config.n_conditional_status - 1))
        )
        cat_vec = cat.mean(dim=2)

        if self.config.n_numeric_fields:
            num_fields = self.numeric_field_emb(self.numeric_field_ids[: self.config.n_numeric_fields].to(device))[None, None, :, :]
            num = num_fields + self.numeric_value_proj(numeric_values.unsqueeze(-1))
            num = num * numeric_mask.unsqueeze(-1).float()
            denom = numeric_mask.sum(dim=2).clamp(min=1).unsqueeze(-1).float()
            num_vec = num.sum(dim=2) / denom
        else:
            num_vec = torch.zeros_like(cat_vec)

        if self.config.n_ordinal_fields:
            ord_fields = self.ordinal_field_emb(self.ordinal_field_ids[: self.config.n_ordinal_fields].to(device))[None, None, :, :]
            ord_vecs = ord_fields + self.ordinal_cbe_proj(ordinal_cbe)
            ord_vecs = ord_vecs * ordinal_mask.unsqueeze(-1).float()
            denom = ordinal_mask.sum(dim=2).clamp(min=1).unsqueeze(-1).float()
            ord_vec = ord_vecs.sum(dim=2) / denom
        else:
            ord_vec = torch.zeros_like(cat_vec)

        drug_vec = self.drug_encoder(drug_name_ids, drug_class_ids, drug_mask)
        service_vec = self.service_emb(service_state.clamp(min=0, max=self.config.n_service_states - 1))
        year_vec = self.year_proj(year_features(visit_year, self.config.year_min, self.config.year_max))
        era_vec = self.era_emb(era_tokens(visit_year).clamp(0, self.config.n_eras - 1))
        slots = torch.stack([cat_vec, num_vec, ord_vec, drug_vec, service_vec, year_vec, era_vec], dim=2)
        slots = slots + self.axis_slot_emb
        flat_slots = slots.reshape(bsz * seq_len, 7, -1)
        attended, _ = self.axis_attn(flat_slots, flat_slots, flat_slots, need_weights=False)
        flat_slots = self.axis_norm(flat_slots + attended)
        flat_slots = self.axis_norm(flat_slots + self.axis_ff(flat_slots))
        visit_vec = flat_slots.mean(dim=1).reshape(bsz, seq_len, -1)
        return self.out_norm(visit_vec), conditional


class EraTimeStateSelfAttention(nn.Module):
    def __init__(self, config: SCTMv4p3Config, service_prior_bias: torch.Tensor | None = None) -> None:
        super().__init__()
        if config.d_model % config.n_heads != 0:
            raise ValueError("d_model must be divisible by n_heads")
        self.config = config
        self.head_dim = config.d_model // config.n_heads
        self.qkv = nn.Linear(config.d_model, config.d_model * 3)
        self.out = nn.Linear(config.d_model, config.d_model)
        self.dropout = nn.Dropout(config.dropout)
        self.time_alpha = nn.Parameter(torch.zeros(config.n_heads))
        init = torch.zeros(config.n_heads, config.n_eras, config.n_eras, config.n_service_states, config.n_service_states)
        if service_prior_bias is not None:
            if service_prior_bias.shape != (config.n_service_states, config.n_service_states):
                raise ValueError("service_prior_bias must be [n_service_states, n_service_states]")
            prior = service_prior_bias.float().to(device=init.device)
            init = init + prior[None, None, None, :, :]
        self.state_era_bias = nn.Parameter(init)

    def forward(
        self,
        x: torch.Tensor,
        service_state: torch.Tensor,
        visit_year: torch.Tensor,
        time_since_start_days: torch.Tensor,
        valid_mask: torch.Tensor,
    ) -> torch.Tensor:
        bsz, seq_len, d_model = x.shape
        qkv = self.qkv(x).view(bsz, seq_len, 3, self.config.n_heads, self.head_dim)
        q, k, v = qkv[:, :, 0].transpose(1, 2), qkv[:, :, 1].transpose(1, 2), qkv[:, :, 2].transpose(1, 2)
        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        tdiff = (time_since_start_days[:, :, None] - time_since_start_days[:, None, :]).abs().clamp(min=0)
        time_penalty = -F.softplus(self.time_alpha)[None, :, None, None] * torch.log1p(tdiff)[:, None, :, :]
        scores = scores + time_penalty

        qs = service_state.clamp(0, self.config.n_service_states - 1)
        ks = qs
        qe = era_tokens(visit_year).clamp(0, self.config.n_eras - 1)
        ke = qe
        qid = qe * self.config.n_service_states + qs
        kid = ke * self.config.n_service_states + ks
        bias_table = self.state_era_bias.reshape(
            self.config.n_heads,
            self.config.n_eras * self.config.n_service_states,
            self.config.n_eras * self.config.n_service_states,
        )
        bias = bias_table[:, qid[:, :, None], kid[:, None, :]].permute(1, 0, 2, 3)
        scores = scores + bias

        causal = torch.triu(torch.ones(seq_len, seq_len, dtype=torch.bool, device=x.device), diagonal=1)
        scores = scores.masked_fill(causal[None, None, :, :], torch.finfo(scores.dtype).min)
        scores = scores.masked_fill(~valid_mask[:, None, None, :], torch.finfo(scores.dtype).min)
        attn = torch.softmax(scores, dim=-1)
        attn = torch.nan_to_num(attn, nan=0.0)
        y = torch.matmul(self.dropout(attn), v).transpose(1, 2).contiguous().view(bsz, seq_len, d_model)
        return self.out(y)


class EraTrajDecoderLayer(nn.Module):
    def __init__(self, config: SCTMv4p3Config, service_prior_bias: torch.Tensor | None = None) -> None:
        super().__init__()
        d = config.d_model
        self.attn = EraTimeStateSelfAttention(config, service_prior_bias=service_prior_bias)
        self.norm1 = nn.LayerNorm(d)
        self.norm2 = nn.LayerNorm(d)
        self.ff = nn.Sequential(nn.Linear(d, d * 4), nn.GELU(), nn.Dropout(config.dropout), nn.Linear(d * 4, d))
        self.dropout = nn.Dropout(config.dropout)

    def forward(
        self,
        x: torch.Tensor,
        service_state: torch.Tensor,
        visit_year: torch.Tensor,
        time_since_start_days: torch.Tensor,
        valid_mask: torch.Tensor,
    ) -> torch.Tensor:
        x = x + self.dropout(self.attn(self.norm1(x), service_state, visit_year, time_since_start_days, valid_mask))
        x = x + self.dropout(self.ff(self.norm2(x)))
        return x


class EraTrajDecoder(nn.Module):
    def __init__(self, config: SCTMv4p3Config, service_prior_bias: torch.Tensor | None = None) -> None:
        super().__init__()
        self.config = config
        self.start_visit = nn.Parameter(torch.zeros(1, 1, config.d_model))
        self.layers = nn.ModuleList([EraTrajDecoderLayer(config, service_prior_bias=service_prior_bias) for _ in range(config.n_layers)])
        self.final_norm = nn.LayerNorm(config.d_model)

    def causal_decode(
        self,
        u: torch.Tensor,
        service_state: torch.Tensor,
        visit_year: torch.Tensor,
        time_since_start_days: torch.Tensor,
        valid_mask: torch.Tensor,
    ) -> torch.Tensor:
        x = u
        for layer in self.layers:
            x = layer(x, service_state, visit_year, time_since_start_days, valid_mask)
        return self.final_norm(x)

    def history_inputs(
        self,
        u: torch.Tensor,
        service_state: torch.Tensor,
        visit_year: torch.Tensor,
        time_since_start_days: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        bsz = u.shape[0]
        shifted_u = torch.cat([self.start_visit.expand(bsz, 1, -1), u[:, :-1, :]], dim=1)
        shifted_service = torch.cat([torch.zeros_like(service_state[:, :1]), service_state[:, :-1]], dim=1)
        shifted_year = torch.cat([visit_year[:, :1], visit_year[:, :-1]], dim=1)
        shifted_time = torch.cat([torch.zeros_like(time_since_start_days[:, :1]), time_since_start_days[:, :-1]], dim=1)
        return shifted_u, shifted_service, shifted_year, shifted_time

    def forward(
        self,
        u: torch.Tensor,
        service_state: torch.Tensor,
        visit_year: torch.Tensor,
        time_since_start_days: torch.Tensor,
        valid_mask: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        h_u, h_s, h_y, h_t = self.history_inputs(u, service_state, visit_year, time_since_start_days)
        z_history = self.causal_decode(h_u, h_s, h_y, h_t, valid_mask)
        z_post = self.causal_decode(u, service_state, visit_year, time_since_start_days, valid_mask)
        return z_history, z_post


class ContactInjector(nn.Module):
    def __init__(self, config: SCTMv4p3Config) -> None:
        super().__init__()
        self.config = config
        d = config.d_model
        self.arrival_proj = nn.Linear(4, d)
        self.era_emb = nn.Embedding(config.n_eras, d)
        self.norm = nn.LayerNorm(d)
        self.scale = nn.Parameter(torch.tensor(float(config.contact_injection_init)))

    def forward(self, z_history: torch.Tensor, batch: dict[str, torch.Tensor]) -> torch.Tensor:
        prev_log = previous_interval_log(batch["time_since_start_days"])
        year_feat = year_features(batch["visit_year"], self.config.year_min, self.config.year_max)
        if "visit_indices" in batch:
            visit_idx = batch["visit_indices"].float().clamp(min=0.0) / 256.0
        else:
            visit_idx = torch.zeros_like(prev_log)
        arrival = torch.cat([prev_log.unsqueeze(-1), year_feat, visit_idx.unsqueeze(-1)], dim=-1)
        contact = self.arrival_proj(arrival) + self.era_emb(era_tokens(batch["visit_year"]).clamp(0, self.config.n_eras - 1))
        return self.norm(z_history + self.scale * contact)


class SCTMv4p3(nn.Module):
    """SCTM v4.3 L1/L2-aware architecture.

    The raw v3p5 tensor remains the compatibility contract. v4.3-specific
    targets (3-cause K, 5-state missingness, conditional structural inactivity)
    are derived in the loss from that tensor and the metadata-built config.
    """

    def __init__(
        self,
        config: SCTMv4p3Config,
        field_value_mask: torch.Tensor,
        service_prior_bias: torch.Tensor | None = None,
    ) -> None:
        super().__init__()
        if config.n_pwe_bins != len(PWE_BIN_EDGES_DAYS):
            raise ValueError("v4.3 uses the 7-bin data-verified PWE grid")
        self.config = config
        self.tab_encoder = AxisSlotTabEncoderV4p3(config)
        self.traj_decoder = EraTrajDecoder(config, service_prior_bias=service_prior_bias)
        self.contact_injector = ContactInjector(config)
        self.static_encoder = PatientStaticEncoder(config)
        self.static_injection_scale = nn.Parameter(torch.zeros(1))
        d = config.d_model
        self.next_delta_condition = nn.Linear(1, d)
        self.pwe_log_lambda_pre_head = nn.Linear(d, config.n_pwe_causes * config.n_pwe_bins)
        self.pwe_log_lambda_contact_head = nn.Linear(d, config.n_pwe_causes * config.n_pwe_bins)
        self.pwe_log_lambda_post_head = nn.Linear(d, config.n_pwe_causes * config.n_pwe_bins)
        self.active_state_head = nn.Linear(d, config.n_active_states)
        self.missingness_head = nn.Linear(d, config.n_cat_fields * config.n_missing)
        self.deployable_risk_pre_head = nn.Linear(d, config.n_events)
        self.deployable_risk_contact_head = nn.Sequential(
            nn.Linear(d * 2, d),
            nn.GELU(),
            nn.Dropout(config.dropout),
            nn.Linear(d, config.n_events),
        )
        self.deployable_risk_post_head = nn.Linear(d, config.n_events)
        self.field_head = FieldSpecificCategoricalHead(d, config.n_cat_fields, config.cat_vocab_size, field_value_mask)
        self.numeric_head = nn.Linear(d, max(1, config.n_numeric_fields))
        self.ordinal_head = ProportionalOddsOrdinalHead(d, config.n_ordinal_fields, config.cbe_dim)
        self.ontology_event_head = nn.Linear(d, config.n_events)
        self.history_event_head = nn.Linear(d, config.n_events)
        self.contact_event_head = nn.Linear(d, config.n_events)
        self.color_signal_head = nn.Linear(d, 4)
        self.rollout_prefix_proj = nn.Sequential(
            nn.Linear(config.n_pwe_causes + config.n_active_states + config.n_events + 2, d),
            nn.GELU(),
            nn.LayerNorm(d),
        )
        # These v4-only visit-encoder terms have no v3 checkpoint analogue. Keep
        # them neutral during Stage 1 backbone-freeze warmup; Stage 2 can learn
        # non-zero effects once the backbone is unfrozen.
        nn.init.zeros_(self.tab_encoder.conditional_status_emb.weight)
        nn.init.zeros_(self.tab_encoder.era_emb.weight)
        self._init_pwe_hazard_bias()

    def _init_pwe_hazard_bias(self) -> None:
        # Rates are per day. Initialize near quarterly next-contact hazards and
        # rare absorbing hazards so early training is not dominated by PWE NLL.
        prior = torch.full((self.config.n_pwe_causes, self.config.n_pwe_bins), -10.0)
        prior[0, :] = -4.4
        prior[1, :] = -9.0
        prior[2, :] = -10.5
        with torch.no_grad():
            for head in (self.pwe_log_lambda_pre_head, self.pwe_log_lambda_contact_head, self.pwe_log_lambda_post_head):
                head.bias.copy_(prior.reshape(-1))

    def _teacher_forced_next_context(self, z_post: torch.Tensor, batch: dict[str, torch.Tensor]) -> torch.Tensor:
        """Condition generated next-visit observations on sampled/teacher-forced Δt only.

        The previous version also injected the true `service_state[t+1]`, which
        made `active_state_head` a label copier. A_t is now predicted from
        z_post plus the generated/teacher-forced time interval, and no next
        service label enters any deployable risk path.
        """

        delta_ctx = self.next_delta_condition(batch["delta_t_next_log"].unsqueeze(-1).float())
        return z_post + delta_ctx

    def _heads_from_states(self, z_history: torch.Tensor, z_contact: torch.Tensor, z_post: torch.Tensor, batch: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
        bsz, seq_len, _ = z_post.shape
        tf_ctx = self._teacher_forced_next_context(z_post, batch)
        pwe_pre = self.pwe_log_lambda_pre_head(z_history).view(bsz, seq_len, self.config.n_pwe_causes, self.config.n_pwe_bins)
        pwe_contact = self.pwe_log_lambda_contact_head(z_contact).view(bsz, seq_len, self.config.n_pwe_causes, self.config.n_pwe_bins)
        pwe_post = self.pwe_log_lambda_post_head(z_post).view(bsz, seq_len, self.config.n_pwe_causes, self.config.n_pwe_bins)
        deployable_pre = self.deployable_risk_pre_head(z_history)
        deployable_contact = self.deployable_risk_contact_head(torch.cat([z_history, z_contact], dim=-1))
        deployable_post = self.deployable_risk_post_head(z_post)
        ontology_event_logits = self.ontology_event_head(tf_ctx)
        out = {
            "z_history": z_history,
            "z_contact": z_contact,
            "z_post": z_post,
            "z_minus": z_history,
            "z_plus": z_post,
            "pwe_log_lambda_pre": pwe_pre,
            "pwe_log_lambda_contact": pwe_contact,
            "pwe_log_lambda_post": pwe_post,
            "pwe_log_lambda": pwe_post,
            "active_state_logits": self.active_state_head(tf_ctx),
            "missingness_logits": self.missingness_head(tf_ctx).view(bsz, seq_len, self.config.n_cat_fields, self.config.n_missing),
            "risk_event_logits": deployable_contact,
            "deployable_risk_logits": deployable_contact,
            "deployable_risk_logits_pre": deployable_pre,
            "deployable_risk_logits_contact": deployable_contact,
            "deployable_risk_logits_post": deployable_post,
            "field_logits": self.field_head(tf_ctx),
            "numeric_mu": self.numeric_head(tf_ctx)[..., : self.config.n_numeric_fields],
            "ordinal_cum_logits": self.ordinal_head(tf_ctx),
            "ontology_event_logits": ontology_event_logits,
            "history_event_logits": self.history_event_head(z_history),
            "contact_event_logits": self.contact_event_head(z_contact),
            "color_cum_logits": self.color_signal_head(tf_ctx),
            # Backward-compatible aliases used by generic v2/v3-style export or
            # risk-head scripts must remain deployable. Keep the teacher-forced
            # ontology head available only under its explicit diagnostic name.
            "event_logits": deployable_contact,
            "future_risk_logits": self.history_event_head(z_history),
        }
        pwe_ci = pwe_closed_form_cif(pwe_post, torch.tensor(PWE_BIN_EDGES_DAYS[:-1], dtype=z_post.dtype, device=z_post.device))
        out["pwe_bin_survival_probs"] = pwe_ci["survival"]
        out["pwe_bin_cif_probs"] = pwe_ci["cif"]
        return out

    def _rollout_prefix_embedding(self, out: dict[str, torch.Tensor], temperature: float) -> torch.Tensor:
        cause = F.gumbel_softmax(out["pwe_log_lambda"].amax(dim=-1), tau=max(temperature, 1.0e-3), hard=False, dim=-1)
        active = F.gumbel_softmax(out["active_state_logits"], tau=max(temperature, 1.0e-3), hard=False, dim=-1)
        event = torch.sigmoid(out["risk_event_logits"])
        delta_mu = torch.log1p(F.softplus(out["pwe_log_lambda"]).sum(dim=(2, 3)).reciprocal().clamp(max=365.0))
        next_contact_prob = cause[..., :1]
        features = torch.cat([cause, active, event, delta_mu.unsqueeze(-1), next_contact_prob], dim=-1)
        return self.rollout_prefix_proj(features)

    def forward(
        self,
        batch: dict[str, torch.Tensor],
        rollout_steps: int = 1,
        teacher_forcing_rate: float = 1.0,
        gumbel_temperature: float = 1.0,
    ) -> dict[str, torch.Tensor]:
        u, conditional_status = self.tab_encoder(
            batch["cat_value_ids"],
            batch["missing_ids"],
            batch["numeric_values"],
            batch["numeric_mask"],
            batch["ordinal_cbe"],
            batch["ordinal_mask"],
            batch["drug_name_ids"],
            batch["drug_class_ids"],
            batch["drug_mask"],
            batch["service_state"],
            batch["visit_year"],
        )
        z_history, z_post = self.traj_decoder(u, batch["service_state"], batch["visit_year"], batch["time_since_start_days"], batch["valid_mask"])
        static_vec = self.static_encoder(batch.get("static_value_ids"), z_post)
        static_injection = self.static_injection_scale * static_vec
        z_history = z_history + static_injection[:, None, :]
        z_post = z_post + static_injection[:, None, :]
        z_contact = self.contact_injector(z_history, batch)
        out = self._heads_from_states(z_history, z_contact, z_post, batch)
        out["u"] = u
        out["conditional_status"] = conditional_status
        out["static_vec"] = static_vec
        out["static_injection_scale"] = self.static_injection_scale
        out["contact_injection_scale"] = self.contact_injector.scale

        if rollout_steps > 1:
            # Differentiable scheduled-sampling prefix used for Stage 2b/2c.
            # It re-enters the causal decoder with predicted visit embeddings;
            # raw Stage0 field resynthesis can replace this hook later without
            # changing the loss interface.
            rollout_logits: list[torch.Tensor] = []
            rollout_source = out
            eps = float(max(0.0, min(1.0, teacher_forcing_rate)))
            for _ in range(2, int(rollout_steps) + 1):
                prefix = self._rollout_prefix_embedding(rollout_source, gumbel_temperature)
                shifted_prefix = torch.cat([u[:, :1, :], prefix[:, :-1, :]], dim=1)
                u_mix = eps * u + (1.0 - eps) * shifted_prefix
                rz_history, rz_post = self.traj_decoder(
                    u_mix,
                    batch["service_state"],
                    batch["visit_year"],
                    batch["time_since_start_days"],
                    batch["valid_mask"],
                )
                rz_history = rz_history + static_injection[:, None, :]
                rz_post = rz_post + static_injection[:, None, :]
                rz_contact = self.contact_injector(rz_history, batch)
                r_out = self._heads_from_states(rz_history, rz_contact, rz_post, batch)
                rollout_source = r_out
                rollout_logits.append(r_out["pwe_log_lambda"])
                out["_last_rollout_context"] = r_out["z_post"]
            out["rollout_pwe_log_lambda"] = torch.stack(rollout_logits, dim=0) if rollout_logits else out["pwe_log_lambda"].new_zeros(0)
        return out

    @torch.no_grad()
    def predict_fixed_horizon_risk(
        self,
        batch: dict[str, torch.Tensor],
        horizons_days: torch.Tensor | list[float],
        mode: str = "contact",
    ) -> dict[str, torch.Tensor]:
        out = self.forward(batch)
        if mode in {"pre_visit", "history", "pre"}:
            log_lambda = out["pwe_log_lambda_pre"]
        elif mode in {"contact", "contact_present"}:
            log_lambda = out["pwe_log_lambda_contact"]
        elif mode in {"field_end", "post", "post_visit"}:
            log_lambda = out["pwe_log_lambda_post"]
        else:
            raise ValueError(f"unknown fixed-horizon risk mode: {mode}")
        h = torch.as_tensor(horizons_days, dtype=log_lambda.dtype, device=log_lambda.device)
        return pwe_closed_form_cif(log_lambda, h)


def pwe_closed_form_cif(log_lambda_kj: torch.Tensor, horizons_days: torch.Tensor) -> dict[str, torch.Tensor]:
    """Closed-form PWE survival and cause-specific cumulative incidence.

    Args:
        log_lambda_kj: [..., K, J] raw log-rate parameters.
        horizons_days: [H] positive horizons in days.
    """

    if horizons_days.ndim == 0:
        horizons_days = horizons_days[None]
    # CIFs are probabilities; computing the closed form in bf16/fp16 can make
    # cause-specific masses sum above 1 when next-contact hazards dominate.
    # Keep training/export autocast enabled around the model, but evaluate this
    # probability algebra in fp32 and enforce the total CIF <= 1 - S(t).
    calc_dtype = torch.float32 if log_lambda_kj.dtype in {torch.float16, torch.bfloat16} else log_lambda_kj.dtype
    lambda_kj = F.softplus(log_lambda_kj.to(calc_dtype)) + 1.0e-8
    lambda_j = lambda_kj.sum(dim=-2)
    lower = torch.tensor(PWE_FINITE_LOWER_DAYS, dtype=lambda_kj.dtype, device=lambda_kj.device)
    upper = torch.tensor(PWE_BIN_EDGES_DAYS[:-1], dtype=lambda_kj.dtype, device=lambda_kj.device)
    h = horizons_days.to(dtype=lambda_kj.dtype, device=lambda_kj.device).clamp(min=0.0)
    finite_exposure = (torch.minimum(h[:, None], upper[None, :]) - lower[:-1][None, :]).clamp(min=0.0)
    last_exposure = (h - lower[-1]).clamp(min=0.0)[:, None]
    exposure = torch.cat([finite_exposure, last_exposure], dim=1)
    base_ndim = lambda_j.ndim - 1
    hazard_exposure = lambda_j.unsqueeze(-2) * exposure.reshape((1,) * base_ndim + exposure.shape)
    cum_before = torch.cumsum(hazard_exposure, dim=-1) - hazard_exposure
    survival_before = torch.exp(-cum_before)
    event_in_bin = -torch.expm1(-hazard_exposure)
    share = lambda_kj / lambda_j.unsqueeze(-2).clamp(min=1.0e-8)
    cif = (share.unsqueeze(-3) * survival_before.unsqueeze(-2) * event_in_bin.unsqueeze(-2)).sum(dim=-1)
    survival = torch.exp(-hazard_exposure.sum(dim=-1)).clamp(min=0.0, max=1.0)
    max_cif_total = (1.0 - survival).unsqueeze(-1)
    cif_total = cif.sum(dim=-1, keepdim=True)
    scale = torch.where(cif_total > max_cif_total, max_cif_total / cif_total.clamp(min=1.0e-12), torch.ones_like(cif_total))
    cif = (cif * scale).clamp(min=0.0, max=1.0)
    return {"survival": survival, "cif": cif}


__all__ = [
    "SCTMv4p3",
    "SCTMv4p3Config",
    "PWE_BIN_EDGES_DAYS",
    "build_service_attention_prior",
    "era_tokens",
    "pwe_closed_form_cif",
    "previous_interval_log",
]