validation_agent/core/feasibility_judge.py

"""FeasibilityJudge — V0/V1/V2/V3 gate per 02_protocol.md §2-§5.

Inputs: ``NormalizedEdge`` from :mod:`validation_agent.core.edge_loader`.
Outputs: ``FeasibilityReport`` matching ``feasibility.schema.v0.2``
(``validation_agent/schemas/feasibility.schema.json``).

Gates implemented:
    V0 — relies on EdgeLoader (already passed by the time edge reaches here).
    V1 — Mappability: status ∈ {exact, close, tentative}; dataset/field
         exist in v7; medications/dietary excluded (memory feedback-physiological-only);
         pid space joinability ≥ 5 (04 §10.1.1).
    V2 — Container filter: when Y (or X) is a container column, validate
         upstream value_filter OR infer from edge text (Y / Y_original_text /
         Y_qualifiers.definition).
    V3 — Equation inference: from Y_qualifiers.measurement_scale +
         X_contrast_type → equation_type + candidate_estimator; reject
         if estimator not in M1 supported set.

Out of scope (delegated):
    - V4/V5/V6 (Plan / Codegen / Result schema gates) live in their own modules.
    - Data layer inspection (n_per_stratum, value distribution) → DataInspector.

Dataset-selection anchor (generic, reserved — NOT yet implemented; North Star
SPEC §1 "选在哪个数据集上验证"):
    This gate judges runnability against the *single registered* validation target
    — HPP, via its v7 synthetic isomorph (``v7_root`` / ``exists_in_v7`` /
    ``hpp_to_v7_path``). The design is dataset-GENERIC: feasibility is conceptually
    evaluated *per candidate dataset*, and a design-period, data-/inventory-driven
    ``DatasetSelector`` (**never a rule table**) chooses which feasible dataset(s) to
    validate on once the data is inventoried. M1a registry = {HPP} → selection is the
    identity. Adding a dataset later = register it + run this same gate against its
    inventory + let the selector pick; no rule-based / paper-specific branch is ever
    introduced. (per 01 §3.2 anchor.)

The reasoning text composed in §_compose_reasoning is human-readable and used
by Planner / PlanReviewer audit; not parsed downstream.
"""
from __future__ import annotations

import json as _json
import re
from dataclasses import dataclass, field
from datetime import datetime, timezone
from pathlib import Path
from typing import Any

import pyarrow.parquet as pq

from validation_agent import __metadata_version__
from validation_agent.configs import excluded_dataset_reasons, excluded_value_filter_reasons
from validation_agent.core.edge_loader import CONTAINER_FIELDS, NormalizedEdge
from validation_agent.path_mapping import (
    DEFAULT_V7_ROOT,
    hpp_to_v7_path,
    split_dataset_subtable,
    strip_hpp_prefix,
)


# ===========================================================================
# Domain constants
# ===========================================================================

# Exclusions (medications / dietary / behavioral self-report) are loaded per-instance
# from configs/excluded_datasets.json (11 §6) — the single source of truth — so they
# are no longer hardcoded here. See FeasibilityJudge._exclusion_blocker.

# 02 §5.1 equation inference table.
# Keys: (Y_qualifiers.measurement_scale, X_contrast_type) — "any" matches any contrast.
# Values: (equation_type, candidate_estimator, milestone).
_EQUATION_INFERENCE: dict[tuple[str, str], tuple[str, str, str]] = {
    ("continuous", "level_contrast"): ("E1-linear-binary", "e1_linear", "M1"),
    ("continuous", "unit_increment"): ("E1-linear", "e1_linear", "M1"),
    ("continuous", "continuous_linear"): ("E1-linear", "e1_linear", "M1"),
    ("continuous", "category_vs_reference"): ("E1-linear-category", "e1_linear", "M1"),
    ("continuous", "threshold"): ("E1-linear-binary", "e1_linear", "M1"),
    ("continuous", "quantile"): ("E1-linear-category", "e1_linear", "M1"),
    ("continuous", "per_sd"): ("E1-linear", "e1_linear", "M1"),
    ("continuous", "any"): ("E1-linear", "e1_linear", "M1"),
    ("binary", "any"): ("E1-logistic-binary", "e1_logistic", "M1"),
    ("proportion", "any"): ("E1-logistic-binary", "e1_logistic", "M1"),
    ("time_to_event", "any"): ("E2-cox", "e2_cox", "M2"),
    ("count", "any"): ("E1-poisson", "e1_poisson", "M2"),
    ("rate", "any"): ("E1-poisson", "e1_poisson", "M2"),
}

# Estimators registered in M1 (per 04 §3-§4 + configs/registry/).
_M1_SUPPORTED_ESTIMATORS = frozenset({"e1_linear", "e1_logistic"})

# Y text qualifier prefixes commonly stripped before value_filter inference.
_Y_PREFIXES_TO_STRIP = (
    "early-onset ",
    "late-onset ",
    "incident ",
    "current ",
    "history of ",
    "diagnosed with ",
    "new-onset ",
    "recurrent ",
)


# ===========================================================================
# Output dataclass
# ===========================================================================


@dataclass
class FeasibilityReport:
    """Structured V0/V1/V2/V3 result.

    The ``to_json_dict`` projection emits a dict that validates against
    ``feasibility.schema.v0.2`` (validation_agent/schemas/feasibility.schema.json).
    """

    edge_id: str
    judged_at: str
    metadata_version: str
    is_runnable: bool
    blockers: list[str]
    warnings: list[str]
    x_mapping: dict[str, Any]
    y_mapping: dict[str, Any]
    z_summary: dict[str, int]
    inferred_equation_type: str | None = None
    candidate_estimator: str | None = None
    reasoning: str = ""
    data_gap: dict[str, Any] | None = None
    longitudinal: dict[str, Any] | None = None
    # Internal extras (not part of feasibility.schema.json):
    z_details: list[dict[str, Any]] = field(default_factory=list)
    pid_overlap_count: int | None = None

    def to_json_dict(self) -> dict[str, Any]:
        """Serialize to feasibility.schema.v0.2 shape (no internal extras)."""
        return {
            "edge_id": self.edge_id,
            "judged_at": self.judged_at,
            "metadata_version": self.metadata_version,
            "is_runnable": self.is_runnable,
            "blockers": list(self.blockers),
            "warnings": list(self.warnings),
            "x_mapping": _project_role_mapping(self.x_mapping),
            "y_mapping": _project_role_mapping(self.y_mapping),
            "z_summary": dict(self.z_summary),
            "inferred_equation_type": self.inferred_equation_type,
            "candidate_estimator": self.candidate_estimator,
            "reasoning": self.reasoning,
            "data_gap": self.data_gap,
            "longitudinal": self.longitudinal,
        }


def _project_role_mapping(m: dict[str, Any]) -> dict[str, Any]:
    """Drop internal keys (sub_table, parquet_path) before schema-validation."""
    allowed = {
        "status",
        "dataset",
        "field",
        "exists_in_v7",
        "is_container",
        "value_filter_required",
        "value_filter_present",
        "inferred_value_filter",
        "v7_match_count_expected",
    }
    return {k: v for k, v in m.items() if k in allowed}


# ===========================================================================
# Main class
# ===========================================================================


class FeasibilityJudge:
    """V0/V1/V2/V3 gate evaluator. Caches parquet schemas + pid sets in-process."""

    def __init__(
        self,
        v7_root: Path | str = DEFAULT_V7_ROOT,
        *,
        metadata_version: str = __metadata_version__,
        min_pid_overlap: int = 5,
        supported_estimators: frozenset[str] = _M1_SUPPORTED_ESTIMATORS,
    ) -> None:
        # ── ANCHOR (01 §3.2, generic dataset selection) ──────────────────────
        # v7_root identifies the *current single* validation target — HPP via its
        # v7 synthetic isomorph. Generalizing to a dataset registry + a design-period,
        # data-/inventory-driven DatasetSelector (never a rule table) is the reserved,
        # non-rule-based extension point (see module docstring). M1a registry = {HPP}.
        self.v7_root = Path(v7_root)
        self.metadata_version = metadata_version
        self.min_pid_overlap = min_pid_overlap
        self.supported_estimators = supported_estimators
        # Exclusion source of truth: configs/excluded_datasets.json (11 §6).
        # bare dataset name -> blocker code; value_filter.type -> blocker code.
        self._excluded_datasets: dict[str, str] = excluded_dataset_reasons()
        self._excluded_vf_types: dict[str, str] = excluded_value_filter_reasons()
        self._parquet_columns_cache: dict[str, frozenset[str]] = {}
        self._parquet_pids_cache: dict[str, frozenset[Any]] = {}
        self._parquet_stages_cache: dict[str, list[str]] = {}

    # ---- public API ---------------------------------------------------

    def judge(self, edge: NormalizedEdge) -> FeasibilityReport:
        blockers: list[str] = []
        warnings: list[str] = []

        x_mapping = self._check_role(edge, "X", blockers)
        y_mapping = self._check_role(edge, "Y", blockers)

        # V1.x — pid joinability (only when both roles resolve to v7 paths)
        pid_overlap_count: int | None = None
        if x_mapping.get("exists_in_v7") and y_mapping.get("exists_in_v7"):
            pid_overlap_count = self._pid_overlap(
                x_mapping["__full_id"], y_mapping["__full_id"]
            )
            if pid_overlap_count < self.min_pid_overlap:
                blockers.append("v1_pid_space_disjoint")

        # V2 — container value_filter handling (Y only; X is rarely container in M1)
        for role, mapping in (("X", x_mapping), ("Y", y_mapping)):
            self._handle_container_filter(edge, role, mapping, blockers, warnings)

        # V3 — equation_type inference (only when Y was structurally mappable)
        eq_type, estimator = self._infer_equation_type(edge, y_mapping)
        if estimator is not None and estimator not in self.supported_estimators:
            blockers.append("v3_estimator_not_registered_in_m1")
        elif eq_type is None and y_mapping.get("exists_in_v7"):
            blockers.append("v3_cannot_infer_equation_type")

        # V3b — E3-longitudinal substrate (additive; independent of the cross-sectional
        # equation above). Runnable when a v7-resolved role table carries >=2 research_stage
        # timepoints (e3_change for 2 waves, e3_lmm for >=3). Adds NO blocker — it is an
        # additive analysis path the planner may plan as a within-subject change dose_response.
        longitudinal = self._build_longitudinal(x_mapping, y_mapping)

        # Z mapping summary + per-item categorization
        z_summary, z_details = self._categorize_z(edge)

        # data_gap pruning when not runnable
        data_gap = self._build_data_gap(edge, x_mapping, y_mapping, blockers)

        reasoning = self._compose_reasoning(
            edge,
            x_mapping=x_mapping,
            y_mapping=y_mapping,
            z_summary=z_summary,
            eq_type=eq_type,
            estimator=estimator,
            blockers=blockers,
            warnings=warnings,
            pid_overlap=pid_overlap_count,
        )

        is_runnable = not blockers

        return FeasibilityReport(
            edge_id=edge.edge_id,
            judged_at=datetime.now(timezone.utc).isoformat(),
            metadata_version=self.metadata_version,
            is_runnable=is_runnable,
            blockers=blockers,
            warnings=warnings,
            x_mapping=x_mapping,
            y_mapping=y_mapping,
            z_summary=z_summary,
            inferred_equation_type=eq_type,
            candidate_estimator=estimator,
            reasoning=reasoning,
            data_gap=data_gap,
            longitudinal=longitudinal,
            z_details=z_details,
            pid_overlap_count=pid_overlap_count,
        )

    # ---- V1: per-role mappability ------------------------------------

    def _check_role(
        self,
        edge: NormalizedEdge,
        role: str,
        blockers: list[str],
    ) -> dict[str, Any]:
        role_code = role.lower()
        m = edge.hpp_mapping.get(role, {}) or {}
        status = m.get("status")
        dataset = m.get("dataset")
        field_name = m.get("field")

        out: dict[str, Any] = {
            "status": status,
            "dataset": dataset,
            "field": field_name,
            "exists_in_v7": False,
            "is_container": field_name in CONTAINER_FIELDS if field_name else False,
            "value_filter_required": False,
            # Internals (stripped before schema serialization):
            "__full_id": None,
            "__parquet_path": None,
        }

        # Physiological-only scope is absolute (memory feedback-physiological-only +
        # 11 §6 / decision #8): medications / dietary / behavioral self-report are
        # excluded REGARDLESS of mapping status, BEFORE the status guard, so an
        # unmapped excluded.* edge is still attributed to its exclusion code (not the
        # generic v1_x_not_mappable). Source of truth: configs/excluded_datasets.json.
        excl = self._exclusion_blocker(dataset, m.get("value_filter"))
        if excl:
            blockers.append(excl)
            return out

        if status not in ("exact", "close", "tentative"):
            blockers.append(f"v1_{role_code}_not_mappable")
            return out

        if not dataset or not field_name:
            blockers.append(f"v1_{role_code}_not_mappable")
            return out

        try:
            parquet_path = hpp_to_v7_path(dataset, v7_root=self.v7_root)
        except ValueError:
            blockers.append(f"v1_{role_code}_field_not_in_v7")
            return out

        columns = self._parquet_columns(parquet_path)
        if field_name not in columns:
            blockers.append(f"v1_{role_code}_field_not_in_v7")
            return out

        out["exists_in_v7"] = True
        out["is_container"] = field_name in CONTAINER_FIELDS
        out["value_filter_required"] = out["is_container"]
        out["__full_id"] = dataset
        out["__parquet_path"] = parquet_path
        # Reflect upstream value_filter presence for downstream introspection.
        vf = m.get("value_filter")
        if isinstance(vf, dict) and isinstance(vf.get("include"), list) and vf.get("include"):
            out["value_filter_present"] = True
        elif vf is None:
            out["value_filter_present"] = False
        else:
            out["value_filter_present"] = False
        return out

    # ---- V2: container value_filter -----------------------------------

    def _handle_container_filter(
        self,
        edge: NormalizedEdge,
        role: str,
        mapping: dict[str, Any],
        blockers: list[str],
        warnings: list[str],
    ) -> None:
        if not mapping.get("is_container") or not mapping.get("exists_in_v7"):
            return
        role_code = role.lower()
        m = edge.hpp_mapping.get(role, {}) or {}
        vf = m.get("value_filter")
        # Upstream-provided dict (02 §4.1/§4.3): each failure maps to exactly ONE code.
        if isinstance(vf, dict):
            include = vf.get("include")
            if isinstance(include, list) and include:
                # Validate match enum (02 §4.2).
                match_kind = vf.get("match")
                if match_kind not in (None, "exact", "contains", "regex", "code_prefix"):
                    self._mark_v2_blocker(mapping, blockers, "v2_unsupported_filter_match")
                return
            if isinstance(include, list):
                # include present but empty → distinct code (02 §4.3).
                self._mark_v2_blocker(mapping, blockers, "v2_value_filter_include_empty")
                return
            # dict without a usable include list → value_filter structurally absent.
            self._mark_v2_blocker(mapping, blockers, f"v2_{role_code}_missing_value_filter")
            return
        # Upstream-provided plain string is acceptable (loose typing in gpt5.5).
        if isinstance(vf, str) and vf.strip():
            mapping["inferred_value_filter"] = vf.strip()
            return
        # No upstream filter → try inference from edge text (02 §4.4 flow).
        inferred = self._infer_container_filter(edge, role)
        if inferred:
            mapping["inferred_value_filter"] = inferred
            mapping["v7_match_count_expected"] = 0
            warnings.append(f"v7_{role_code}_value_filter_expected_to_match_zero")
            return
        # Inference failed → single canonical code (02 §4.4 step 4), not two.
        self._mark_v2_blocker(mapping, blockers, "v2_cannot_infer_value_filter")

    @staticmethod
    def _mark_v2_blocker(mapping: dict[str, Any], blockers: list[str], code: str) -> None:
        """Append a V2 blocker and stamp it on the role mapping (for data_gap
        attribution). The internal ``_v2_blocker`` key is stripped before schema
        serialization (see ``_project_role_mapping``)."""
        blockers.append(code)
        mapping["_v2_blocker"] = code

    def _infer_container_filter(self, edge: NormalizedEdge, role: str) -> str | None:
        eqf = edge.equation_formula_reported
        if role == "Y":
            candidates = [
                eqf.get("Y"),
                eqf.get("Y_original_text"),
                (eqf.get("Y_qualifiers") or {}).get("definition"),
            ]
        else:
            candidates = [
                eqf.get("X"),
                eqf.get("X_original_text"),
                (eqf.get("X_qualifiers") or {}).get("definition"),
            ]
        for src in candidates:
            if not isinstance(src, str):
                continue
            text = src.strip()
            if not text:
                continue
            lowered = text.lower()
            for prefix in _Y_PREFIXES_TO_STRIP:
                if lowered.startswith(prefix):
                    text = text[len(prefix):]
                    lowered = lowered[len(prefix):]
                    break
            # Truncate to ≤ 6 words to keep the filter compact.
            words = text.split()
            if not words:
                continue
            return " ".join(words[: min(len(words), 6)])
        return None

    # ---- V3: equation inference ---------------------------------------

    def _infer_equation_type(
        self,
        edge: NormalizedEdge,
        y_mapping: dict[str, Any],
    ) -> tuple[str | None, str | None]:
        if not y_mapping.get("exists_in_v7"):
            return None, None
        yq = edge.equation_formula_reported.get("Y_qualifiers") or {}
        scale: str | None = yq.get("measurement_scale")
        if not scale and y_mapping.get("is_container"):
            # 02 §5.1 fallback: container Y treated as binary prevalence.
            scale = "binary"
        if not scale:
            return None, None
        xct = edge.equation_formula_reported.get("X_contrast_type")
        candidate = _EQUATION_INFERENCE.get((scale, xct))
        if candidate is None:
            candidate = _EQUATION_INFERENCE.get((scale, "any"))
        if candidate is None:
            return None, None
        eq_type, estimator, _milestone = candidate
        return eq_type, estimator

    # ---- Z categorization (no L1 blocker; produces z_summary + details)

    def _categorize_z(
        self,
        edge: NormalizedEdge,
    ) -> tuple[dict[str, int], list[dict[str, Any]]]:
        z_items = edge.hpp_mapping.get("Z") or []
        details: list[dict[str, Any]] = []
        n_mapped = 0
        n_fuzzy = 0
        n_unmapped = 0
        for zm in z_items:
            if not isinstance(zm, dict):
                continue
            d = self._categorize_z_item(zm)
            details.append(d)
            cat = d["category"]
            if cat == "mapped":
                n_mapped += 1
            elif cat == "fuzzy":
                n_fuzzy += 1
            else:
                n_unmapped += 1
        return (
            {
                "n_total": len(z_items),
                "n_mapped": n_mapped,
                "n_fuzzy": n_fuzzy,
                "n_unmapped": n_unmapped,
            },
            details,
        )

    def _categorize_z_item(self, zm: dict[str, Any]) -> dict[str, Any]:
        concept = zm.get("name") or zm.get("concept") or "<unknown>"
        dataset = zm.get("dataset")
        field_name = zm.get("field")
        status = zm.get("status")
        out: dict[str, Any] = {
            "concept": concept,
            "dataset": dataset,
            "field": field_name,
        }

        # Exclusions first (11 §6 / decision #8): medications / dietary / behavioral
        # self-report. Z reason = blocker code without the v1_ prefix
        # (medications_excluded / dietary_excluded / self_report_excluded).
        excl = self._exclusion_blocker(dataset, zm.get("value_filter"))
        if excl:
            out["category"] = "unmapped"
            out["reason"] = excl[3:] if excl.startswith("v1_") else excl
            return out

        if status not in ("exact", "close", "tentative"):
            out["category"] = "unmapped"
            out["reason"] = "unmapped"
            return out

        if dataset is None:
            out["category"] = "unmapped"
            out["reason"] = "unmapped"
            return out

        # v7 dataset existence check
        try:
            parquet_path = hpp_to_v7_path(dataset, v7_root=self.v7_root)
        except ValueError:
            bare = strip_hpp_prefix(dataset).split(".", 1)[0]
            out["category"] = "unmapped"
            out["reason"] = f"v7_dataset_not_available_{bare}"
            return out

        # field=null while dataset exists → fuzzy (02 §7.2)
        if not field_name:
            out["category"] = "fuzzy"
            out["reason"] = "fuzzy_mapping_field_null_cannot_resolve"
            return out

        columns = self._parquet_columns(parquet_path)
        if field_name not in columns:
            out["category"] = "unmapped"
            out["reason"] = "unmapped"
            return out

        # Container columns require a value_filter; without one, can't enter
        # adjustment_set as a regular Z (02 §7.2).
        if field_name in CONTAINER_FIELDS:
            vf = zm.get("value_filter")
            has_filter = (
                isinstance(vf, dict) and isinstance(vf.get("include"), list) and vf.get("include")
            ) or (isinstance(vf, str) and vf.strip())
            if not has_filter:
                out["category"] = "unmapped"
                out["reason"] = "container_column_no_inferable_filter"
                return out

        out["category"] = "mapped"
        out["mapping_quality"] = "mapped"
        return out

    # ---- helpers: dataset classifiers + parquet IO --------------------

    def _exclusion_blocker(self, dataset: str | None, value_filter: Any) -> str | None:
        """Return the V1 blocker code if (dataset, value_filter) is permanently
        excluded per configs/excluded_datasets.json (11 §6: medications / dietary /
        behavioral self-report), else None. Single source of truth — no hardcoding.
        """
        if isinstance(dataset, str):
            bare = strip_hpp_prefix(dataset).split(".", 1)[0]
            code = self._excluded_datasets.get(bare)
            if code:
                return code
        if isinstance(value_filter, dict):
            t = value_filter.get("type")
            if isinstance(t, str):
                code = self._excluded_vf_types.get(t)
                if code:
                    return code
        return None

    def _parquet_columns(self, parquet_path: str) -> frozenset[str]:
        if parquet_path in self._parquet_columns_cache:
            return self._parquet_columns_cache[parquet_path]
        schema = pq.read_schema(parquet_path)
        cols: set[str] = set(schema.names)
        # Pandas-metadata index columns (MultiIndex) — these become regular
        # columns after .reset_index() in run.py, so they're addressable.
        meta = schema.metadata or {}
        pandas_md = meta.get(b"pandas")
        if pandas_md:
            try:
                pmd = _json.loads(pandas_md.decode("utf-8"))
            except (UnicodeDecodeError, _json.JSONDecodeError):
                pmd = None
            if isinstance(pmd, dict):
                for ix in pmd.get("index_columns", []) or []:
                    if isinstance(ix, str):
                        cols.add(ix)
                    elif isinstance(ix, dict):
                        name = ix.get("name")
                        if isinstance(name, str):
                            cols.add(name)
        result = frozenset(cols)
        self._parquet_columns_cache[parquet_path] = result
        return result

    def _parquet_pids(self, full_dataset_id: str) -> frozenset[Any]:
        if full_dataset_id in self._parquet_pids_cache:
            return self._parquet_pids_cache[full_dataset_id]
        parquet_path = hpp_to_v7_path(full_dataset_id, v7_root=self.v7_root)
        # Use pyarrow to peek the column without loading everything else.
        table = pq.read_table(parquet_path, columns=None)
        # Convert to pandas to leverage reset_index for MultiIndex.
        df = table.to_pandas()
        if df.index.names != [None]:
            df = df.reset_index()
        if "participant_id" not in df.columns:
            result = frozenset()
        else:
            pids = df["participant_id"].dropna().unique().tolist()
            result = frozenset(pids)
        self._parquet_pids_cache[full_dataset_id] = result
        return result

    def _research_stages(self, parquet_path: str) -> list[str]:
        """Sorted distinct research_stage values on a role parquet (mirrors
        DataInventory._distinct_strings(df, 'research_stage')). [] when the column is absent."""
        if parquet_path in self._parquet_stages_cache:
            return self._parquet_stages_cache[parquet_path]
        table = pq.read_table(parquet_path, columns=None)
        df = table.to_pandas()
        if df.index.names != [None]:
            df = df.reset_index()
        if "research_stage" not in df.columns:
            result: list[str] = []
        else:
            vals = df["research_stage"].dropna().unique().tolist()
            result = sorted(str(v) for v in vals)
        self._parquet_stages_cache[parquet_path] = result
        return result

    def _build_longitudinal(
        self, x_mapping: dict[str, Any], y_mapping: dict[str, Any]
    ) -> dict[str, Any] | None:
        """E3-longitudinal feasibility (02 §5.1 extension / REBUILD E3 row). Reads the distinct
        research_stage count of the v7-resolved role tables; picks the role with the MOST stages
        (so a static annotation role can't mask a longitudinal one). >=3 -> e3_lmm (LMM slope),
        ==2 -> e3_change (change-score), <2 -> not_runnable. Returns None when no role resolves
        to v7 (cannot inspect timepoints). Aligns with cev/schemas/feasibility_per_equation."""
        stages: list[str] = []
        for mapping in (y_mapping, x_mapping):
            if mapping.get("exists_in_v7") and mapping.get("__parquet_path"):
                s = self._research_stages(mapping["__parquet_path"])
                if len(s) > len(stages):
                    stages = s
        if not stages:
            return None
        n = len(stages)
        if n >= 3:
            status, estimator, missing = "ok", "e3_lmm", []
        elif n == 2:
            status, estimator, missing = "ok", "e3_change", []
        else:
            status, estimator, missing = "not_runnable", None, ["research_stage_single_timepoint"]
        return {
            "equation_type": "E3-longitudinal",
            "timepoints": stages,
            "n_timepoints": n,
            "candidate_estimator": estimator,
            "status": status,
            "missing": missing,
        }

    def _pid_overlap(self, x_full_id: str, y_full_id: str) -> int:
        return len(self._parquet_pids(x_full_id) & self._parquet_pids(y_full_id))

    # ---- data_gap / reasoning ----------------------------------------

    def _build_data_gap(
        self,
        edge: NormalizedEdge,
        x_mapping: dict[str, Any],
        y_mapping: dict[str, Any],
        blockers: list[str],
    ) -> dict[str, Any] | None:
        if not blockers:
            return None
        eqf = edge.equation_formula_reported
        x_reason = self._role_reason(x_mapping, blockers, "x")
        y_reason = self._role_reason(y_mapping, blockers, "y")
        return {
            "x_concept": eqf.get("X", "") or "",
            "x_reason": x_reason,
            "y_concept": eqf.get("Y", "") or "",
            "y_reason": y_reason,
            "recommended_v7_extensions": [],
        }

    @staticmethod
    def _role_reason(mapping: dict[str, Any], blockers: list[str], role_code: str) -> str:
        """data_gap reason for one role: 'ok' only if it both exists in v7 AND
        cleared the V2 container check; otherwise the attributable blocker."""
        if not mapping.get("exists_in_v7"):
            return FeasibilityJudge._pick_role_blocker(blockers, role_code)
        v2_blocker = mapping.get("_v2_blocker")
        return v2_blocker if v2_blocker else "ok"

    @staticmethod
    def _pick_role_blocker(blockers: list[str], role_code: str) -> str:
        for b in blockers:
            if b.startswith(f"v1_{role_code}_") or b.startswith(f"v2_{role_code}_"):
                return b
        return "not_runnable"

    @staticmethod
    def _compose_reasoning(
        edge: NormalizedEdge,
        *,
        x_mapping: dict[str, Any],
        y_mapping: dict[str, Any],
        z_summary: dict[str, int],
        eq_type: str | None,
        estimator: str | None,
        blockers: list[str],
        warnings: list[str],
        pid_overlap: int | None,
    ) -> str:
        parts: list[str] = []
        parts.append(
            f"X status={x_mapping.get('status')} -> "
            f"{x_mapping.get('dataset')}.{x_mapping.get('field')} "
            f"(exists_in_v7={x_mapping.get('exists_in_v7')});"
        )
        parts.append(
            f"Y status={y_mapping.get('status')} -> "
            f"{y_mapping.get('dataset')}.{y_mapping.get('field')} "
            f"(exists_in_v7={y_mapping.get('exists_in_v7')}, "
            f"is_container={y_mapping.get('is_container')})."
        )
        if y_mapping.get("inferred_value_filter"):
            parts.append(
                f"Inferred Y value_filter={y_mapping['inferred_value_filter']!r}; "
                "v7 placeholder substrate -> match_count expected 0."
            )
        parts.append(
            f"Z mapping: {z_summary['n_mapped']} mapped / {z_summary['n_fuzzy']} fuzzy / "
            f"{z_summary['n_unmapped']} unmapped (of {z_summary['n_total']} total)."
        )
        if pid_overlap is not None:
            parts.append(f"pid_overlap_count={pid_overlap}.")
        if eq_type and estimator:
            parts.append(
                f"Inferred equation_type={eq_type}, candidate_estimator={estimator}."
            )
        if blockers:
            parts.append(f"Blockers: {', '.join(blockers)}.")
        if warnings:
            parts.append(f"Warnings: {', '.join(warnings)}.")
        return " ".join(parts)


__all__ = [
    "FeasibilityJudge",
    "FeasibilityReport",
]