"""Load extraction-stage table JSON using existing WaterScope media resolver.

Not a validated cell grid when structure_extracted=False. Preserve PDF word
coordinates, original strings and extraction quality flags; never silently
interpret position-only data as exact, verified spreadsheet cells.
"""
from __future__ import annotations
import json
from pathlib import Path
from typing import Callable, Any


def load_table_evidence(folder: str, label: str, resolver: Callable[[str,str], Any]) -> dict | None:
    manifest_path = resolver(folder, "manifest.json")
    if manifest_path is None:
        return None
    manifest = json.loads(Path(manifest_path).read_text(encoding="utf-8-sig"))
    matches = [x for x in manifest.get("items", []) if x.get("label") == label and x.get("kind") == "table"]
    if len(matches) != 1:
        return None
    filename = matches[0].get("data_file")
    if not filename or Path(filename).name != filename or not filename.lower().endswith(".json"):
        return None
    path = resolver(folder, filename)
    if path is None:
        return None
    raw = Path(path).read_bytes()
    if len(raw) > 3_000_000:
        raise ValueError("Table JSON exceeds 3MB limit")
    table = json.loads(raw.decode("utf-8-sig"))
    if table.get("kind") != "table" or table.get("label") != label:
        raise ValueError("Table JSON does not match manifest label")
    rows = table.get("text_rows") or []
    cells = table.get("cells")
    if not isinstance(rows, list) or len(rows) > 500:
        raise ValueError("Invalid/oversized table text rows")
    # Preserve every source row, but replace redundant per-word bounding boxes
    # with a compact word@horizontal-position representation. A text row is NOT
    # a verified table cell row, especially for merged/multi-line groups.
    compact_rows = []
    for row in rows:
        words = []
        for word in (row.get("words") or [])[:100]:
            bbox = word.get("bbox_points") or []
            token = str(word.get("text", ""))[:120]
            if len(bbox) == 4:
                words.append([token, round(float(bbox[0]), 1)])
            else:
                words.append([token, None])
        compact_rows.append({"text": str(row.get("text", ""))[:1600], "positioned_words": words})
    # Include all nonempty rows, including continuation rows and classifications.
    # The original JSON remains in the dataset; this is only an inference view.
    return {
        "label": label, "data_file": filename,
        "structure_extracted": bool(table.get("structure_extracted")),
        "row_count": table.get("row_count"), "column_count": table.get("column_count"),
        "cells": cells if table.get("structure_extracted") and isinstance(cells, list) else None,
        "text_rows": compact_rows,
        "source_quality": "structured_cells" if table.get("structure_extracted") and isinstance(cells,list)
                          else "positioned_text_unverified",
    }