from __future__ import annotations

import importlib.util
import json
from pathlib import Path
from types import ModuleType

import numpy as np

REPOSITORY_ROOT = Path(__file__).resolve().parents[1]
RUNNER_PATH = REPOSITORY_ROOT / "experiments/perception/run_m48t_upstream_parity_worker.py"
PROFILE_PATH = REPOSITORY_ROOT / "config/perception/m48t-upstream-parity-v1.json"


def load_runner() -> ModuleType:
    specification = importlib.util.spec_from_file_location("m48t_upstream_parity", RUNNER_PATH)
    assert specification is not None and specification.loader is not None
    module = importlib.util.module_from_spec(specification)
    specification.loader.exec_module(module)
    return module


def test_upstream_parity_profile_freezes_official_and_deployed_providers() -> None:
    profile = json.loads(PROFILE_PATH.read_text("utf-8"))

    assert profile["model"]["package_version"] == "1.9.4"
    assert profile["model"]["resolution"] == [704, 704]
    assert profile["dataset"]["image_count"] == 5000
    assert profile["providers"]["pytorch"]["confidence_prefilter"] == 0.0
    assert profile["providers"]["tensorrt"]["confidence_prefilter"] == 0.0
    assert profile["authority"]["navigation_or_safety_accepted"] is False


def test_tensorrt_decoder_uses_sparse_coco_ids_and_original_image_geometry() -> None:
    runner = load_runner()
    boxes = np.zeros((1, 300, 4), dtype=np.float16)
    logits = np.full((1, 300, 91), -20, dtype=np.float16)
    boxes[0, 4] = np.asarray((0.5, 0.5, 0.2, 0.4), dtype=np.float16)
    logits[0, 4, 1] = np.float16(4.0)  # COCO sparse id 1: person
    logits[0, 7, 12] = np.float16(5.0)  # sparse gap: must never escape

    rows = runner.decode_tensorrt_coco_rows(
        image_id=42,
        image_width=1000,
        image_height=500,
        boxes=boxes,
        logits=logits,
        maximum_detections=2,
    )

    assert len(rows) == 1
    assert rows[0]["image_id"] == 42
    assert rows[0]["category_id"] == 1
    assert np.allclose(rows[0]["bbox"], [400.0244, 149.9756, 199.9512, 200.0488], atol=0.1)


def test_parity_decision_localizes_upstream_and_deployment_failures() -> None:
    runner = load_runner()
    profile = json.loads(PROFILE_PATH.read_text("utf-8"))

    passed = runner.build_parity_decision(
        profile=profile,
        pytorch_metrics={"ap_50_95": 0.565, "ap_50": 0.751},
        tensorrt_metrics={"ap_50_95": 0.563, "ap_50": 0.749},
        full_admission_run=True,
    )
    assert passed["passed"] is True
    assert passed["diagnosis"] == "upstream-and-tensorrt-parity-passed"

    deployment_failure = runner.build_parity_decision(
        profile=profile,
        pytorch_metrics={"ap_50_95": 0.565, "ap_50": 0.751},
        tensorrt_metrics={"ap_50_95": 0.54, "ap_50": 0.72},
        full_admission_run=True,
    )
    assert deployment_failure["passed"] is False
    assert deployment_failure["diagnosis"] == "tensorrt-deployment-parity-failed"

    smoke = runner.build_parity_decision(
        profile=profile,
        pytorch_metrics={"ap_50_95": 0.565, "ap_50": 0.751},
        tensorrt_metrics={"ap_50_95": 0.565, "ap_50": 0.751},
        full_admission_run=False,
    )
    assert smoke["passed"] is False
    assert smoke["diagnosis"] == "bounded-smoke-only"