tests/test_intake.py

"""WP-11 real-scan intake tests (Fixes A–D).

Every assert compares two independently computed quantities.
Teeth constructions prove the naive/old path fails the bar.
"""

from __future__ import annotations

import io
import warnings

import numpy as np
import pytest
from PIL import Image, ImageOps

from app.preprocessing import (
    INTAKE_MAX_MEGAPIXELS,
    _detect_negative_inversion,
    _to_float_rgb,
    preprocess_negative,
    trim_uniform_border,
)


def _solid_pil(val: float, size: int = 32) -> Image.Image:
    """Solid gray RGB PIL image with luminance ≈ val (sRGB display value)."""
    arr = np.full((size, size, 3), int(np.clip(val, 0, 1) * 255), dtype=np.uint8)
    return Image.fromarray(arr, mode="RGB")


# ---------------------------------------------------------------------------
# Fix A — loader (16-bit + EXIF + size guard)
# ---------------------------------------------------------------------------

def test_i16_preserves_precision_beyond_8bit_teeth():
    """I;16 mid-gray levels that collapse under convert('RGB')/255 stay distinct.

    Independent: high-bit loader values vs 8-bit-truncated naive values.
    Teeth: old convert("RGB") path fails to distinguish the levels.
    """
    # Two mid-gray levels that map to the same 8-bit code after /255 truncation
    # 32768 and 32895 differ by 127 in 16-bit; after //256 both become 128 → same 8-bit
    v1, v2 = 32768, 32895
    assert (v1 // 256) == (v2 // 256), "setup: levels must collapse under 8-bit"

    arr1 = np.full((8, 8), v1, dtype=np.uint16)
    arr2 = np.full((8, 8), v2, dtype=np.uint16)
    img1 = Image.fromarray(arr1, mode="I;16")
    img2 = Image.fromarray(arr2, mode="I;16")

    # Independent high-bit load
    hi1 = float(_to_float_rgb(img1).mean())
    hi2 = float(_to_float_rgb(img2).mean())

    # Independent naive 8-bit path (the old convert("RGB")/255)
    naive1 = float(np.asarray(img1.convert("RGB"), dtype=np.float32).mean() / 255.0)
    naive2 = float(np.asarray(img2.convert("RGB"), dtype=np.float32).mean() / 255.0)

    # Teeth: naive collapses
    assert abs(naive1 - naive2) < 1e-6, "teeth: old convert('RGB') must collapse levels"
    # High-bit path preserves distinction
    assert abs(hi1 - hi2) > 1e-4, f"16-bit path must distinguish levels (Δ={abs(hi1-hi2):.6f})"
    # Sanity: values near 0.5
    assert 0.4 < hi1 < 0.6 and 0.4 < hi2 < 0.6


def test_exif_orientation_transposes_teeth():
    """EXIF orientation tag rotates content; without transpose, layout differs.

    Independent: content after exif_transpose vs raw array layout.
    Teeth: skipping exif_transpose yields a different spatial arrangement.
    """
    # Build a 2×4 image with unique pixel sequence so orientation is observable
    # Row-major: [[0,1,2,3],[4,5,6,7]] → after rotate-90-CW via EXIF tag 6 → 4×2
    raw = np.arange(8, dtype=np.uint8).reshape(2, 4)
    rgb = np.stack([raw, raw, raw], axis=-1)
    img = Image.fromarray(rgb, mode="RGB")

    # Attach EXIF Orientation=6 (rotate 90 CW)
    # Minimal EXIF with Orientation tag
    exif = img.getexif()
    exif[274] = 6  # Orientation
    buf = io.BytesIO()
    img.save(buf, format="JPEG", exif=exif.tobytes(), quality=95)
    buf.seek(0)
    loaded = Image.open(buf)

    # Independent: apply transpose (as preprocess does) vs not
    transposed = ImageOps.exif_transpose(loaded)
    arr_with = np.asarray(transposed.convert("RGB"))
    arr_without = np.asarray(loaded.convert("RGB"))  # no transpose

    # Teeth: without transpose shape/content differs from transposed
    assert arr_with.shape != arr_without.shape or not np.array_equal(arr_with, arr_without), (
        "teeth: without exif_transpose the image must differ"
    )
    # With transpose: orientation applied → 4×2
    assert arr_with.shape[0] == 4 and arr_with.shape[1] == 2, (
        f"expected 4×2 after orientation-6, got {arr_with.shape}"
    )


def test_size_guard_downscales_over_50mp():
    """Uploads beyond INTAKE_MAX_MEGAPIXELS are downscaled (with warning)."""
    # Simulate a large logical size without allocating 50MP of pixels:
    # monkey the size check via a real modest image whose .size we override
    # by resizing a small image UP then relying on the guard.
    # Allocate ~2.5 MP and temporarily lower the constant via a unit that
    # still exercises the path: create 3000×3000 (~9 MP) and lower? 
    # Spec: beyond 50 MP. Creating 50MP is heavy (~150 MB RGB). Use a
    # synthetic path that triggers _guard_intake_size with a temporarily
    # patched constant.
    from app import preprocessing as prep

    small = Image.fromarray(np.zeros((100, 100, 3), dtype=np.uint8), mode="RGB")
    # Temporarily set a tiny cap so 100×100 exceeds it
    old = prep.INTAKE_MAX_MEGAPIXELS
    try:
        prep.INTAKE_MAX_MEGAPIXELS = 0.005  # 0.005 MP = 5000 px; 100×100=10000 > cap
        with warnings.catch_warnings(record=True) as w:
            warnings.simplefilter("always")
            out = prep._guard_intake_size(small)
        assert out.size[0] * out.size[1] < 100 * 100
        assert any("intake cap" in str(x.message) for x in w)
    finally:
        prep.INTAKE_MAX_MEGAPIXELS = old


def test_8bit_path_byte_identical_to_legacy():
    """Ordinary RGB load must match convert('RGB')/255 exactly (synthetic path)."""
    rng = np.random.default_rng(0)
    arr = rng.integers(0, 256, (32, 32, 3), dtype=np.uint8)
    img = Image.fromarray(arr, mode="RGB")
    legacy = np.asarray(img.convert("RGB"), dtype=np.float32) / 255.0
    modern = _to_float_rgb(img)
    assert np.array_equal(legacy, modern)


# ---------------------------------------------------------------------------
# Fix B — inversion detection + scan_type override
# ---------------------------------------------------------------------------

def test_two_stat_heuristic_bright_negative():
    """Synthetic bright negative (high mean AND median) → detected as negative."""
    # Independent construction: solid bright field
    lum = np.full((64, 64), 0.70, dtype=np.float32)
    assert float(lum.mean()) > 0.55 and float(np.median(lum)) > 0.50
    assert _detect_negative_inversion(lum) is True


def test_two_stat_heuristic_normal_positive():
    """Normal-key positive (mid tones) → not inverted."""
    lum = np.full((64, 64), 0.40, dtype=np.float32)
    assert float(lum.mean()) <= 0.55 or float(np.median(lum)) <= 0.50
    assert _detect_negative_inversion(lum) is False


def test_two_stat_rejects_bright_outlier_only():
    """Bright mean from outliers but median dim → not negative (two-stat)."""
    lum = np.full((64, 64), 0.30, dtype=np.float32)
    # Spike enough pixels so mean > 0.55 while median stays at 0.30
    # 40% of pixels at 1.0: mean = 0.3*0.6 + 1.0*0.4 = 0.58
    lum[:26, :] = 1.0
    mean_l, med_l = float(lum.mean()), float(np.median(lum))
    assert mean_l > 0.55 and med_l <= 0.50, f"setup mean={mean_l:.3f} med={med_l:.3f}"
    # Independent: old single-stat would say True; new rule says False
    old_single = mean_l > 0.55
    new_two = _detect_negative_inversion(lum)
    assert old_single is True
    assert new_two is False


def test_scan_type_override_wins_over_heuristic():
    """scan_type='positive'/'negative' force outcome regardless of luminance.

    Two independent constructions: a bright field (heuristic → negative) forced
    positive, and a dark field (heuristic → positive) forced negative.
    """
    bright = _solid_pil(0.80)  # heuristic would invert
    dark = _solid_pil(0.30)    # heuristic would not invert

    # Independent: auto outcomes
    auto_bright = preprocess_negative(bright, scan_type="auto")
    auto_dark = preprocess_negative(dark, scan_type="auto")
    assert auto_bright.was_inverted is True
    assert auto_dark.was_inverted is False

    # Override wins
    force_pos = preprocess_negative(bright, scan_type="positive")
    force_neg = preprocess_negative(dark, scan_type="negative")
    assert force_pos.was_inverted is False, "override positive must win over bright heuristic"
    assert force_neg.was_inverted is True, "override negative must win over dark heuristic"


# ---------------------------------------------------------------------------
# Fix C — auto_exposed densitometry anchor (load-bearing)
# ---------------------------------------------------------------------------

def test_auto_exposed_anchor_beats_linear_on_brightness_shift():
    """On a brightness-shifted scan, auto_exposed density is closer to un-shifted truth.

    Truth = density of the ORIGINAL un-shifted fixture (independent densitometry call).
    Linear anchor on shifted scan is measurably wrong (teeth).
    Auto_exposed on shifted is closer to truth than linear on shifted.
    """
    from pathlib import Path
    from densitometry import scan_to_density
    from film_physics import get_film_curve

    fix = sorted(Path("synth/fixtures").glob("case_*.npz"))[0]
    data = np.load(fix)
    scan = data["scan"].astype(np.float32)  # un-shifted truth source
    stock = str(data["stock"]) if "stock" in data else "Generic"
    curve = get_film_curve(stock)
    d_min = float(curve.d_min.item())

    # Independent truth density on the un-shifted scan (linear path = synthetic truth)
    d_truth, _ = scan_to_density(scan, stock=stock)

    # Simulate auto-exposure shifting the white point (global brightness scale)
    k = 0.55
    shifted = np.clip(scan * k, 0.0, 1.0).astype(np.float32)

    # Independent: linear anchor on shifted
    d_linear, _ = scan_to_density(shifted, stock=stock)
    # Independent: auto_exposed anchor on shifted
    d_auto, _ = scan_to_density(shifted, stock=stock, d_min_override=d_min)

    # Median absolute error vs truth (VALID-ish: use all finite pixels)
    err_linear = float(np.median(np.abs(d_linear - d_truth)))
    err_auto = float(np.median(np.abs(d_auto - d_truth)))

    # Teeth: linear anchor is measurably wrong on the shifted scan
    assert err_linear > 0.05, (
        f"teeth: linear anchor must be wrong on brightness-shifted scan "
        f"(err={err_linear:.4f})"
    )
    # Auto_exposed closer to un-shifted truth
    assert err_auto < err_linear, (
        f"auto_exposed err {err_auto:.4f} must beat linear err {err_linear:.4f}"
    )


def test_linear_calibration_default_unchanged_on_fixture():
    """Default scan_calibration='linear' keeps densitometry path on fixtures."""
    from pathlib import Path

    fix = sorted(Path("synth/fixtures").glob("case_*.npz"))[0]
    data = np.load(fix)
    scan = (data["scan"] * 255).clip(0, 255).astype(np.uint8)
    pil = Image.fromarray(scan)
    # Default path (linear) must populate density
    pre = preprocess_negative(pil, stock="Generic", scan_calibration="linear")
    assert pre.density is not None
    # auto_exposed also populates (different anchor)
    pre_ae = preprocess_negative(pil, stock="Generic", scan_calibration="auto_exposed")
    assert pre_ae.density is not None


# ---------------------------------------------------------------------------
# Fix D — uniform-border trim (opt-in, default OFF)
# ---------------------------------------------------------------------------

def test_trim_uniform_border_removes_black_border():
    """Image with black uniform border → trimmed to content bbox."""
    # Content 20×20 gray at 0.5, surrounded by 10 px black border → 40×40 total
    canvas = np.zeros((40, 40, 3), dtype=np.float32)
    canvas[10:30, 10:30] = 0.5
    cropped, bbox = trim_uniform_border(canvas, tol=1e-3, max_frac=0.25)
    top, bottom, left, right = bbox
    # Independent expected content region
    assert top == 10 and left == 10 and bottom == 30 and right == 30, (
        f"expected content bbox (10,30,10,30), got {bbox}"
    )
    assert cropped.shape[:2] == (20, 20)
    assert float(cropped.mean()) > 0.4


def test_trim_uniform_border_identity_no_border():
    """No uniform border → returned unchanged (identity)."""
    rng = np.random.default_rng(7)
    content = rng.uniform(0.2, 0.8, (32, 32, 3)).astype(np.float32)
    cropped, bbox = trim_uniform_border(content, tol=1e-3, max_frac=0.25)
    # Independent: shape matches original; pixel values equal
    assert cropped.shape == content.shape
    assert bbox == (0, 32, 0, 32)
    assert np.allclose(cropped, content)


def test_trim_max_frac_cap_never_empty():
    """All-uniform pathological input: max_frac cap respected, never empty."""
    solid = np.zeros((40, 40, 3), dtype=np.float32)
    cropped, bbox = trim_uniform_border(solid, tol=1e-3, max_frac=0.25)
    top, bottom, left, right = bbox
    # Independent: remaining region non-empty; each edge trim ≤ max_frac
    assert cropped.size > 0
    assert bottom > top and right > left
    assert top <= int(40 * 0.25)
    assert (40 - bottom) <= int(40 * 0.25)
    assert left <= int(40 * 0.25)
    assert (40 - right) <= int(40 * 0.25)


def test_auto_trim_default_off_preserves_size():
    """auto_trim=False (default) does not crop a bordered image."""
    canvas = np.zeros((40, 40, 3), dtype=np.uint8)
    canvas[10:30, 10:30] = 128
    pil = Image.fromarray(canvas, mode="RGB")
    pre_off = preprocess_negative(pil, auto_trim=False)
    pre_on = preprocess_negative(pil, auto_trim=True)
    # Independent: off keeps full size; on is smaller
    assert pre_off.rgb.shape[0] == 40 and pre_off.rgb.shape[1] == 40
    assert pre_on.rgb.shape[0] < 40 or pre_on.rgb.shape[1] < 40
    # trim_bbox_frac contract: None when off, fractions of pre-trim size when on
    assert pre_off.trim_bbox_frac is None
    tf, bf, lf, rf = pre_on.trim_bbox_frac
    # Independent expected fractions from the known 10-px border on 40 px
    assert abs(tf - 10 / 40) < 1e-6 and abs(bf - 30 / 40) < 1e-6
    assert abs(lf - 10 / 40) < 1e-6 and abs(rf - 30 / 40) < 1e-6


# ---------------------------------------------------------------------------
# WP-11 post-review — full-res export must match the intake pipeline geometry
# ---------------------------------------------------------------------------

def _fixture_scan_uint8() -> np.ndarray:
    from pathlib import Path

    fix = sorted(Path("synth/fixtures").glob("case_*.npz"))[0]
    scan = np.load(fix)["scan"].astype(np.float32)
    return (np.clip(scan, 0.0, 1.0) * 255).astype(np.uint8)


def test_fullres_export_follows_exif_orientation_teeth():
    """Full-res A/B must have the EXIF-transposed geometry, not the raw one.

    The working images come from the exif_transpose'd pipeline (Fix A); the
    full-res guide is built from the ORIGINAL upload. Independent expectation:
    a (48h, 64w) raw with Orientation=6 displays as (64h, 48w), so full-res
    output must be 48×64 (PIL w×h). Teeth: pre-fix the raw orientation leaked
    through and the output was 64×48.
    """
    from app.main import process_negative

    scan8 = _fixture_scan_uint8()[:48, :64]  # non-square: h=48, w=64
    img = Image.fromarray(scan8, mode="RGB")
    exif = img.getexif()
    exif[274] = 6  # Orientation: rotate 90 CW on display
    buf = io.BytesIO()
    img.save(buf, format="JPEG", exif=exif.tobytes(), quality=95)
    buf.seek(0)
    upload = Image.open(buf)

    out = process_negative(upload, "Generic", 1.0, 0.5, 3, False, True)
    status, full_a = out[4], out[9]
    assert "Full-res export failed" not in status
    assert full_a is not None
    # Independent: expected dims from raw dims + orientation-6 transpose rule
    assert full_a.size == (48, 64), (
        f"full-res must follow EXIF-transposed geometry (48w,64h), got {full_a.size}"
    )


def test_fullres_export_uses_trimmed_region_teeth():
    """With auto_trim on, full-res must span the trimmed content, not the border.

    Independent expectation: 64×64 content + 12 px black border = 88×88 upload;
    trim recovers the 64×64 content, so full-res output must be 64×64.
    Teeth: pre-fix the untrimmed original leaked through and output was 88×88.
    """
    from app.main import process_negative

    content = _fixture_scan_uint8()  # 64×64
    canvas = np.zeros((88, 88, 3), dtype=np.uint8)
    canvas[12:76, 12:76] = content
    upload = Image.fromarray(canvas, mode="RGB")

    out = process_negative(
        upload, "Generic", 1.0, 0.5, 3, False, True, auto_trim=True
    )
    status, full_a = out[4], out[9]
    assert "Full-res export failed" not in status
    assert full_a is not None
    assert full_a.size == (64, 64), (
        f"full-res must span the trimmed 64×64 content, got {full_a.size}"
    )