"""Tests for harness/vision.py — the readability judge's pure logic + fail-open behaviour.

No moviepy / no live API: frame extraction and the vision call are injected, so we exercise
parsing, aggregation, and the all-important fail-open guarantees deterministically.
"""

import pytest

from harness import vision
from harness.vision import ReadabilityReport, NOT_ASSESSED, assess_readability


_FAKE_FRAMES = [(0.25, b"png0"), (0.5, b"png1"), (0.75, b"png2")]


def _extract_ok(_path):
    return list(_FAKE_FRAMES)


# ── JSON parsing ──────────────────────────────────────────────────────────────

def test_parse_plain_json():
    score, issues = vision._parse_score_json('{"score": 0.82, "issues": ["text too small"]}')
    assert score == 0.82 and issues == ["text too small"]


def test_parse_fenced_json_with_prose():
    raw = "Here is my verdict:\n```json\n{\"score\": 0.4, \"issues\": [\"bubble clipped\", \" \"]}\n```"
    score, issues = vision._parse_score_json(raw)
    assert score == 0.4
    assert issues == ["bubble clipped"]          # blank issue dropped


def test_parse_clamps_score():
    assert vision._parse_score_json('{"score": 1.7}')[0] == 1.0
    assert vision._parse_score_json('{"score": -3}')[0] == 0.0


# ── assess_readability: happy path + aggregation ──────────────────────────────

def test_assess_happy_path_injected():
    rep = assess_readability(
        "x.mp4",
        extract=_extract_ok,
        assess=lambda frames: (0.73, ["one issue"]),
    )
    assert rep.assessed and rep.score == 0.73
    assert rep.issues == ["one issue"]
    assert rep.frames_assessed == 3
    assert rep.error == ""


def test_assess_passed_threshold_semantics():
    rep = ReadabilityReport(score=0.5)
    assert rep.passed(0.4) is True
    assert rep.passed(0.6) is False


# ── fail-open guarantees (a vision/extraction problem must never block) ───────

def test_assess_fails_open_on_extract_error():
    def boom(_path):
        raise RuntimeError("no moviepy")

    rep = assess_readability("x.mp4", extract=boom, assess=lambda f: (0.9, []))
    assert rep.score == NOT_ASSESSED
    assert not rep.assessed
    assert "frame_extract_failed" in rep.error
    assert rep.passed(0.99) is True               # a non-assessment never fails the gate


def test_assess_fails_open_on_vision_error():
    def boom(_frames):
        raise RuntimeError("api 500")

    rep = assess_readability("x.mp4", extract=_extract_ok, assess=boom)
    assert rep.score == NOT_ASSESSED
    assert "vision_failed" in rep.error
    assert rep.frames_assessed == 3               # we still recorded that frames were extracted


def test_assess_no_frames_is_not_assessed():
    rep = assess_readability("x.mp4", extract=lambda p: [], assess=lambda f: (0.9, []))
    assert rep.score == NOT_ASSESSED and "no frames" in rep.error


def test_assess_no_api_key_with_default_assessor(monkeypatch):
    # Default (real) assessor but no key anywhere → not assessed, never raises.
    monkeypatch.delenv("GEMINI_API_KEY", raising=False)
    rep = assess_readability("x.mp4", api_key="", extract=_extract_ok)
    assert rep.score == NOT_ASSESSED
    assert "GEMINI_API_KEY" in rep.error


def test_summary_is_human_readable():
    assert "not assessed" in ReadabilityReport(error="boom").summary()
    assert "0.42" in ReadabilityReport(score=0.42, issues=["x"]).summary()