"""W18D bg→fp region ref slice — minimal safety tests only. 3 tests: 1. happy path join shape + coverage on synthetic W18A2 + W18C inputs. 2. missing bg in W18A2 overlay OR base_marker not in W18C read causes invariant fail. 3. methodology grep + production/api zero guard. """ from __future__ import annotations import json import re import sys from pathlib import Path _REPO_ROOT = Path(__file__).resolve().parents[3] _SCRIPTS_DIR = _REPO_ROOT / "backend" / "scripts" if str(_SCRIPTS_DIR) not in sys.path: sys.path.insert(0, str(_SCRIPTS_DIR)) def _synthetic_w18a_fp_entry() -> dict: return { "base_fp_t2i_prompt_text": "schematic plan.", "included_marker_legend": [ {"marker_number": 1, "source_candidate_number": 1, "base_layer_decision": "base_structural_unit", "label": "u1", "unit_id": "U1", "visual_encoding": "filled_area", "must_be_legible": True}, {"marker_number": 2, "source_candidate_number": 2, "base_layer_decision": "base_structural_unit", "label": "u2", "unit_id": "U2", "visual_encoding": "filled_area", "must_be_legible": True}, ], "excluded_transient_elements": [ {"marker_number": 14, "source_candidate_number": 14, "base_layer_decision": "state_overlay_transient_object", "label": "Tx", "unit_id": "U1", "excluded_reason": "scene state", "overlay_instruction_hint": "describe in bg body"}, ], "bg_state_overlay_payload_by_bg": { "BGa": { "bg_id": "BGa", "fp_id": "fp_l05_01", "target_unit_ids": ["U1"], "base_markers_to_reference": [1], "transient_markers_to_describe": [14], "prompt_appendix_hint": "anchor on living", }, "BGb": { "bg_id": "BGb", "fp_id": "fp_l05_01", "target_unit_ids": ["U2"], "base_markers_to_reference": [2], "transient_markers_to_describe": [], "prompt_appendix_hint": "clean view", }, }, "base_fp_contract_notes": "compact zones", "production_prompt_delta_recommendations": "review", } def _synthetic_w18c_readback() -> dict: return { "read_markers": [ {"marker_number": 1, "visible": True, "confidence": "high", "observed_label_or_area": "zone1", "approximate_10x10_rect": [0, 0, 4, 4], "notes": ""}, {"marker_number": 2, "visible": True, "confidence": "high", "observed_label_or_area": "zone2", "approximate_10x10_rect": [5, 0, 9, 4], "notes": ""}, ], "missing_or_ambiguous_markers": [], "base_unit_layout": [ {"marker_number": 1, "unit_id": "U1", "observed_area_label": "u1", "approximate_10x10_rect": [0, 0, 4, 4], "adjacency_notes": "left"}, {"marker_number": 2, "unit_id": "U2", "observed_area_label": "u2", "approximate_10x10_rect": [5, 0, 9, 4], "adjacency_notes": "right"}, ], "structural_relationship_summary": {}, "unexpected_transient_markers": [], "readback_conflicts": [], } def _build_region_ref_via_helper(w18a_fp, w18c_readback): from experiment_floor_plan_bg_to_fp_region_ref_slice import ( _build_region_ref_for_fp, ) return _build_region_ref_for_fp( fp_id="fp_l05_01", w18a_fp_entry=w18a_fp, w18c_readback=w18c_readback, w18b_png_relative_to_repo="scripts_output/fake/png/fp_l05_01.png", ) def _build_report(*, region_ref_by_bg, w18a_fp_entry, w18c_readback, target_fp_ids={"fp_l05_01"}, production_diff_empty=True, image_import_seen=False, image_api_call_count=0, vlm_api_call_count=0, llm_api_call_count=0): from experiment_floor_plan_bg_to_fp_region_ref_slice import ( _build_w18d_compatibility_report, ) return _build_w18d_compatibility_report( region_ref_by_bg=region_ref_by_bg, w18a_fp_entry=w18a_fp_entry, w18c_readback=w18c_readback, target_fp_ids=target_fp_ids, production_diff_empty=production_diff_empty, db_write_count=0, image_import_seen=image_import_seen, image_api_call_count=image_api_call_count, vlm_api_call_count=vlm_api_call_count, llm_api_call_count=llm_api_call_count, missing_inputs=[], prev_run_id="fakeW18C", ) def test_w18d_happy_path_join_shape_and_coverage(): """Synthetic W18A2 + W18C → W18D join produces per-bg payloads with target_unit_refs / base_marker_refs / transient_overlay_to_describe populated, and all 5 invariants pass.""" w18a = _synthetic_w18a_fp_entry() w18c = _synthetic_w18c_readback() region_ref, diags = _build_region_ref_via_helper(w18a, w18c) assert set(region_ref.keys()) == {"BGa", "BGb"} bga = region_ref["BGa"] assert bga["base_fp_png_path_relative_to_repo"].endswith( "/png/fp_l05_01.png" ) assert any( r["unit_id"] == "U1" and r["marker_number"] == 1 for r in bga["target_unit_refs"] ) assert bga["base_marker_refs"][0]["marker_number"] == 1 assert bga["base_marker_refs"][0]["visible"] is True assert bga["base_marker_refs"][0]["confidence"] == "high" assert ( bga["transient_overlay_to_describe"][0]["marker_number"] == 14 ) assert "#1" in bga["region_prompt_hint"] assert "#14" in bga["region_prompt_hint"] # BGb has no transient marker — region hint must say (none) for it. bgb = region_ref["BGb"] assert bgb["transient_overlay_to_describe"] == [] assert "(none)" in bgb["region_prompt_hint"] rep = _build_report( region_ref_by_bg=region_ref, w18a_fp_entry=w18a, w18c_readback=w18c, ) assert rep["all_pass"] is True def test_w18d_missing_bg_or_unresolved_base_marker_fails_invariants(): """If W18A2 overlay carries a bg the W18D output is missing → invariant 3 fails. If a base_marker is not in W18C read_markers → invariant 4 fails.""" w18a = _synthetic_w18a_fp_entry() w18c = _synthetic_w18c_readback() region_ref, _ = _build_region_ref_via_helper(w18a, w18c) # (a) drop BGb from W18D output → cover invariant fails. missing_bg = json.loads(json.dumps(region_ref)) del missing_bg["BGb"] rep_miss = _build_report( region_ref_by_bg=missing_bg, w18a_fp_entry=w18a, w18c_readback=w18c, ) assert rep_miss["invariants"][ "bg_region_refs_cover_w18a2_overlay_bgs" ]["pass"] is False # (b) drop marker #1 from W18C read_markers → resolve invariant fails. w18c_bad = json.loads(json.dumps(w18c)) w18c_bad["read_markers"] = [ e for e in w18c_bad["read_markers"] if e["marker_number"] != 1 ] region_ref_bad, _ = _build_region_ref_via_helper(w18a, w18c_bad) rep_resolve = _build_report( region_ref_by_bg=region_ref_bad, w18a_fp_entry=w18a, w18c_readback=w18c_bad, ) assert rep_resolve["invariants"][ "referenced_base_markers_resolve_in_w18c" ]["pass"] is False # (c) base marker visible=False → resolve invariant fails. w18c_invis = json.loads(json.dumps(w18c)) w18c_invis["read_markers"][0]["visible"] = False region_ref_invis, _ = _build_region_ref_via_helper(w18a, w18c_invis) rep_invis = _build_report( region_ref_by_bg=region_ref_invis, w18a_fp_entry=w18a, w18c_readback=w18c_invis, ) assert rep_invis["invariants"][ "referenced_base_markers_resolve_in_w18c" ]["pass"] is False def test_w18d_production_api_zero_guard_and_methodology_grep(): """All API call counts must be zero; production diff empty; image imports not seen. Plus script source must not embed scenario-specific tokens.""" w18a = _synthetic_w18a_fp_entry() w18c = _synthetic_w18c_readback() region_ref, _ = _build_region_ref_via_helper(w18a, w18c) rep_ok = _build_report( region_ref_by_bg=region_ref, w18a_fp_entry=w18a, w18c_readback=w18c, ) assert rep_ok["invariants"][ "production_diff_zero_db_write_zero_llm_vlm_image_zero" ]["pass"] is True rep_llm = _build_report( region_ref_by_bg=region_ref, w18a_fp_entry=w18a, w18c_readback=w18c, llm_api_call_count=1, ) assert rep_llm["invariants"][ "production_diff_zero_db_write_zero_llm_vlm_image_zero" ]["pass"] is False rep_img = _build_report( region_ref_by_bg=region_ref, w18a_fp_entry=w18a, w18c_readback=w18c, image_api_call_count=1, ) assert rep_img["invariants"][ "production_diff_zero_db_write_zero_llm_vlm_image_zero" ]["pass"] is False rep_vlm = _build_report( region_ref_by_bg=region_ref, w18a_fp_entry=w18a, w18c_readback=w18c, vlm_api_call_count=1, ) assert rep_vlm["invariants"][ "production_diff_zero_db_write_zero_llm_vlm_image_zero" ]["pass"] is False rep_prod = _build_report( region_ref_by_bg=region_ref, w18a_fp_entry=w18a, w18c_readback=w18c, production_diff_empty=False, ) assert rep_prod["invariants"][ "production_diff_zero_db_write_zero_llm_vlm_image_zero" ]["pass"] is False script_path = ( _SCRIPTS_DIR / "experiment_floor_plan_bg_to_fp_region_ref_slice.py" ) assert script_path.exists(), f"script missing: {script_path}" forbidden_tokens = [ "b" + "edroom", "ki" + "tchen", "blood" + "stain", "cur" + "tain", "coo" + "ktop", "tele" + "vision", "cri" + "me", "vi" + "lla", "roo" + "ftop", "foot" + "print", "pol" + "ice", "de" + "ck", "wheel" + "house", "ba" + "throom", "su" + "ri-young", ] pat = re.compile(r"(?i)\b(" + "|".join(forbidden_tokens) + r")\b") m = pat.search(script_path.read_text()) assert m is None, ( f"{script_path.name}: scenario-specific token leaked → " f"{m.group(0) if m else ''}" )