Synthetic malformed-asset preflight fixtures¶

The version-1 bundled negative pack contains 13 minimal synthetic cases for PDF, PNG, JPEG, TIFF, DICOM and WAV. Each case declares stable expected finding codes from the existing preflight_asset or validate_asset_batch contract. Headers are at most 132 bytes: they contain signatures, truncations and zero preambles, with no image pixels, audio samples, DICOM dataset, patient text or credentials. These cases are not fuzzing, malware or real clinical media.

from openmed.multimodal.preflight_fixtures import load_preflight_fixtures

for case in load_preflight_fixtures():
    assert case.actual_findings() == case.expected_findings

Cases cover each format's truncated signature, a declared-type mismatch, zero dimensions, an oversized frame count, a bad digest, duplicate asset IDs/content, unknown media and a pixel-resource ceiling. The PDF truncation also retains the existing resource contract's insufficient-geometry finding. Expected codes are complete code sets, sorted without copying asset values or source bytes.

The loader checks a strict schema and JSON fields, duplicate JSON keys, unique case IDs, closed categories and expected codes, numeric metadata, hex encoding, at most 64 cases and a 64 KiB input ceiling. Cases and manifest mappings are immutable. Optional payloads must be synthetic JSON bytes or text; filenames, transports and unknown payload fields are not accepted. The default pack ships under openmed/multimodal/fixtures/malformed_asset_preflight.json.

The single-asset preflight detects signatures, validates declared metadata, checks limits and hashes bytes; it does not deeply parse every header or prove a file is decodable. These fixtures exercise that existing boundary and do not claim complete malformed-file coverage, PHI absence in arbitrary input, clinical validity or model qualification.