Architectural Breakdown: A logo at 1.00:1 contrast passed every check we had

The original implementation of the logo validation code was fine in theory, but failed to perform well in practice. The code suffered from several key defects.

First, the decompression routine `_decode_idat()` read the entire concatenated IDAT payload into memory before processing each row, resulting in excessive memory usage. For a 5 MB RGBA PNG, this could consume up to 40 MB of memory per request, quickly exhausting system resources under heavy load.

Second, the queue mechanism `QUEUE_MAX_SIZE` was defined but never instantiated. As a result, incoming assets were allowed to pile up in unbounded local buffers, with no back-pressure to regulate ingestion rates. This created a system with no natural limits, prone to resource exhaustion when faced with a traffic burst.

Third, the `validate()` function accepted `width` and `height` parameters that were never used. Callers could pass invalid dimensions, and the function would fall back to using the static IHDR values. While this worked by accident initially, it would break down when the validator was reused across different asset types.

Fourth, the code incorrectly assumed a value of 3 for grayscale images based on the color type. Grayscale images should use a value of 1, but the code defaulted to 3, leading to out-of-bounds access when processing grayscale scans.

Finally, the decompression process could block the event loop indefinitely if handed oversized or malformed IDAT blocks. On an 8 GB node, this could cause multiple workers to stall simultaneously, collapsing overall throughput and making the service unusable.

These defects combined to create a logo validation system that looked correct on paper but was fundamentally flawed when run in production. The fixes involved enforcing hard memory limits, properly instantiating the queue, validating dimensions, correcting the grayscale handling, and ensuring the decompression process did not block the event loop.