It's a very simple format on a surface level. The file starts with a 12-byte header:

struct DxvkStateCacheHeader {
  char     magic[4]   = { 'D', 'X', 'V', 'K' };
  uint32_t version    = 10; /* incremented any time something changes */
  uint32_t entrySize  = 0; /* no longer meaningful, always 0 */
};

Followed by an array of entries until the end of the file, so the design allows us to just append data to the file when compiling new pipelines without rewriting the whole thing.

Each entry has the following layout, note that the header contains a bit field:

struct {
  uint32_t stageMask : 8; /* VkShaderStageFlags */
  uint32_t entrySize : 24; /* size of the data array, in bytes, not including header */
  uint8_t sha1Hash[20]; /* data hash, not including header */
  char data[];
};

The data array contains a compact serialization of the pipeline state data, but I'd highly discourage trying to parse that, especially since the exact contents may change with future releases. If your intention is to merge caches, just use the SHA1 hash for deduplication purposes.

We don't really do a huge amount of validation at compile time.

I don't think it's the file that breaks, it's more likely just that we're trying to compile a pipeline that's not supported on the given driver. Which, yes, is a problem that needs to be addressed.