2.5.0

2.5.0

• Deprecated the legacy mmdbdata.Unmarshaler callback,
  UnmarshalMaxMindDB(*mmdbdata.Decoder) error. It remains supported throughout
  v2, but new handwritten decoders should implement
  mmdbdata.CursorUnmarshaler so nested decoding can return a proven successor
  without rescanning the value. When a type implements both interfaces, the
  cursor callback takes precedence. Removal is planned for v3. GitHub #224.
  Legacy callbacks must not retain the supplied decoder or its iterators after
  returning; decoder instances may now be pooled and reused.
• Added the optional maxminddb-gen command for reproducible generation of
  reflection-free decoders for application-owned types, together with cursor
  primitives that avoid rescanning completely consumed containers and a
  pool-free cursor unmarshaling interface whose opaque successor supports
  single-pass nested custom decoding. The command discovers exported structs in
  its input source file and writes a matching
  <source>_maxminddb.go file by default while preserving build constraints and
  recognized filename build suffixes. Generated struct decoders use lightweight
  counted map traversal and compact pointer-string fast paths. Output-path
  migrations ignore superseded generated methods while analyzing replacements,
  MaxMind tag validation remains isolated from unrelated tags, and output
  replacement requires an exact generated ownership marker.
• Fixed valid four-byte data pointers whose ignored high address bits produce
  control values 29 through 31 so they are not misread as extended value sizes.
• Fixed the string cache so overlapping string encodings that share a payload
  offset remain distinct and cannot return the wrong cached string or map key.
• Fixed nested struct fields containing a non-map value so decoding reports the
  correct type error at the field offset instead of retrying from the record
  root.
• Reduced IPv4 and IPv6 lookup time for databases with 28-bit search-tree
  records.
• Reduced allocations when recurring decoded strings share a primary cache
  slot.
• Reduced struct decoding time by using compact field-name fingerprints before
  falling back to full string hashing.
• Rejected impossible or malformed large container sizes before allocating
  destination maps and slices, while reducing preflight overhead for common
  strings and booleans and avoiding preflight when caller-provided slice
  capacity already prevents an allocation.
• Kept readers reachable through memory-mapped lookup, decode, and iteration
  operations so runtime cleanup cannot unmap active data.
• Reduced opening memory by decoding metadata without a string cache and added
  DisableStringCache for readers that favor lower memory over repeated-decode
  allocation savings.
• Released decoder-owned data and cache references when a reader is closed.
• Rejected invalid netip.Addr lookup values.
• Made verification reject invalid UTF-8 strings and made empty-value filtering
  reject pointer-to-pointer records consistently with other decoder paths.
• Corrected cold-cache and concurrent-lookup benchmarks so they measure steady
  cache misses and lookup work rather than warm caches and goroutine setup.