[MarshalX]

What is done

• Reduce model memory up to 39% by sharing __pydantic_fields_set__ between models
• Optimize string format validation
• Optimize unknown-type field decoding
• Optimize CID validation
• Optimize bytes field decoding
• Optimize firehose frame parsing
• Optimize get_or_create
• Optimize DotDict (5× faster, less memory)
• Reduce import time and memory with lazy imports
• Fix hasattr() on the models package raising ModuleNotFoundError

Results

Workload	Time	Retained memory
Firehose frame parsing	−15.5%	-
Firehose: frame + CAR + record decode	−8.1%	6.8 MB -> 4.6 MB (−32%)
Jetstream messages	−6.3%	15.0 MB -> 9.2 MB (−39%)
Feed response (100 posts)	±0%	12.5 MB -> 7.8 MB (−38%)
Followers response (100 profiles)	−2.9%	3.5 MB -> 2.2 MB (−37%)
DotDict (records without a model)	−78.6% (5× faster)	6.5 MB -> 6.0 MB

Import	Before	After
from atproto import Client	209 ms, 63 MB RSS	150 ms, 51 MB RSS
from atproto import FirehoseSubscribeReposClient	206 ms, 63 MB RSS	143 ms, 45 MB RSS
from atproto import JetstreamClient	207 ms, 63 MB RSS	171 ms, 54 MB RSS