How to reproduce the behaviour

nlp = English()
doc = nlp("I can't believe you have done this")

"can't" is a tokenizer exception because of the funky contraction (yay english)

Spans that contain these exceptions are marked as has_special
declared here:

And has_special spans are not cached:

The problem is that has_special, once set to a nonzero value, is never reset. And this means that once the tokenizer encounters a special case, every subsequent span is also marked as special, and none of them get cached, even if they should be.

This has some fairly significant tokenizer performance implications. It should be much faster than it is. I'll put some benchmarking in my PR.

Your Environment

Info about spaCy

• spaCy version: 3.8.14
• Platform: macOS-26.3.1-arm64-arm-64bit
• Python version: 3.12.4
• Pipelines: en_core_web_md (3.8.0), en_core_web_sm (3.8.0)