Possible schema issue with enum output

I'm working with the new schema support (I just added it for llm-ollama, for context), and I am running into something I don't quite understand which may be a bug. I am using the llm-ollama plugin for all these examples.

Take this brief demo snippet:

#!/usr/bin/env python

import llm

CATEGORY_SCHEMA = {
    "type": "string",
    "pattern": "^[a-zA-Z0-9]*$",
}

model = llm.get_model("llama3.2")
response = model.prompt(
    "Is 11:30 AM in the night or the day?",
    schema=CATEGORY_SCHEMA
)
print(f"Original response: '{response.text()}'")
stripped = response.text().replace('"', '')
print(f"Stripped quotes: '{stripped}'")
print("Ordinals:")
for x in response.text():
    print(f"'{x}': {ord(x)}")

When I run it, I sometimes get an output like:

$ ~/tmp/schema-demo.py
Original response: '"Night" '
Stripped quotes: 'Night '
Ordinals:
'"': 34
'N': 78
'i': 105
'g': 103
'h': 104
't': 116
'"': 34
' ': 32

The schema is functioning, since if I exclude it I usually get a sentence-long response. However, the schema-constrained response contains quote and space characters which are disallowed by the pattern.

I am still learning about LLMs, so if this is just an expected sort of weirdness I'll accept it and add some additional validation on my end. I just wanted to report it in case it represented a real issue that was worth addressing. Thanks for your hard work on this library, it's been a joy to work with!

Fun note: if I use the LLM command-r7b, I get even more bonkers output:

$ ~/tmp/schema-demo.py
Original response: '"Night"
                                          '
Stripped quotes: 'Night
                                          '
Ordinals:
'"': 34
'N': 78
'i': 105
'g': 103
'h': 104
't': 116
'"': 34
'
': 10
'       ': 9
' ': 32
' ': 32
' ': 32
' ': 32
' ': 32
' ': 32
' ': 32
' ': 32
' ': 32
' ': 32
'       ': 9
' ': 32
' ': 32
' ': 32
' ': 32
'       ': 9
'       ': 9
' ': 32
' ': 32

Reproduce the schema-demo.py example with the llm-ollama plugin, the llama3.2 model, and command-r7b, comparing responses with and without the schema. Start by tracing the schema-support behavior involved in model.prompt(); done means establishing whether the quoted or whitespace-padded output violates the requested pattern and documenting or addressing the behavior accordingly.