BUG: BrailleConverter silently drops characters outside its lookup table

[feiiiiii5]

Summary

BrailleConverter silently drops any character that is not in its Braille lookup table (and not an escape character or a mapped uppercase letter). Since the converter is used to obfuscate prompts for red-teaming, the model receives a different instruction than intended without any warning.

Reproduction

import asyncio
from pyrit.converter import BrailleConverter

async def main():
    c = BrailleConverter()
    for prompt in ["What\x27s 2+2?", "a@b.com", "100% sure", "café 中文"]:
        r = await c.convert_async(prompt=prompt, input_type="text")
        print(repr(r.output_text))

asyncio.run(main())

Output on current main:

\x27⠠⠺⠓⠁⠞⠄⠎ ⠼⠃⠼⠃⠦\x27      # "+" dropped → "22?"
\x27⠁⠃⠲⠉⠕⠍\x27              # "@" dropped → "ab.com"
\x27⠼⠁⠚⠚ ⠎⠥⠗⠑\x27        # "%" dropped
\x27⠉⠁⠋ \x27                    # "é 中文" dropped → "caf "

+, @, %, <, &, accented letters and CJK characters are all silently removed.

Expected behavior

Characters outside the Braille table should pass through unchanged, matching every other converter in this codebase (e.g. AtbashConverter uses str.translate which leaves unmapped characters as-is, and ArabiziConverter/CharacterSpaceConverter keep unknown characters). The encoded prompt should preserve the original instruction verbatim apart from the Braille-translatable parts.