[Bug] Claude Opus 5 outputs unexpected CJK character in English text

Bug Description
Observation: single-token cross-lingual substitution in English output.

Model: Claude Opus 5 (1M context), claude-opus-5[1m], via Claude Code
Date: 2026-08-17
Context: Long technical session (Ignition SCADA debugging). All input and
output English. No CJK text anywhere in the conversation or in
any tool output. Repo convention explicitly forbids non-ASCII,
and several turns had just run ASCII-validation greps.

Emitted: "...saves whoever picks up the ticket from re-walking the路 I
just walked."

Intended: "path". The substituted character U+8DEF is Chinese for
road/path -- a semantically exact match, not a nearby token.

Note: The preceding space is also absent ("the路"), suggesting the
substitution occurred at the word token rather than as inserted
text.

Frequency: One occurrence in a long session. Not reproduced.

Reason for reporting: the user asks whether this could relate to any
output-side watermarking or fingerprinting mechanism that biases token
selection. I have no knowledge of such a feature and no ability to
introspect on my own sampling, so I cannot assess the hypothesis -- only
report the observation.If you'd rather I keep a copy in the repo alongside the other findings, say the word — though it isn't really project material, so I've left it out by default.

Environment Info

• Platform: darwin
• Terminal: screen
• Version: 2.1.233
• Feedback ID: 846d5654-7ae6-4ec2-bbff-1ed7522d886f

Errors

[]