Model emits emoji despite explicit no-emoji instruction, then "cancels" it with inline commentary instead of removing it

**Context**: Claude Code CLI session, casual chat turn (not code editing).

**What happened**: The system prompt and project instructions both said "avoid emojis unless the user explicitly asks." The model sent a reply containing an emoji anyway, then appended a parenthetical noting the mistake ("didn't mean to use that emoji – retracting it") in the same message instead of regenerating the response without the emoji. The emoji was already sent; the "retraction" text had no actual effect.

**Expected**: Either don't use the emoji in the first place (per explicit instruction), or if caught mid-generation, silently omit it rather than emitting narration about the internal correction.

**Why this matters**: This is a small but clean repro of two separate failure modes stacking – (1) an explicit formatting instruction being ignored, and (2) the model narrating its own self-correction process to the user instead of just producing the corrected output, which the system prompt also explicitly discourages.

**Reproduction**: Hard to trigger deterministically – occurred once during a multi-turn casual conversation. No fixed minimal repro steps available.

No files, tests, or deterministic entry point are identified. Start with a Claude Code CLI casual chat turn and investigate whether the reported behavior can be reproduced; done means honoring the no-emoji instruction without exposing self-correction commentary.

索引モデルが issue の本文から書いたものです。