Ordered-list numbering in user messages is silently rewritten before the model sees it

What happens: When I type a numbered reply like

3. See it
4. Works
5. Confirmed

the assistant receives it as 1. / 2. / 3. instead. The actual digits I typed are lost. The model has no way to know I was answering items 3, 4, 5 specifically — it sees a list starting from 1 and infers I'm answering the first three items.

Repro:

1. Ask the assistant a question whose answer involves a numbered list of more than ~3 items.
2. As the user, reply with a subset of those items using their original numbers (e.g., 3. ..., 4. ...).
3. Observe the model's next turn — it treats your reply as if you'd answered items 1, 2, ...

Expected: The user's literal numbering is preserved end-to-end. Markdown rendering can renumber for visual display in the UI if desired, but the content delivered to the model should match what the user typed.

Why it matters: This corrupts the semantic content of the user message. In a long checklist (e.g., a test plan, an audit, a code-review pass), the user routinely answers a subset of items by their original numbers. Silent renumbering causes the assistant to misread completion state and make wrong follow-up decisions.

Severity: Low individually (the model can sometimes recover from context) but high in aggregate, because the assistant has no way to detect that it happened and the user has no way to know their input was modified.