Function sanitize_audio_markers 

pub fn sanitize_audio_markers(
    messages: &[ChatMessage],
) -> Cow<'_, [ChatMessage]>

Strip loadable audio markers (see strip_unplayable_audio_markers) across every message in messages, logging one degradation warning when any are removed. Returns the input borrowed when no candidate marker is present (the common, allocation-free path) or an owned rebuilt vector otherwise.

Assistant messages in native tool-call history are rewritten field-wise — only their content string, through rewrite_model_visible_text — so signed reasoning inside the envelope survives the seam.

This is the shared seam keeping a raw audio path out of provider payloads, whichever route the history takes:

• the main iteration prep (prepare_messages_for_provider, via prepare_messages_inner), and
• one-shot queries that dispatch history directly without full prep (the max-iteration graceful summary and the other run_model_query callers).

Non-audio media markers pass through untouched; see AUDIO_MARKER_KINDS for why the split falls where it does.

A DECLARED carrier never has its text regex-stripped: its audio attachments drop from the declared list (the count the strict parser checks stays consistent, so a mixed declaration cannot degrade to legacy and lose its images), and its body is verbatim. A LEGACY carrier is cloned unchanged — its body is text and no seam rewrites it. Ordinary messages keep the regex strip.