Pure streaming filter that removes ~~~tool_call ... ~~~ and ~~~tool_result ... ~~~ fenced blocks from chunked assistant text.

Non-native-tool-call providers emit tool calls as text fences (the ExAthena.ToolCalls.TextTagged protocol, ~~~tool_call\s*\n ... \n~~~). Some models additionally roleplay the result side, echoing ~~~tool_result fences (often with invented tool output) into their visible text. When this output streams to hosts as {:content, chunk} deltas, the raw fence markup leaks into the transcript — the loop already surfaces real tool activity as {:tool_call, ...} / {:tool_result, ...} events, so both fence kinds must be stripped from the text stream.

Chunks may split a marker at any byte boundary, so the filter holds the longest chunk tail that could still become a marker as pending text:

• Outside a fence, a tail like "~~", "~~~tool_" (shared by both markers), "~~~tool_c" (only ~~~tool_call), or "~~~tool_r" (only ~~~tool_result) — any strict prefix of either open marker — is held instead of emitted. A complete ~~~tool_call / ~~~tool_result is also held while only optional whitespace follows — the open marker is <marker>\s*\n (mirroring the TextTagged regex), so until the newline arrives we can't tell a fence from literal text like ~~~tool_caller. If a non-whitespace char arrives first, the held text is released as visible.
• Inside a fence, everything is swallowed until the close marker \n~~~; a tail like "\n~" is held the same way.

flush/1 ends the stream: pending text held outside a fence turned out to be real text and is returned; an unterminated fence is swallowed entirely (half a fence is still tool markup, and showing it is worse).