Fix NameError in _handle_message_with_agent on already-sent media delivery

[jorgeblanc9]

Summary

• _history_media_paths was only defined inside the unrelated run_sync() closure (~line 21981), not in _handle_message_with_agent, so any already-streamed response that also needed to deliver media raised a NameError at the _deliver_media_from_response(...) call.
• This surfaced on essentially every Telegram message as a second "Sorry, I encountered an unexpected error." reply following the normal streamed response.
• Fix computes the same set from history, the pre-turn transcript already loaded earlier in _handle_message_with_agent, mirroring how run_sync derives agent_history for the same dedup purpose (so already-delivered media isn't double-counted, and media generated in the current turn is still delivered).

Traceback fixed

File "gateway/run.py", line 14561, in _handle_message_with_agent
    history_media_paths=_history_media_paths,
NameError: name '_history_media_paths' is not defined

Test plan

• Reproduced live on a Telegram adapter instance: every message triggered the streamed reply followed by the generic error.
• Applied fix, restarted the gateway process, confirmed the traceback no longer appears in gateway.log on subsequent messages.
• Would appreciate a maintainer's eyes on whether history (vs. some other pre-turn history source) is the right dedup input in all call paths that reach this branch.

🤖 Generated with Claude Code

[alt-glitch]

This was generated by AI during triage.

Duplicate of #72321: both replace the undefined _history_media_paths in the same post-stream delivery call with history-derived paths. #72321 also adds focused coverage that preserves prior-turn MEDIA deduplication.

[teknium1]

Fixed on main via PR #72341 (salvaged from #72321 by @TheEpTic, merge commit c2ee503). Your fix was character-for-character identical — same _collect_history_media_paths(history) approach — but #72321 arrived with a regression test covering the already_sent branch, so we salvaged that one. Earliest report credit goes to @Ieqzya in #72228. Thanks for the correct fix — closing as a duplicate.