fix(#55): close YoutubeDL, geoip reader, and the audio-clip retain graph

[Rushaway]

Addresses the memory-leak report by fixing every concrete leak found on inspection. Best-effort: the remaining question ("is it fully gone?") needs a tracemalloc / objgraph pass on a live server over days — but each item below is a real retained-resource bug.

1. YoutubeDL never released (YouTube path)

get_url_youtube_info built a yt_dlp.YoutubeDL and never called close(). Every !yt / !yts — and every retry inside get_first_valid_entry — left the extractor's HTTP connection pools and cookie jar pinned until GC. Matches the report that memory grows with each new external-lookup command but not on a repeat (yt-dlp serves the repeat from cache).

Fix: with yt_dlp.YoutubeDL(ydl_opts) as ydl:.

2. Audio-clip object graph pinned after playback

FFmpegAudioPlayer.Stop() now clears self.callbacks after firing them. The Play/Stop/Update callbacks close over the AudioClip and the AudioManager (and the !say / !dec temp-file cleanup), and the player is reachable back through those closures. While a pending _updater / _read_stream / _reap_process task still held the player, that whole graph stayed alive — refcounting alone wouldn't free it. Simulated: with the clear, the graph is released as soon as Stop runs; without it, it waits on cyclic GC.

3. AntiSpam / Advertiser last_clips growth

last_clips was only pruned inside SpamCheck / Think, which only run while a dominant clip is updating, so a burst that leaves no dominant clip behind kept its entries until the next one played. Extracted the existing prune into _prune_last_clips() and also call it on every OnStop(). OnUpdate() now uses .get() and returns cleanly instead of raising KeyError (then logged) when the entry is already gone.

A randomized equivalence check (20k cases) confirms _prune_last_clips() + the untouched SpamCheck keeps the same entry set and counted duration as the original inline loop — detection behaviour is unchanged, only the cleanup cadence.

4. geoip2.database.Reader leaked on reload

OpenWeather opened a Reader (an mmap of the GeoIP DB) that was never closed, and CommandHandler.Setup() builds a fresh command instance on every !reload. Added BaseCommand.close() (no-op) with an OpenWeather override that closes the reader, and CommandHandler.Setup() now calls close() on the outgoing commands before dropping them.

Context

Much of the original 2023 report was already chipped away by #137-146 and #160-170 (zombie ffmpeg, unclosed sockets/writers, cancellation leaks). This picks up what was left.

Test plan

• ruff check / ruff format / mypy clean.
• Retained-graph release simulated (item 2); prune equivalence fuzzed 20k cases (item 3).
• yt_dlp / aiohttp not installed in the check env — YoutubeDL has been a context manager for years, but a live !yt + !weather + !reload smoke test before merge is worth it.
• Live tracemalloc run over a few days to confirm the curve is flat.

Merge order

Batch with #164 (#171) and #169 (#172); sequential VERSION patch bumps (this one -> 1.8.25). One-line VERSION rebase if it doesn't merge last.

Closes #55

🤖 Generated with Claude Code

[Dolly132]

LGTM, tested and works fine without errors/exceptions