fix(client): bound actor sound channels with a recycling pool (#489)

[CoreyRDean]

Closes #489.

Non-technical summary

After a sustained play session with several audible actors moving and fighting, the game client (bin/Client.exe) would hard-crash — no error dialog, empty log. This PR fixes that. Actor sound effects (footsteps, combat/death cries, and server-triggered positional sounds) now reuse a small fixed set of audio channels instead of allocating a fresh one every time and never giving it back. The client can now play actor audio indefinitely without exhausting the sound backend.

Technical summary

Per-step actor sounds were fire-and-forget: three sites called EmitSound(...) and discarded the channel handle it returns, so the OpenAL backend's sources accumulated until allocation failed and the process crashed (issue #489). The managed SoundZone path already avoided this by keeping its handle and ChannelPlaying()-checking it; the per-step actor paths did not.

• New EmitActorSound(Snd, EN) in Actors3D.bb — routes every actor sound through a fixed-size ring (ActorSoundPoolSize = 24) of channel handles. Before reusing a slot it StopChannel()s whatever is still playing there (LRU reclaim), so the live actor-sound channel count is capped at the pool size regardless of session length or actor count. Mirrors the existing SoundZone reuse pattern.
• Three call sites routed through the pool: footstep EmitSound (Client.bb), the attack/hit/death EmitSound in PlayActorSound (Actors3D.bb), and the P_Sound packet handler (ClientNet.bb) — the last is a server-triggerable sibling of the same leak, included per the repo's sibling-asymmetry doctrine.
• A Snd = 0 guard skips emitting silence for a missing/unregistered sound id — a small improvement over the old raw EmitSound(0, ...).
• The ClientNet.bb comment shifted downstream handler line numbers, so the auto-generated docs/protocol/index.md is regenerated (line-number shifts only — no packet semantics changed).

No breaking changes. The helper is additive; the call-site swaps are behavior-preserving except for the channel cap and the Snd=0 skip.

Acceptance criteria + test results

New non-Strict regression test src/Tests/Modules/ActorSoundPoolTest.bb (passes under test.bat ActorSoundPool) replicates the pool bookkeeping verbatim with the audio primitives stubbed, and pins:

• Snd=0 short-circuits — returns 0, no channel consumed, cursor unchanged, nothing stopped.
• First fill reclaims nothing — the initial poolSize emits hit empty slots; cursor wraps to 0.
• Cursor stays in [0, poolSize) across many emits (Dim index always in bounds).
• LRU reclaim on wrap — emit #(poolSize+1) stops the oldest channel, #(poolSize+2) the next-oldest.
• Live channel count capped — after K emits, exactly K - poolSize channels were stopped, so at most poolSize are ever live.
• Conditional reclaim — a slot whose sound already finished (ChannelPlaying False) is overwritten without a needless StopChannel.
• Full compile.bat clean across all five engine targets + all seven tools (Actors3D.bb is shared by Client and GUE).
• Both generator staleness gates (gen_packet_index.sh --check, gen_bvm_reference.sh --check) pass.

Trade-offs, deferred follow-ups, remaining gap

• Runtime confirmation of the crash fix is walkthrough-level, not end-to-end. Agent-launched Graphics3D is blocked in this environment, so the no-crash outcome is established by the issue's own bisection + the unit-tested channel cap + static reasoning, not a live multi-actor session. The fix is the approach the issue itself recommends.
• Re-enabling the per-step actor sounds in data/ (currently disabled as the data-side workaround) is deferred to the default-project owner — it's a content change, not an engine change, and is safe to flip once this lands.
• Pool size (24) is a conservative constant; if a future zone needs more simultaneous distinct actor sounds it can be raised, with the same cap guarantee.

[rockinraymond]

Dang this one is huge nice work!