feat(a2b): audio — wire play-bgm/play-voice, fix BGM addressing

Wire audio end-to-end (OGG plays natively in Godot; no Frida, no decode):
- IHost.PlayBgm/PlayVoice + VM dispatch for play-bgm(0xbf)/play-voice(0xc4).
  Non-Godot hosts no-op them so --selftest + engine 8/8 stay byte-identical.
- GodotAdvHost resolves id->OGG; Main plays via two AudioStreamPlayer nodes
  (BGM looping; voice interrupt-on-new).

Key finding (by-ear, systematic-debugging): the two audio ops use DIFFERENT
addressing — the earlier "unified manifest" assumption was wrong for BGM.
- play-voice -> per-scene manifest files[base+id] (offset 0, same as textures).
  Confirmed by ear; upgraded med->HIGH.
- play-bgm  -> DIRECT LITERAL NAME BGM{id:03d}.OGG, NOT the manifest.
  Real game: play-bgm 5->BGM005, 8->BGM008 (manifest gave +1). Proven by
  play-bgm 0x23->BGM035 (real standalone track; BGM set skips 030-034) that the
  manifest mis-resolved to a graphics entry. Fix is BGM-only:
  ResourceMap.BgmPathById; voices/textures unchanged.

Also: Lily's silence root-caused as correct form-gating (G[0xa57/0xa58/0xa59]),
left unseeded by choice (no dummy state). Added diagnostic
`Age.Cli audio <SCENE.BIN> [0xADDR=VAL ...]`. Corrected opcodes.toml (play-bgm
direct-name, play-voice HIGH) + docs/memory (dropped the bogus unified-manifest
/ Frida-BGM006 claims).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/phase-a-slice-plan.md

@@ -210,6 +210,41 @@ dims) + the sprite position/registration ops so geometry is correct; add alpha/a
 green chromakey; likely move to a proper canvas/blit compositor. Fixes sprites, background placement, and
 fades together. (Superseded: the id-specific plan in `docs/superpowers/plans/2026-07-06-a2b-background.md`.)
 
+### A2b-Audio — WIRED, plays end-to-end (2026-07-06)
+Audio wired, OGG plays natively in Godot (no Frida, no decode/geometry work). **KEY FINDING — the two audio
+ops use DIFFERENT addressing (the initial "unified manifest" assumption was WRONG for BGM):**
+- **`play-voice`** → per-scene manifest `files[base+id]`, **offset 0** (same as `set-texture`).
+- **`play-bgm`** → **DIRECT LITERAL NAME** `id → BGM{id:03d}.OGG` (DATA3), NOT the manifest.
+
+**Shipped:** `IHost.PlayBgm/PlayVoice`; VM dispatch routes `play-bgm`(0xbf)/`play-voice`(0xc4) (both argc 1);
+the three non-Godot hosts (`CaptureHost`, test `RecHost`/`CountHost`) no-op them so `--selftest` + engine 8/8
+stay byte-identical (audio ops still `pc+1`, step count unchanged); `ResourceMap.BgmPathById(id)` (direct
+name) for BGM + `ResourceMap.AudioPath(AssetEntry)` (manifest `Resolve`) for voice; `Main` loads via
+`AudioStreamOggVorbis.LoadFromBuffer` into two `AudioStreamPlayer` nodes (BGM `Loop=true`; voice `Loop=false`,
+interrupt-on-new). Headless run: 0 OGG-load failures, selftest byte-parity OK.
+
+**BY-EAR VALIDATED (2026-07-06, systematic-debugging).** User confirmed voices play on their lines
+(`play-voice` med→HIGH). Two reports root-caused:
+- **BGM off-by-one → FIXED (real root cause, resolver changed for BGM only).** Real game plays BGM005 for
+  `play-bgm 0x5` and BGM008 for `0x8`; we mis-played BGM006/009 because we resolved BGM via the manifest
+  (`files[5]=BGM006`). BGM is actually addressed by **direct name** `BGM{id:03d}.OGG`. **Proof:** `play-bgm
+  0x23 → BGM035.OGG`, a real standalone track (the BGM set skips 030-034) that the manifest mis-resolved to a
+  graphics entry (`files[35]=EV049AA.AGF`). Voices are NOT off-by-one — the manifest interleaves graphics/
+  voice (`files[35]=EV049AA`, `[36]=MAN999`, `[37]=EV052CA`, `[38]=SYL0001`), so `id-1` would land voices on
+  `.AGF` (silent) but they play ⇒ voice offset is exactly 0. So the fix is BGM-specific; voices/textures
+  unchanged. Corrects the earlier "Frida-confirmed play-bgm 5→BGM006" record (a mis-attribution).
+- **Lily silent = correct, form-gated (NOT a bug).** Her lines use a 3-way dispatch on form flags
+  `G[0xa57]`(A)/`G[0xa58]`(B)/`G[0xa59]`(C): exactly one is 1 in the real game (her current form), else the
+  line `jmp`s past with no voice. Our harness seeds no globals → all zero → every Lily line skipped. Proven
+  by seeding: `audio SC0000.BIN 0xa57=1` → 35 LILA clips fire in order (form B→LILB, C→LILC). Left unseeded
+  by user choice (no dummy state); Lily stays silent until real cross-scene state flow (Phase B) exists.
+
+**Diagnostic tool added:** `Age.Cli audio <SCENE.BIN> [0xADDR=VAL ...]` — runs a scene and dumps executed
+`play-bgm`/`play-voice` ops in order (BGM direct-name, voice manifest), optional global seeding. Used for all
+of the above. **Watch-items:** BGM looping is whole-file for now (Eushully OGGs may carry `LOOPSTART`/
+`LOOPLENGTH` Vorbis comments — refine later); `play-sound-effect`(0xb4, argc 2) left stubbed (arg roles
+unconfirmed).
+
 ---
 
 ## Risks / open questions for A0