Dual encoding for both WVOX enums via the readEnumField
template: eventKind (int 0..8 OR token "greeting"/
"aggro"/"death"/"queststart"/"questprogress"/
"questcomplete"/"goodbye"/"special"/"phase") and
genderHint (int 0..2 OR token "male"/"female"/"both").
volumeDb serializes as signed int8 since boss-call
clips use +5dB above ambient and quiet whisper clips
can use -10dB or lower.
transcript serializes as a plain JSON string —
operators editing the JSON sidecar can rewrite voice
transcripts for accessibility (TTS engines and chat-
bubble subtitles) without binary tooling.
All 3 presets (questgiver/boss/vendor) byte-identical
roundtrip OK. CLI flag count 1225 -> 1227.
Novel replacement for the implicit per-NPC voice dialog
system vanilla WoW encoded across CreatureTextSounds
(server-side aggro/death barks), npc_text (gossip audio
cross-references), and per-quest dialog blobs. Each
entry binds one NPC to one voice clip for one
triggering event with metadata covering audio path,
duration, volume, gender hint for randomized casts,
variant index for multiple lines per event, and a
transcript field for accessibility (TTS engines + chat-
bubble subtitles).
Nine eventKind values cover the full NPC dialog
surface: Greeting / Aggro / Death / QuestStart /
QuestProgress / QuestComplete / Goodbye / Special /
Phase. The Phase kind specifically supports boss-fight
percentage milestones (75%/50%/25% transitions) where
multiple Phase entries with distinct variantIndex
disambiguate the boss-encounter scripting.
Three preset emitters: makeQuestgiver (5-clip canonical
quest dialog flow), makeBoss (6-clip Lich King fight
with phase milestones at 75/50/25%, special mechanic
call at +5dB for raid audibility, death line),
makeVendor (4-clip vendor interaction).
Validator's most novel check is per-(npcId, eventKind,
variantIndex) triple uniqueness — two clips with all
three matching would be ambiguous when the trigger
handler picks one randomly. The vendor preset
originally bound both Buy and Sell to (Special, 0)
which the validator caught and flagged before commit;
fix uses variantIndex 0 for Buy and 1 for Sell so the
trigger handler can distinguish.
Validator also warns on durationMs=0 with non-empty
audioPath (subtitle sync impossible), volumeDb outside
[-20,+6] (clip risk), and empty transcript (TTS +
chat-bubble subtitle would be blank).
Format count 113 -> 114. CLI flag count 1220 -> 1225.
