Warm GPU models on session start to hide Modal cold start

Field Notes README + anonymous gameplay trace logging (Sharing-is-Caring)

Split People/Things, add friend/study/lonely (28 tones), swap to Qwen2.5-7B

Redesign vocabulary (25 tones, no links) + coherence, loading text, act cards

UX polish: single-click glyph edit, tap-to-ask, evidence scroll, tab NEW dots

Rework Lily's voice: leaner prompt, helping questions, empathy meter

Fix first-turn race: ensure a session exists before /api/turn

Ending: use the user-provided final_words.wav; fix first-reply tooltip clipping

Ending polish: warmer/slower voice; memorial on warm dark, not white

Guarantee reachability: tone coverage + progression valves; expose runEnding

Phase 7.3: ending sequence — finale tones, fade to white, her voice, memorial

Ambient: replace piano bed with a soft drone pad; dim the cursor glow

UI: remove glyph category dot; add low ambient bed + cursor glow

UI: double-click glyph to edit in lexicon; click-outside/Esc closes drawer

Drop off-topic moderation gate; keep harmful as a binary safety check

Story pivot: dreams + travel + Mount Fuji; 3 connective tones; Her Things tab

Phase 6: wire Nemotron moderation into /api/turn

Pivot: reframe Lily from visual artist to musician

Music: overlap in-game tone playback so replies flow as a melody

Gameplay: give Lily a persistent mood across turns

Gameplay: enforce Lily's word order + associative answers

Phase 5.2: Nemotron-Mini-4B semantic lexicon scoring

Core-loop fun + story enrichment pass

Phase 3: real Gemma 4 E4B inference on Modal

Phase 0-2 + atmosphere: playable vertical slice on mock inference