feat: gTTS as primary TTS — Coqui XTTS-v2 was never installed on Space

fix: snapshot mode (drop streaming) — HF proxy can't sustain frame uploads

fix: webcam stream actually fires + JS auto-arms record toggle

fix: webcam capture — global cache, .change() event, drop share=True

fix: global frame-cache so click handler can actually see live frames

fix: read webcam directly on click instead of state-stashing

ux: streaming webcam + Take-image button — one-time permission grant

ux: auto-dismiss photo after recognition so next capture is one tap

ux: 2-button Snapshot tab (Take Photo auto-fires recognition; remove debug accordion)

fix: Clear button now resets webcam too + accurate help-text labels

fix: 4 Space runtime bugs found in line-by-line audit

fix: libgles2/libegl1 in Dockerfile + handle missing mp.solutions namespace

fix: pin starlette<0.41 — gradio 4.44.1 incompatible with new TemplateResponse signature

fix: share=True + SPACE_ID env to skip gradio's localhost-check pre-flight

docs: README — fine-tune narrative + hybrid pipeline diagram + model URLs

fix: set SYSTEM=spaces so gradio skips its localhost-check on HF Docker

fix: switch to Docker SDK so Python + launch are fully under our control

fix: remove python_version pin (was breaking gradio 4.44.1 launch on HF Spaces)

fix: app.py — let HF Spaces gradio runtime handle server_name/port

fix: pin python 3.10 (3.11 hit gradio localhost-check runtime error)

fix: requirements.txt — un-merge concatenated lines + drop server-only deps

fix: download MLP + MediaPipe weights from HF Hub instead of bundling

feat: MediaPipe+MLP landmark classifier — 90% ASL fingerspelling accuracy

fix: websockets<13 to satisfy gradio-client 1.3.0 dep range

fix: pillow<11.0 to satisfy gradio 4.44.1 dep range

Switch to Qwen3-only stack: VLM=Qwen3-VL-32B, composer=Qwen3-8B

docs: pitch deck + demo video script + lablab submission form content

docs(readme): tighten Why AMD + add For Deaf-led teams section

docs(walkthrough): Deaf-community ethics + MI300X comparison + future work

fix: threading.Lock on LandmarkExtractor + composer prompt format match

feat(ui): Record-sign tab — multi-frame burst capture for full ASL words

feat(backend): /recognize accepts frames list + body-size cap

feat(vlm): add recognize_sign_from_frames for multi-image VLM input

feat(prompts): NVIDIA-pattern single-frame + multi-frame VLM prompts

perf: cache OpenAI-compatible client per (provider, model) tuple

refactor: extract closed vocabulary into signbridge.vocab module

fix: 4 BLOCKERs + 6 IMPORTANTs from deep-check audit

docs(rules): add 'research before decisions' + 'deep-check before checkpoints' to standing rules

feat(submission): generate 1280x640 cover image for HF Space + lablab

fix: clean up stale Apache 2.0 mentions + restore walkthrough spacing

fix(submission): move HF Space to event org lablab-ai-amd-developer-hackathon

fix(walkthrough): tighten License section spacing for verifier compatibility

fix(license): switch from Apache 2.0 to MIT to satisfy hackathon rule

docs: snapshot competition requirements as authoritative section in CLAUDE.md

fix: composite alpha onto white before VLM round-trip

feat: end-to-end smoke test + gold-set accuracy harness

docs: fill in GitHub + HF Space URLs in CLAUDE.md

fix: shorten README short_description to <60 chars (HF Space requirement)

chore: drop Build-in-Public extra challenge

feat: SignBridge initial scaffold + tests