docs(backlog): why metrics are relative — absolute-score ceiling + arousal vs valence

fix(logs): suppress per-queue-join StarletteDeprecationWarning flooding HF logs

docs: record chunk-fix verification — full-res neeraj_reel_2 scored at 63.2GB peak

feat(mps): chunk the frame processor() call — bound the long/HD memory transient

docs: instagram-curve contrast guide + chunk-processor memory-fix design

feat(ui): lock Score button while scoring + persist/download curves + label fix

fix(progress): n_pass = measured sweep count (cache-on=2, cache-off=2*windows)

fix(dedup): correct the cache-key rationale comment (Theory A proven)

docs: backlog — cross-video engagement-benchmark comparator (exemplar-based)

feat(mps): dedup-cache machinery — hardened key, clear-per-Score, n_pass (gated)

docs: dedup-cache enable plan (tribe-brain reviewed, parity-gated)

feat(ui): timeline playhead follows video playback

docs: playhead-follow design (tribe-brain APPROVED, pending implementation)

feat(ui): in-card determinate progress bar via gr.Timer (Option T)

feat(ui): determinate per-clip encode progress bar (Option X)

refine(mps): blocked attention BLOCK=1024 + stock-fallback for edge cases (mask/causal/dropout/training/small-N)

feat(mps): blocked/streaming V-JEPA2 attention (Approach C) — structural memory bound, replaces per-layer empty_cache hook

fix(windowing): global z-score post-stitch (cross-window scale consistency) to cure WIN-1 seam

feat(quality): surface 'no speech track -> video-only' notice when no Word events

feat(mps): wire vm_stat memguard into the dedup encode loop (start gate + periodic abort)

perf(mps): HIGH watermark 0.6->0.45 (graceful OOM under the gate) + document B=1

feat(mps): per-layer empty_cache hook on V-JEPA2 encoder layers (TRIBE_MPS_EMPTY_EVERY)

feat(memguard): vm_stat system-memory governor for local MPS runs

docs: implementation plan for local MPS long-video Quality scoring (P1-P3, 8 tasks)

docs: incorporate tribe-brain spec review (ASR 0it re-diagnosed, WIN-1 verify-first, hook path + dedup-cache guard, P4 split out)

docs: design spec for local MPS long-video Quality scoring (V-JEPA2 OOM fix + adaptive governor + Quality path + WIN-1)

docs: refine dedup residual (low-risk w/ cache off) + cheap A/B closure method

docs: note frame-dedup ON + cache OFF; flag ≥150s multi-window parity as residual

docs: add KNOWN_ISSUES.md (status, known bugs, gaps, roadmap)

feat(text): preview synthesized speech + remove profiler harness

feat: all modes working on HF — fix Quality hub-writable EACCES

feat: local MPS all-modes-working (audio + text); video deferred to HF

perf: breakdown — add use_dedup + keep_in_ram A/B toggles

perf: breakdown v3 — split loop into H2D vs loader-PULL (CPU)

perf: breakdown v2 — reliable head timing via nn.Module instance hook

perf: breakdown profiler — split W2V-BERT build vs forward, 60s, billed-tagged

perf: add CLEAN breakdown profiler (coarse timers, real-clip upload)

feat: local MPS port + ZeroGPU perf toolkit (dedup, backbone-reuse, pre-warm)

docs: expand README — modes, local/deploy, perf, AOT status, licensing

revert: disable torch.compile on V-JEPA2 (dynamo tracer bug broke video runs)

perf: torch.compile (static AOTInductor graph) on V-JEPA2 encode

perf: keep bf16 encode win; drop clip-batching (correct but compute-bound/slower)

fix(perf): batched encode B=4 + correct exca-descriptor fallback

perf+fix: batched V-JEPA2 encode (~3-8x) + writable hub cache for opt-in ASR/LLaMA

fix: default video/audio to fast no-ASR path + theme/css -> launch() (Gradio 6)

perf: bf16 V-JEPA2 encode via scoped monkeypatch (~2x, fp32 output)

fix: revert blanket bf16 autocast (broke numpy on bf16 outputs)

perf: bf16 autocast on backbones (~1.5-2x encode + ~half activation memory)

perf: preload_from_hub (bake ungated backbones) + TF32 fast matmul

perf: raise extractor batch sizes to exploit ZeroGPU VRAM (~48GB) + VRAM log