maintenance page: hero uses the full-res photo (640px thumb upscaled blurry)

maintenance page english-only (hero + 2x2 grid); merge session-init into the warmup stopwatch (measured 10-15s, not 30s); trim stale comments

quota banner: the 60s->30s->shared-pool ladder and the 60s-balance advice, zh/en; app pool relabeled shared (it is not owner-billed)

ladder drops the 10s rung (cannot outlive warmup); countdown anchors at schedule time — ZeroGPU counts the queue wait into the duration window

trim dead paths: one-level-up layout, JVE/MIMO env knobs shadowed by the /data fast path, CKPT_ROOT detector fallback, jve_p indirection, error.left_s the client never reads; docstring ladder para lives at ws()

lease countdown: fork ticks lease_left each second; client shows remaining during warmup and run

trim: required params, no dead-guard, fewer comments

queue abandon: granted fork bails when the client already left; dead sessions stop laddering into the app pool

acquisition ladder: visitor 60/30/10 then app pool 60/30/10 on quota refusals; client shows pool/seconds/2x for wait+grant; lease expiry is a normal end, not an error; busy vs quota split; /ws honors maintenance

maintenance page behind JOYOMNI_MAINTENANCE: notice + team gallery

wan555-equivalent acquisition: one visitor-billed @spaces.GPU per session, nothing else

camp for a card: one visitor window, then the app pool re-queues every 60s window (10 min cap)

JOYOMNI_GPU_SIZE env picks the ZeroGPU tier (xlarge default, large half-card)

docstring and turnstile comment match the visitor-pays design

visitor quota pays for usage (wan555 pattern): mint identity restored; free waiting room + visitor-then-pool single schedule

gradio-pattern acquisition: free in-app waiting room, single app-pool schedule per turn, visitors never billed; drop the mint/probe machinery

remove the dead gpu_busy state

send the full error message, no truncation

docstring matches the probe ladder

lease ladder 60/30/10 via free probes; 120s per session is too expensive

billing-free waiting: poll enable_queue=false probes; gpu_wait/gpu_ready/quota states in the client

visitors never wait in a billed queue: app pool first, instant-refusal fallback only, any queue timeout aborts

queue congestion: skip to the app pool instead of re-burning visitor reservations; live queue status in the client

quota exhaustion: structured error + bilingual banner instead of a raw traceback

team gallery: add wenxun-x screenshot

team gallery: strip above the fold + lightbox originals, served only where /team.json exists

trim kernel-path comment

drop the sageattention wheel: this GPU class runs cuDNN (8/20 bench); runtime falls back gracefully

pro6000 runs plain cuDNN SDPA: drop the sage-on default; remove dead JOYOMNI_TXT_PARALLEL switch

drop redundant fallbacks: single error-report layer, all-or-nothing AOTI preload, faithful finalize flush

enable full local-style recording (owner's decision): input/output mp4 segments + prompts.json to /data/recordings, finalize_recording message, /download_last endpoint

drop external space citation from the no-Blocks note

prune settled forensics and stale comments; dedupe AOTI dispatch closures; drop dead pe_report init

print diagnostics in full (drop error/log truncation); align joyomni_streaming with repo HEAD (drop dormant skip-warmup gate)

speak the zerogpu-headers postMessage handshake with the huggingface.co parent (gradio protocol); /mint stashes the parent-minted quota headers, /ws redeems the ticket

claim user identity over GET /mint (edge translates __sign on HTTP only) and hand /ws a one-time ticket; visitor tier now runs user-scoped

forward HF __sign JWT on the WS URL so logged-in users bill their own quota; log/report token identity; add ?probe=1 identity check

drop the skip-warmup default (ZeroGPU packing hangs without load-time VAE warmups — verified over five cycles); remove the finished token probe

revert gradio-lease experiment: any Blocks instance breaks the FastAPI-demo startup path (4 cycles verified; reference streaming Space uses the no-Blocks topology). Keep the boot speedup + ladder.

trigger ZeroGPU packing + startup report manually at import (the launch hook never fires for a FastAPI demo once a mounted Blocks exists)

bulletproof blocks tail: bracket prints per step, full traceback to stdout, automatic fallback to direct scheduling if the gradio path fails

JOYOMNI_SKIP_VAE_COMPILE_WARMUP: skip emulation-time VAE warmups on Space boot (forks never use the compile wrappers) — ~4-6 min faster cold start

restore @spaces.GPU startup sentinel (dropped in the blocks rewrite -> RUNTIME_ERROR: no GPU function detected)

lease claim failure surfaces to the client instead of hanging the session

native gradio lease path (LTX-style): frontend claims the GPU via a real gradio queue endpoint (ZeroGPU handshake = personal billing/PRO priority); direct ladder kept as 6s fallback

re-add token probe: user-blob presence + cookie/auth-header flags (who actually gets billed?)

ladder steps 120/60/20; log which rung acquired the lease; silence the harmless gradio-queue warning

adaptive lease ladder (120/55/25s per tier): the scheduler caps requests at the identity's remaining balance, so a fixed duration locks out low-balance identities entirely

remove token-claims probe (experiment concluded: HF-page visitors get user-scoped tokens; tier-1 already bills the visitor personally)

log x-ip-token claims (auth level + user) to determine whether the edge user-scopes tokens for HF-page visitors