Header/footer: dataset and project-page links, model in footer; 2B T1 to the three-seed value

Space card: short description, pin, dataset link, paper and code links

Load LoRA adapter weights on CPU (ZeroGPU spaces>=0.51 fix)

Unpin spaces (platform force-installs its version); keep peft==0.19.1

Pin spaces==0.50.4 + peft==0.19.1 (ZeroGPU startup regression)

Open header/footer paper and repo links in a new tab (iframe breakout)

demo: wait for scenario re-render before scrolling on system select

demo: scroll scenarios into view on system select (mobile, parent-aware)

demo: reveal Tap-for-result via inline style, not body-class (Gradio scoping)

demo: harden parent-scroll helper (scrollToOffset -> scrollTo -> scrollIntoView)

demo: '3. Tap for result' button + parent-aware scroll (works in embed)

demo: page labels v23 -> v24 + T1 stat to v24 held-out numbers

demo: re-curate scenarios (first 3 per system ticket+issue, blockers demoted)

demo: raise mobile chat cap 340px -> 400px (~60vh of a 667 phone)

demo: cap mobile chat panel with fixed px (vh breaks in iframe-resizer embed)

demo: tighten mobile chat panel cap 60vh -> 50vh

demo: cap chat/output panel at 60vh on mobile

demo: hide Plan-trip bar via inline style (beats Gradio scoped CSS)

demo: step-label colons + tighter padding, hide Plan-trip bar on output view

demo: numbered step labels, hide map CTA while running, shorter map

demo: switch to Qwen3.5-9B-metro-v24 (leakage-free retrain)

demo-v3: reserve mobile bottom space for Play now pill

tighten kiosk disclaimer + add AI caveat

demo-v3: 'Play now' affordance on each scenario card

scenarios: lead each system with iconic disruption (mirrors monorepo)

revert landing replay; add first-scenario pulse cue instead

fix landing replay never firing (Textbox not State, drop async JS)

landing replays for all 6 systems + render helpers

landing-page replay infrastructure (no-op until traces commit)

bump torch 2.5.1 → 2.8.0 (ZeroGPU now requires 2.8+)

sync: HF Space ↔ /simulate prompt-construction parity

roll back torch 2.6 + fla path (TTFT + duration cap crashes)

revert regex throttle (TTFT regressed to ~20s)

throttle the regex/mutation inside the streamer loop

bump torch 2.5.1 → 2.6.0 + re-add fla + causal-conv1d

revert fla + causal-conv1d (Triton 3.1/3.2 ABI mismatch)

add flash-linear-attention + causal-conv1d (hybrid layers)

probe runtime attn_implementation post-PEFT merge

per-round timing + tok/s line in chat for perf comparison

switch to flash-attn prebuilt wheel (source build was failing)

flash-attn for ZeroGPU + regional scenario flavor + subtitle

mobile UX pass — padding, scroll, dots, route, disclaimer

single popover on picked station (drop redundant pick label)

clear stale destination marker when restarting origin pick

4:5 portrait map on mobile

reclaim horizontal mobile padding (chrome was eating ~112px)

drop run-lock from chains — was blocking inference

fix run-lock chain stalling Plan/Scenario clicks

tonal soften + BART default + tz auto-pick + map Plan CTA

empty-state hint inside Output pane