[dmccoystephenson]

Makes the walkable hall feel alive: the nearest screens play their game's trailer, each nearby cabinet plays a soft attract loop placed in the room, the most played (and liked) games' marquees glow a little warmer, and a Hall of Fame board by the door lists the best score on each leaderboard. Four commits, each reviewable on its own.

1. Live trailers on the screens (81aa57f, 068387b)

• The screen of the cabinet nearest the visitor plays the game's trailer (silent, looping) instead of its poster, and goes back to the poster on walking away. Two screens at most on a desktop, one on a phone or in economy (at 10 fps there), only screens in front of the visitor and facing them, with a little hysteresis so the edge of range does not flicker.
• None plays under reduced motion, while a game is open, while the page is hidden, or with ?trailers=0.
• Clips: hall_assets/clips/<slug>.mp4, 29 clips, 1.4 MB in all (384x216, 15 fps, H.264, no audio, faststart), re-encoded by the new tools/make_hall_clips.py from danielstephenson.dev/trailers. They are served same-origin, like the posters, because that host sends no CORS header and a cross-origin video would taint the canvas. The CSP needed no change (default-src 'self' covers same-origin media).
• Hall assets now answer single byte ranges (206/416, Accept-Ranges: bytes): Safari plays a video only from a server that does. Compressed variants are never ranged. server.byteRange() is unit-tested.
• Frames reach the texture asynchronously through VideoFrame.copyTo({format: "RGBA"}) where available, so the hall's frame never stalls on a GPU readback. A synchronous canvas drawImage/getImageData was measured at 70–140 ms per frame under SwiftShader; the async path costs about 3–5 ms per upload. Elsewhere frames go through a small canvas.
• No engine change: hall_texture already updates a same-sized texture in place and regenerates its mipmaps.

2. Spatial attract-loop sound (ff708d9)

• With Sound on (the existing button / "Arcade hum" setting; still off by default), the nearest cabinets (3, or 2 on a phone or in economy) each play a short chiptune phrase synthesised with WebAudio, with no audio files. The phrase is made up from the slug, so it is the same on every visit, and flavoured by the game: a tak text adventure gets a slow minor triangle-wave figure, a Halloween-tagged game a minor one, and the rest a brighter pulse, bell or bass arpeggio. Each plays two bars, then rests for two.
• Placement: a soft inverse-square falloff (silent past ~11 m), panned by direction as the visitor turns, and quieter and low-passed from behind a cabinet. Everything pauses with the hum while a game is open or the page is hidden.
• Measured through an analyser on the output in Chromium: at a cabinet the peaks reach about 0.07, against about 0.04 from across the hall (hum alone, RMS ~0.013). The loops are audible up close and fade to nothing far away.

3. Popularity in the room (0307d03)

• Warm marquees: plays as a share of the most played, plus likes from arcade-social's public /v1/likes/counts, weighted by how many likes exist overall (up to a third, in full from 30 likes in all; with today's 3 likes they barely count). Only the top end warms: today that is env-lib-cpp (0.98), Tidewater (0.29), Beyond Nations (0.13) and FishE (0.05). Dark cabinets in a filtered hall stay dark.
• Hall of Fame board on the near wall, right of the door and across from the counter (layout.json room.fame). It shows each leaderboard's best score (player, game, board, value; at most six, with the most-played games first), plus a "Still open:" line for boards nobody has topped yet. It reads /v1/games/<slug>/boards and then /v1/boards/<slug>/<board>?limit=1 without credentials, three at a time, 2.5 s after load, and keeps the public result in sessionStorage for 10 minutes.

Live data (one score so far)	With sample scores (local only)

Verification

• python3 -m pytest: 375 passed, 1 skipped (after merging main). New tests:
  • node tests hall_trailers_test.cjs (screen choice) and hall_fame_test.cjs (warmth, board rows, fetch and cache, service down), plus additions to hall_life_test.cjs (tunes, spatial).
  • pytest: clips are 16:9 384x216 faststart mp4s under 256 KB each; byte ranges; layout.json names only clips that exist; the Hall of Fame position.
• Browser (tests/browser/hall_atmosphere_check.cjs, Chromium + SwiftShader; this setup forces economy). Desktop 1024x640, iPhone 13, economy, reduce motion and ?trailers=0 were all checked with no console errors. The trailer was shown in every profile except reduce motion and ?trailers=0, where none plays.
• Frame rate A/B in SwiftShader on a shared, noisy box, three runs each, medians:
  • iPhone 13: 40–47 fps with everything on vs 36–49 off.
  • Desktop: 17–25 vs 21–22 off. The desktop shows 2 screens at 15 fps; with SwiftShader every texture upload and mip build runs on the CPU.

Notes for review / integration

• Uses Hall: likes, top scores and achievements on the cabinet card #59 (merged) for <meta name="arcade-social"> and connect-src https://api.<domain>; origin/main (with Hall: likes, top scores and achievements on the cabinet card #59) is merged into this branch and the board and likes were re-checked against it with no shim. Without arcade-social the board reads "Scores are resting" and the marquees go by plays only.
• A trial merge with feat/hall-people conflicts only at the end of tests/test_hall.py (both sides append; keep both).
• The board makes about one request per deployed game (36 today) once per 10 minutes per tab. A single GET /v1/boards/top aggregate in arcade-social would reduce that to one request; it is left as a possible follow-up.
• Open questions resolved toward privacy and economy: the board shows only display names that the public boards already show, with no new data held server-side. Reduced motion stops the trailers, not just the camera. Phones get one trailer at 10 fps and two voices.

---

drafted by Claude on behalf of Daniel Stephenson

🤖 Generated with Claude Code