Disco DeDisco 41217d5438 my-sea voice Phase C: WebRTC mesh signaling app + TURN endpoint + voice-btn wiring + coturn infra — TDD
Phase C (final) of the my-sea invite → spectator → voice blueprint. Self-
hosted WebRTC mesh voice, built room-general but wired for my-sea only; epic
6-seat rooms reuse the same consumer later (key on Room.id). Media never
touches the server — only signaling is relayed. Built from the blueprint's
distilled spec (disco-voice-mesh.pdf unreadable in-env: no poppler/pypdf).

- C1: new apps/voice/ — RoomVoiceConsumer (AsyncJsonWebsocketConsumer):
  signaling-only relay (room group voice.<room_id> + per-peer peer.<uuid>;
  hello→present handshake, offer/answer/ice routed by target/source, left on
  disconnect). room_id is a STRING kwarg (mysea-<owner_id> now). _can_join
  gates: mysea → owner OR present invitee (token deposited, not left); epic
  UUID → seated gamer (later). routing.py ws/voice/<str:room_id>/; asgi.py
  aggregates epic + voice urlpatterns under AuthMiddlewareStack.
  voice-mesh.js: VoiceRoom client (getUserMedia AEC/NS/AGC, mesh
  RTCPeerConnection, newcomer-offers handshake, tuneOpus SDP munge =
  inbandfec+dtx+40kbps cap, mute via getAudioTracks().enabled), lazy-loaded.
- C2: apps/api VoiceTURNCredentialsAPI at /api/voice/turn-credentials/ —
  coturn use-auth-secret REST scheme: username=<expiry>:<user_id>,
  credential=base64(HMAC-SHA1(username, COTURN_SHARED_SECRET)) + stun/turn
  iceServers + ttl. Authenticated-only. 4 ITs (HMAC shape, auth gate).
- C3: settings COTURN_SHARED_SECRET / COTURN_TURN_HOST / COTURN_REALM /
  COTURN_TTL env block.
- C4: #id_voice_btn wiring — _burger.html renders .active + data-room-id when
  voice_active; burger-btn.js bindVoiceBtn (active click → lazy-load
  voice-mesh.js → join / toggle-mute; inactive → existing 2-pulse flash).
  my_sea (owner) + my_sea_visit (spectator) views compute voice_active
  (open 24h window) + voice_room_id=mysea-<owner_id>; spectator page now
  includes the burger. 4 voice-context ITs.
- C5: infra/coturn.conf.j2 (use-auth-secret, the external-ip footgun, relay
  port range, TLS 5349, peer-IP lockdown) + infra/coturn-playbook.yaml
  (dedicated droplet, PySwiss-style split: install coturn, template conf, ufw
  3478/5349/49152-65535, systemd enable) + [coturn] inventory placeholder.
  *** Manual ops step: provision the droplet + fill inventory before voice
  works on staging/prod; CI/local need none of it. ***
- C6: 8 channels ITs (@tag channels) — connect/auth/_can_join gate (owner,
  present invitee, stranger, not-present, anon) + hello/present handshake +
  offer routing + left-on-disconnect. Scope-injected; TransactionTestCase.
- JS: VoiceMeshSpec.js (tuneOpus) + voice-mesh.js registered in SpecRunner.

1440 IT/UT green; voice channels IT + full Jasmine + voice-btn FT green.
Voice infra is code-complete — provision the coturn droplet to go live.

Code architected by Disco DeDisco <discodedisco@outlook.com>
Git commit message Co-Authored-By:
Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-27 13:57:09 -04:00
feat: import Minchiate Fiorentine 1860-1890 deck assets (98 PNGs, alpha-channel transparent bg) + naming convention v2 + pngquant optimization tooling. Public-domain 1860-1890 lithograph scans sourced from Wikimedia (single download series + trump 11 Il Gobbo individually-sourced from same era series); user removed white backgrounds in Photoshop to leave irregular card-shape with transparent canvas. Filenames v2-conformant per [[reference-card-image-naming-convention]] (revised from v1 of 2026-05-24): deck slug carries -1860-1890 publication-year suffix so future Minchiate Fiorentine variants from different publishers/eras coexist cleanly; courts use rank-number-prefix (batons-11-page not batons-page) for linear sort key; trumps carry both numeric rank AND italian-name suffix (trumps-01-papa-uno, trumps-11-il-gobbo) for forensic identification across variant decks the user plans to sell. Il Matto (unnumbered Fool in Minchiate tradition) assigned rank 00 to give it a sortable position. Five trump filenames had elided-apostrophe slugs restored from the download source (-lacqua-l-acqua etc.). Card-back at <deck-slug>-back.png sorts alphabetically before all suit categories — no separate card-back/ subdir needed. pngquant 2.17.0 installed at C:\Users\adamc\AppData\Local\Programs\pngquant\, added to user PATH (effective next session) for future deck imports; ran with --quality=65-85 --speed=1 --strip --skip-if-larger for 57.6% size reduction (86.6 MB → 36.7 MB total, 935 KB → 383 KB avg). Second pass at --quality=40-65 hit pngquant's floor (only 0.5% further reduction — re-quantizing an already-quantized image has little headroom). Il Gobbo from Wikimedia was a dimensional outlier (1426x2366 vs siblings ~620x1024) — resized via System.Drawing HighQualityBicubic to 620x1029 before optimization. Format32bppArgb alpha channel verified intact across samples after optimization pass. Visually validated by user: cards must fill entire screen before any pixelization visible. Sprint A precursor — DeckVariant.has_card_images toggle + image-rendering template branch per [[project-image-based-deck-face-rendering]] follows in subsequent commits, will consume these assets in 6 surfaces (my_sign, my_sea, both billboard applets, room, game_kit). Asset set also unblocks downstream Sprint C+B [[project-deck-segment-model]] (admin form will require image upload + enforce naming convention) and Sprint D [[project-card-deck-icon]] (uses -back.png as the deck-stack icon's repeating card-face). Future: when Sprint B's admin form ships, wire pngquant into an optimize_card_images management command so admin uploads auto-optimize on save. Gitignore line src/apps/epic/static/apps/epic/images/cards-faces/minchiate-fiorentine/ dropped — v1 staging dir deleted (was only ever the rename-staging set; superseded by v2-named optimized set). Total disk delta: +36.7 MB binary content. No code changes — pure asset + convention import
2026-05-24 22:39:21 -04:00
2026-01-13 20:58:25 -05:00
Description
No description provided
155 MiB
Languages
Python 45%
JavaScript 37.6%
HTML 9%
SCSS 8.2%
Jinja 0.1%