v0.6.0 — Centauri Carbon 2 + Canvas multi-filament

Added

• Centauri Carbon 2 support. The CC2 uses MQTT on port 1883
  (instead of the CC1's WebSocket SDCP on 3030) with a JSON-RPC
  envelope, X-Token auth, and a completely different command set
  (1001/1002/1020–1031 vs 0/1/128–512). This release adds:
  • CC2Printer — async MQTT client that subclasses Printer,
    translating CC2 payloads into the same Status / Attributes
    models so every surface (CLI, HTTP, MCP, web UI) works unchanged.
  • connect_auto(host, access_code=...) — factory that detects the
    model (probes :1883 only — a CC1 answers with a harmless kernel
    RST, and the CC1's real WebSocket connect doubles as its own probe)
    and returns the right Printer subclass. Callers never need to
    know which model they're talking to.
  • --access-code / PYCENTAURI_ACCESS_CODE on every CLI command
    and MCP launch (required for CC2, ignored for CC1).
  • paho-mqtt>=2.0 added to core dependencies.
• CC2 status exposes gcode_move.speed — live head speed as the
  commanded speed of the current move, in mm/min (÷60 matches the
  printer screen's mm/s readout)
  and speed_mode (integer 0–3) — fields the CC1 lacks entirely. Both
  are preserved in Status.raw["_cc2"] alongside remaining_time_sec,
  filament_detected, external_device, and exception_status.
• CC2 methods 1028 (SET_TEMPERATURE), 1029 (SET_LIGHT), 1030
  (SET_FAN_SPEED), 1031 (SET_PRINT_SPEED), 1023 (RESUME), and 1036
  (PRINT_TASK_LIST) all confirmed working. 1042 (VIDEO_STREAM) and
  1044 (GET_FILE_LIST) are non-responsive on firmware 01.03.02.51.
• Canvas multi-filament system integration (CC2 only):
  • Printer.canvas_status() — returns CanvasStatus with connected
    units, tray list (filament name/type/color/brand/temp range/loaded
    status), active tray, and auto-refill state. Method 2005.
  • Printer.set_auto_refill(enabled) — toggles auto-refill. Method
    2004. Gated by enable_control.
  • CLI: centauri canvas (read) + centauri refill --on/--off
    (write, requires --enable-control).
  • HTTP: GET /canvas (always) + POST /canvas/refill (control-gated).
  • MCP: get_canvas_status (always) + set_auto_refill (control-gated).
  • CanvasStatus, CanvasUnit, CanvasTray model classes exported
    from the package for typed access.

Fixed (pre-release hardening — none of these shipped in an earlier release)

• CC1 status()/watch() self-heal through a wedged push
  scheduler. Discovered live 2026-07-04 on V0.3.0-o: the firmware can
  enter a state where Cmd 512 subscribes are accepted but no status
  frames are ever pushed, while the request path stays healthy — and a
  reboot does not clear it (verified 2026-07-05). Both methods now fall
  back to explicit Cmd 0 status requests (documented to emit a
  one-shot status frame) whenever a subscribe produces no push within
  one period, instead of hanging forever. See docs/PROTOCOL.md
  "Operational quirks".
• CC2Printer.watch() now polls method 1002 inline whenever pushes go
  quiet instead of silently terminating after one push period — CC2
  centauri watch and the SSE stream no longer die on an idle printer.
• MQTT responses are correlated by the client's exact api_response
  topic and non-push method. Previously any client's response — or the
  printer's own 6000/6008 broadcasts, which reuse small auto-increment
  ids — could resolve a pending request future with the wrong payload
  (including reporting success for a control command).
• Registration is re-published from _on_connect, so paho's automatic
  MQTT reconnect re-registers with the printer and pushes survive a
  printer reboot without restarting centauri server.
• All MQTT-callback work that touches futures, events, queues, or the
  merged status dict is marshalled onto the asyncio loop (fixes the
  cross-thread delta merge and late set_result races).
• GET /stream uses a bounded connect timeout (read stays unbounded
  for MJPEG) instead of hanging forever against a wedged camera port.
• GET /canvas returns 504 for printer timeouts, reserving 501 for
  "no Canvas on this printer"; the web UI now retries transient Canvas
  failures instead of hiding the panel until a page reload.
• A wrong CC2 access code now raises a clear auth error (MQTT reason
  code / HTTP 401 hint) instead of a generic timeout or raw traceback.
• connect_auto() and centauri rtsp no longer probe :3030 — the CC1
  WebSocket connect doubles as the probe, avoiding the connect/close
  churn that port is sensitive to. centauri rtsp and the HTTP server
  both pick the right MJPEG port (CC1 :3031 / CC2 :8080) automatically,
  including reconnects after starting with the printer offline.
• Unknown CC2 machine_status values map to sentinel code 99 instead
  of leaking raw values into the CC1 status-code space; PrintStatus
  gains constants for the CC2 codes 27/28/29.
• CanvasStatus.from_payload survives malformed/absent canvas_info
  payloads; CanvasUnit/CanvasTray are exported from the package
  root; typing_extensions declared as a direct dependency.
• CC2 sessions send the app-level {"type": "PING"} keepalive every
  30 s. The firmware expires a client's registration after several
  quiet minutes and then silently stops answering that session's
  requests — the MQTT connection stays up, responses just never come
  (verified 2026-07-05: a dashboard went request-deaf after ~6 minutes
  while a fresh session answered instantly). MQTT-level keepalive does
  not prevent this; only the app-level PING does.
• centauri server bounds uvicorn's graceful shutdown at 5 s. Open
  SSE/MJPEG streams never close on their own, and without the bound a
  stopped server lingered forever as a zombie still holding a printer
  connection (observed 2026-06-23 and 2026-07-05).
• CC2 lifecycle commands (start/pause/stop/resume) use a 90 s timeout —
  the firmware answers only after the mechanical sequence completes
  (observed 2026-07-05: a resume succeeded but its confirmation
  arrived after the old 15 s window, surfacing a false error). When a
  confirmation still doesn't arrive, POST /print/* returns 504
  ("sent, unconfirmed — check printer status") instead of 502, and the
  web UI shows it as a warning rather than a failure.

Tests

• New tests/test_cc2_mapping.py: pure dict-in/dict-out coverage of
  the CC2→CC1 translation layer — machine_status/sub_status mapping
  (including purge-zone filament-switch detection and its progress
  guard), partial-delta deep merge, full 1002 payload round-trip, and
  Canvas parsing. Plus a server test pinning GET /canvas → 501 on
  CC1.

Documentation

• docs/PROTOCOL.md gains a full CC2 section: MQTT transport, topic
  structure, connection/registration dance, the confirmed method table
  (12 probed, 10 working), error codes, the complete status payload,
  and a CC1-vs-CC2 comparison table. Probed on 2026-06-30 through
  2026-07-04 against firmware 01.03.02.51 (Canvas methods 2004/2005
  verified live with a 4-slot Canvas attached; CC2 webcam confirmed
  unauthenticated MJPEG on :8080).
• docs/ARCHITECTURE.md rewritten for the two-transport reality:
  module map with cc2/connect, per-model ports, CC1-vs-CC2
  /status and SSE flow differences, and an honest test-architecture
  section (no automated live suite exists; live verification is manual
  CLI runs).