beffroi-tools/docs/canvas-pilot/device-check.md

11 KiB
Raw Blame History

Eikos connectivity and Preview check

Verified on 2026-11-13

User-authorized hardware check on an Eikos 4K, firmware 3.3.10. The unit reported this identity through GET /api/tpp/v1/system. Its WebRCS bundle was app.e42fce4879f318b28cde.js. The device address was supplied for this session; it is not a compiled-in application default.

  • HTTP read access and WebRCS WebSocket feedback worked without authentication on this configured device. Do not assume other installations permit this.
  • GET /api/stores/device returned the state tree used by WebRCS, including screen canvas dimensions, layer center positions/container sizes, source assignments, crop/mask fields and transition state.
  • The explicit REST read /api/tpp/v1/screens/2/live-layers/1/presets/preview reported input 4, matching the internal Preview bank. That REST response contained source/status, not the full geometry. Geometry was obtained from the WebRCS state endpoint.
  • Screen 2's transition state was AT_DOWN: Program was DOWN, Preview UP. WebRCS's own bank-selection code maps AT_UP the other way around. These are physical banks, not permanent Program/Preview identifiers.
  • A single Preview layer property was changed over the WebRCS WebSocket: screen 2, bank UP, layer 1, position.pp.posH, 1436 → 1437 → 1436. Readback confirmed the intermediate change and restoration. The complete tested layer matched its original state after restoration. All four screen Program preset trees matched before, during and after this test.
  • No TAKE, CUT, transition, Program, input/global crop, output configuration, source-routing or preset-memory command was sent. Raw snapshots remain temporary local artifacts and are not committed.

The message envelope observed and tested was channel: "DEVICE", with data.path an array of state-tree keys and data.value the new scalar value. The read path for the tested property was:

device/screenList/items/2/presetList/items/UP/liveLayerList/items/1/position/pp/posH

This is evidence of a working bidirectional link, not validation of either advanced-crop export strategy. Only horizontal center movement was exercised. The WebRCS internal endpoint/protocol is firmware-specific, rather than a promise of stable public REST geometry support. Manufacturer references are linked from the official Eikos downloads section. The linked API PDFs were not accessible during this check; protocol findings above come from the installed device and its own WebRCS bundle.

Read-only import bridge

Open Midra connection… in the main menu or Settings → Midra integration. Enter the device IP and click Connect. Then open the separate Import → Midra layers… window. Reads run in a worker thread; canceling discards late results. Choose a screen and choose a destination beside each device layer: Do not import, New layer, or an existing visible, unlocked editor layer. Click Import layers directly in this window; no prior selection or matching layer count is required. Duplicate or removed destinations are rejected without partial changes. The entire import, including new layers, is one undo step. Output resolutions must match. Reading never changes the document; only applying the reviewed import does. An optional Try autoconnecting at startup preference makes one background read attempt using the remembered address. No polling or automatic retries occur. The address is remembered in application settings; no credentials are stored. The footer shows ONLINE after a successful read (last-read status, not continuous monitoring), and OFFLINE when refreshing, after a failed read, or after editing the address.

The Rust adapter was tested against the same Eikos 3.3.10: two 1920×1080 screens; disabled zero-canvas screen slots are filtered. A CLI smoke check is available:

cargo run -p canvas-pilot --example read_midra -- DEVICE_IP

A standard-library Python helper remains available for snapshot capture/offline replay:

python3 tools/midra_probe.py --host DEVICE_IP --save-snapshot /tmp/eikos-state.json
python3 tools/midra_probe.py --host DEVICE_IP --screen 2 --layers 1 2 --svg /tmp/preview.svg
# Replay a saved snapshot without connecting:
python3 tools/midra_probe.py --snapshot /tmp/eikos-state.json --screen 2 --layers 1 2 --svg /tmp/replay.svg

Select the matching number of editor layers, choose Import → Geometry file…, and review mapping before applying. Match the editor output resolution to the device canvas; SVG geometry import does not replace the editor's screen settings. The helper exports container geometry only: center minus half-size becomes top-left x/y, preserving half pixels. Source aspect fitting can make the visible picture differ from this container. Routing, source resolution, source cropping, effects, borders, opacity and zoom are not imported. Nonzero layer crop or mask is refused rather than silently imported as final visible bounds. Device slots are explicitly selected; unused slots are not assumed to be available layers.

The helper has no device write operation, bypasses system proxies, refuses redirects, bounds reads to 32 MiB and never overwrites an existing output file. Snapshots may contain network/private configuration; keep them out of Git and sanitize before sharing. Its exchange contract is documented in interfaces.

Reviewed send and cropped-layer round trip

The user's later authorization permits Program changes and accepts the physical-bank TAKE race. This supersedes the earlier Preview-only restriction above.

Export → Configure Midra export… now supports explicit Preview or Program targeting, per-selected-layer destination mapping, output-sized or source-aspect containers, and exact before/after field review. Each send writes directly, then reads the device back; it does not reject changes made since the snapshot. A partial or uncertain result is reported without automatically repeating or rolling back writes. Local editing, importing, undo/redo, connection and project loading never send.

On Eikos 4K firmware 3.3.10, both sizing modes passed actual write/readback/restore checks on screen 2 Preview layer 1 (UP) and Program layer 2 (DOWN). Cases used a 301 × 251 visible box at 12.5, 34.5, 1920 × 1080 native input corrected to 1920 × 1000, 200% content zoom and offsets +30, −10. All nine fields (size, center, four crop edges, Centered aspect) matched readback; original values were restored after each case. A final snapshot confirmed all screen preset trees matched their pre-test state. The imported visible geometry matched within one pixel. Device-source routing, TAKE, input configuration, output configuration, masks and effects were not changed.

The installed WebRCS bundle maps crop edges linearly from 0–65535 to 0–100% or 0–contentWidth/contentHeight pixels. Its layer size limits are 0–65535 and center limits −67268–67268; this adapter requires positive dimensions. These findings support the device exchange contract.

These tests verify firmware acceptance, readback and mathematical import round trip. They do not replace visual comparison on a physical screen with a known test image; the tested device reported layer state OFF. Platform/display acceptance remains open.

The explicit hardware diagnostic (never part of automated CI) is:

cargo run -p canvas-pilot --example check_midra_send -- --apply-and-restore DEVICE_IP 2 1 preview
# Program writes are deliberate here:
cargo run -p canvas-pilot --example check_midra_send -- --apply-and-restore DEVICE_IP 2 2 program

Centered layers with valid content dimensions and no mask now import visible bounds and source-pixel crop framing. Their combined crop is stored in the local layer crop; the device does not reveal which part originated from this editor's input correction. Other aspect modes retain container-only import and refuse nonzero cropping. Shared input settings and source routing are not synchronized.

The manufacturer's REST API guide v3.0.13 was inspected. It documents logical Preview source assignment, but not geometry writes; this implementation therefore uses the installed WebRCS WebSocket protocol. The physical bank is captured in the review and remains fixed across the send.

Missing preset apply diagnosed

The operator reported that WebRCS geometry changed while the physical multiviewer Preview did not. The bundled WebRCS code calls z1() after geometry edits, which dispatches DEVICE_PRESET_CONTROL_SET_ONE for xUpdate false, then true. The device snapshot had device/preset/control/pp/enableAutoUpdate: false. Earlier tests checked only staging-store readback and therefore did not establish that geometry had reached video processing. The exporter now sends this apply pulse after geometry writes, even when no geometry fields differ. A direct pulse was sent to the connected Eikos to apply its already-staged layout. The operator confirmed that the physical multiviewer Preview then matched WebRCS, establishing the missing apply pulse as the cause. Layer status OFF is not sufficient evidence that the operator's layers are disabled, and must not be used as that diagnosis.

Input preview capture checked on 2026-11-14

The same WebRCS client exposes GET /api/device/snapshots/inputs/:id. A read-only request to active input 3 returned HTTP 200, PNG, 512 × 288 (591155 bytes). Input 1, with no signal, returned a blank 128 × 128 PNG. The input state reported snapshots enabled with size 512_512; the active image retained its 16:9 aspect. The response included Last-Modified and Cache-Control max-age=0. No continuous stream or write was needed. Raw input images/state stay temporary and are not committed.

The implementation fetches a still only on demand, limits concurrent requests to one and enforces a five-second minimum interval. It does not change snapshot enable/size, input routing, capture-library commands or any preset. See input preview interface.

Allocated layers versus firmware slots

The snapshot contains eight live-layer slots per screen, even when the active configuration allocates only one or two. The Rust adapter reads device/preconfig/status/stateList/items/CURRENT/screenList/items/{screen}/pp/layerCount and exposes only IDs 1 through that count for import, mapping and sending. The cached Eikos state reported one allocated layer for screen 1 and two for screen 2; other screen allocations were zero. This is independent of input routing, signal presence, OFF state and the number of physical inputs. Background and foreground are not live-layer destinations. Staged NEW configuration is ignored.