docs: document the tower and why coplanar crests flip the pose; add HUD reading guide

docs/PAGES.md

@@ -26,7 +26,7 @@ Open in `EXHIBIT_MODE=dev`; password-protected in `EXHIBIT_MODE=final`
 | `/admin/characters.html` | Characters | Per-ghost visuals: model, part placement, gradient colours, opacity, height, scale, face/torso textures. |
 | `/admin/landing.html` | Landing Page | Visitor landing screen: title, logo, background, accent colour, about text, disclaimer. |
 | `/admin/preview.html` | Virtual Viewport | Live orbit view of whatever ghosts are currently active, straight off the same WebSocket feed as the exhibit. Ground truth when the AR view disagrees. |
-| `/admin/testsheet.html` | A3 Test Rig | Printable A3 calibration/evaluation sheet, and a one-click load of the matching scene + playlist. See below. |
+| `/admin/testsheet.html` | A3 Test Rig | Two printable A3 pages — calibration sheet plus a foldable marker tower — and a one-click load of the matching scene + playlist. See below. |
 | `/admin/plan.html` | Placement Plan | Measured plan for physically positioning crests and models on the real table. |
 | `/admin/print.html` | Print Crests | Generates printable ArUco markers ("Newbury Crests") at the configured `sizeMM`. |
 | `/admin/calibrate.html` | Calibrate | Camera/marker calibration helper. Detects crests **without** filtering against the scene, so it's the quickest way to prove a printed crest is readable at all. |
@@ -36,28 +36,45 @@ Open in `EXHIBIT_MODE=dev`; password-protected in `EXHIBIT_MODE=final`
 ## The A3 test rig
 
 `/admin/testsheet.html` is a self-contained test environment for evaluating
-tracking, occlusion and lighting without the LEGO town in the way.
+tracking, occlusion and lighting without the LEGO town in the way. Two pages,
+both A3 landscape.
 
-- **Print A3 landscape at 100% ("Actual size"), never "Fit to page."** Check the
-  100 mm scale bar first: if it doesn't measure 100 mm, every crest is the wrong
-  size and every pose is scaled with it.
-- The sheet *is* the table — A3 landscape is exactly 42.0 × 29.7 cm. Origin is
-  the front-left corner, matching the exhibit's world frame.
-- **6 crests** at 60 mm, spread 38.7 cm corner to corner (the spread is what
-  makes the joint solve well-conditioned).
-- **Cylinder occluder**: stand an 85 mm ⌀ × 160 mm tall object on the dashed
-  circle at centre.
-- **Static ghost** at (21, 22) cm, deliberately *behind* the cylinder from the
-  front — walk around it and the ghost should disappear behind the post.
-- **Walking ghost** on an 18 × 12 cm rectangular loop around the cylinder;
-  60 cm perimeter, ~40 s per lap at the default 1.5 cm/s.
-- "Load this layout onto the server" writes both the scene and the playlist,
-  built from the same constants that drew the sheet, so print and config can
-  never drift apart. It **replaces** the live scene and playlist — save your
-  current layout first.
+**Print at 100% ("Actual size"), never "Fit to page."** Check the 100 mm scale
+bar on page 1 first: if it doesn't measure 100 mm, every crest is the wrong size
+and every pose is scaled with it.
 
-Known gap: the layout editor doesn't understand cylinder occluders, so opening
-and saving this layout there will drop the cylinder.
+**Page 1 — the ground sheet.** The sheet *is* the table: A3 landscape is exactly
+42.0 × 29.7 cm, origin at the front-left corner, matching the exhibit's world
+frame. Six 60 mm crests (ids 0–5) spread 38.7 cm corner to corner. Marked on it:
+the tower footprint at centre, the walker's loop, and the static ghost's spot.
+
+**Page 2 — the tower.** Folds into an 85 mm square × 160 mm tall prism with a
+60 mm crest on each of its four sides (ids 6–9, centred 100 mm up) plus one on
+top (id 10). Fold printed side OUT, tape the glue tab, fold the top flap over,
+and stand it on the dashed square with the FRONT panel toward the front edge.
+
+Why the tower exists, and why it isn't optional: **crests lying flat all share
+one plane, and a coplanar point set has two poses that reproject almost equally
+well** — the true one and a mirrored twin. With few crests in frame the solver
+intermittently returns the twin, which looks like ghosts snapping to the wrong
+side of the table and vanishing. Reprojection error can't tell them apart (the
+twin fits the pixels). A crest 100 mm above the sheet makes the geometry properly
+3D, and no mirrored pose can fit it. `cv-worker.js` also gates poses on physics
+(camera must be above the sheet, can't teleport) but that only suppresses the
+symptom — the tower is the cure.
+
+The tower doubles as the occluder. The **static ghost** sits at (21, 22) cm,
+deliberately *behind* it from the front: walk around and the ghost should
+disappear behind the tower. The **walking ghost** runs an 18 × 12 cm loop around
+it, 60 cm perimeter, ~40 s per lap at the default 1.5 cm/s.
+
+"Load this layout onto the server" writes both the scene and the playlist, built
+from the same constants that drew the sheets, so print and config can never
+drift apart. It **replaces** the live scene and playlist — save your current
+layout first.
+
+Known gap: the layout editor doesn't understand cylinder occluders or wall-mounted
+anchors well, so opening and saving this layout there may mangle the tower.
 
 ## Useful API endpoints
 
@@ -77,18 +94,38 @@ the rest need auth in final mode.
 | `/api/client-errors` | Reported client errors (auth in final mode) |
 | `/ws` | WebSocket: `scene`, `active`, `spawn`, `despawn`, `characters`, `pong` |
 
+## Reading the dev HUD
+
+```
+ws open  rx active,pong          WebSocket state + recent message types
+scene 11 anchors  ghost recs 2   scene loaded, ghosts held client-side
+60 fps  detect 20Hz@960  fov 58  render rate, detect rate/resolution, camera fov
+  REJ:jump                       (only when the solver refused a pose)
+eng opencv  moving               engine; FROZEN = ghost motion paused
+axis std  cv:ready               engine mode + OpenCV worker state
+markers 3/4 seen [0,4,6] +1dup   solved/raw crests, ids, duplicate decodes
+raw / view / up / pos            solved camera pose
+m0 px(0.02,-0.05) proj(0.03,-0.04)   AIM CHECK — see below
+ghost ndc(-0.10,0.20) ON-SCREEN  first ghost's screen position
+```
+
+**Aim check.** `px` is where the crest actually sits on screen (`0,0` = centre);
+`proj` is where the solved pose projects it. Point straight at a crest so `px` is
+near `(0,0)`: if `proj` matches, the camera model is right. A sign flip means a
+mirrored axis; a constant ratio means the projection doesn't match the lens (tune
+`localStorage.nbx.hfovDeg`); an intermittent horizontal flip with correct vertical
+is the coplanar ambiguity — stand the tower up.
+
 ## Gotchas worth remembering
 
 **Detection filters against the active scene.** A crest that isn't an anchor in
 the current layout is detected and then discarded, which used to look exactly
-like "the camera can't see anything".
+like "the camera can't see anything". The scene is fetched over HTTP at start so
+a flapping WebSocket can't break tracking, and the HUD distinguishes the cases.
 
-- The scene is fetched over HTTP at start, so a down or flapping WebSocket no
-  longer breaks tracking.
-- The HUD reports `scene N anchors` (or `NOT LOADED`) and `markers 1/2 seen
-  [0,7]` — usable over raw — so a filtered-out crest is visible at a glance.
-- The tracking chip says "Crest N not in this layout" rather than "Point at a
-  Newbury Crest" when this happens.
+**The video stream may be portrait.** iOS often hands back 720×1280, so any FOV
+assumption must apply to the image's LONG axis, not `videoWidth`. Getting that
+wrong made the focal length 1.78× too short.
 
 **Motion amounts live on the server.** `playlist.js` sends explicit `bobAmp`,
 `radius`, `speed` values with every spawn, so the client's defaults never apply