Viewer coordinate and placement contract

The viewer interaction code uses the following authoritative units and state.

Mouse events provide top-origin logical window pixels. They are converted once to top-origin physical framebuffer pixels before projection or picking. The framebuffer size, rather than the logical client size, defines the OpenGL viewport.
Viewer2D world coordinates are metres. zoom is a dimensionless multiplier of the base scale of 25 framebuffer pixels per metre. The stored pan offsets are pixels at zoom 1; therefore their visible framebuffer translation is the stored offset multiplied by zoom.
viewer2d_coordinate_math owns the inverse world/framebuffer formulas for Top, Bottom, Front, and Side. The hidden coordinate is zero when unprojecting. Rendering uses the equivalent orthographic bounds, and overlays and pointer interaction call this utility rather than maintaining another formula.
The authoritative rendered screen bases are Top (+X, +Y), Bottom (-X, +Y), Front (+X, +Z), and Side (-Y, +Z). These signs follow the right vectors produced by the existing gluLookAt cameras. Projection, unprojection, measurement projection, and drag deltas share this basis.
Logical/framebuffer conversion uses the window content scale once and rounds to the nearest integer pixel. Non-finite, zero, negative, or overflowing conversions are invalid; callers retain pending alignment until valid framebuffer dimensions and coordinates are available.
Viewer3D world coordinates are metres. Pointer placement intersects the ray from the current camera matrices with the view plane through the active selection-drag anchor. Logical pointer coordinates are scaled to framebuffer coordinates before unprojection.
A continuous-placement anchor is the raw, unsnapped element origin in world metres. A selection-drag anchor is the effective selection centre in world metres. A Magnet preview is restored before either anchor is updated.
Both viewers maintain a viewport revision. Zoom, pan, orbit, view changes, fit/reset operations, camera interpolation, resize, and framebuffer changes invalidate pointer alignment. The next safe update restores any preview and aligns from the absolute current pointer under the new mapping; it never applies a pixel delta calculated with the old revision.
The revision state is the only alignment truth. Failed attempts do not mark the revision aligned. Paint resolves a pending revision before rendering when continuous placement is active and the canvas has received a valid pointer position.
Re-anchoring first validates projection prerequisites. It then restores the old Magnet preview, translates the raw anchor from an absolute pointer target, and recomputes the preview even for zero raw translation. Preview apply and inverse restoration use the same interactive transform policy; neither operation commits grouping or creates a view-only Undo entry.
Viewer3D derives the raw anchor by removing pending preview translation from its displayed drag anchor without mutating the scene. Ray intersection uses that raw point as its plane point; only a successful intersection begins the restore, absolute alignment, and fresh-preview transaction.
Confirmation preserves the snapped transform of the confirmed element but seeds its next clone from the raw anchor. The new clone remains pending until the normal absolute pointer-alignment path succeeds for the current revision.
Incremental placement uses failure-preserving framebuffer conversion for both pointer samples. Failed conversion or 3D projection retains the newest logical pointer, invalidates alignment, and performs no preview or scene transaction; the next safe update retries by absolute alignment.
All 3D placement geometry uses the raw anchor: pointer planes and projected constraint axes never use the displayed preview-adjusted depth.

The view-dependent hidden-axis Magnet weights are separate from projection and remain intentional: 2D ignores the hidden axis, while 3D progressively reduces the camera-aligned axis weight near an orthogonal view.