test+docs(nav): characterise cursor-anchored LB pan (KD-38) + rotate sky-grab guard

[diarmidmackenzie]

Follow-up to the recent navigation regression fixes (#1855 / #1859 / #1868). No runtime change — docs + one test only.

Why

Tracing #1867 (sky-grab pan lurch) back, the root is that LB pan is cursor-anchored: it grabs the world point under the cursor at mouse-down and keeps that point under the cursor for the whole drag (unlike legacy EditorControls, whose pan rate was cursor-independent). That model was carried over from the original controls but was never written down as a decision — so its consequences (rate scales with anchor depth; a sky grab needs a synthetic anchor on the cursor ray; a near-zero-distance anchor degenerates) only surfaced as the #1867 / TH-80 / TH-81 fixes. This PR states the model explicitly and closes a test-coverage gap on the rotate side.

Changes

• docs/02-key-decisions.md — new KD-38. Characterises cursor-anchored LB pan and, crucially, the pan-vs-rotate asymmetry: on a source === 'fallback' (sky) grab, pan uses the fallback point (why it must lie on the cursor ray, regression: cannot pan up while looking out at the horizon and grabbing top half of screen #1867), while Shift+LB rotate branches on source and orbits the screen-centre ground pivot (_mapModePivot's fallbackCentre), ignoring the point. So the two gestures degenerate differently at the same cursor position — a degenerate-cursor change must be checked against both.
• docs/04-glossary.md — a note on the LB "pan" terms flagging the cursor-anchored dependence, pointing to KD-38.
• test/.../ExperimentalControls.rotation.test.js — a regression guard for the rotate side of the sky grab: Shift+LB over sky orbits fallbackCentre (y ≈ 0), never the cursor-ray fallback point (y > 1). The pan side is already covered by the ExperimentalControls.lbPan sky-grab tests (regression: cannot pan up while looking out at the horizon and grabbing top half of screen #1867); this guards the branch that keeps rotate immune to the same fault.

Testing

• ExperimentalControls.rotation.test.js: 3/3 green (2 existing + the new guard).
• Break-it check: forcing rotate to consume the fallback point makes the new test red (pivot.y 42.7 vs the < 0.5 assertion), confirming the guard bites.
• No src changes, so lint is unaffected.

🤖 Generated with Claude Code