fix: the crow answers its controls, and a night starts on the strand
The browser found it: a live inputMap can be answering the input layer while
the handlers this component put on it are gone. `inputSim fired` showed hop 19
times on the crow map and CF.state.hops never moved; a freshly-registered
onPressed on the SAME map fired 2 of 2 in the same session. The crow could not
be moved at all — not sidled, not hopped, not paused — and every accessor read
healthy, so the round simply starved. Three silent seconds is cheap to
disbelieve, so the climb now re-takes its own controls, a bounded number of
times per round so an idle player never churns the input layer.
The spawn was read off the crow's own transform at awake, so a reload mid-climb
baked wherever it happened to be standing as the start of every round after —
and a crow reloaded over the water restarted in the sea. It is the strand now,
from the layout, which is the same on every peer and after every reload.
Verified after the change: a keyboard round to the roost at 1080x1440 (30 hops,
9 eaten, 3 crumbled, 59 s, the flock lifting on film) and a keyboard round to
the roost in the WASM build in a browser (40 hops, 10 eaten, 51 s).