Two crossings branches were cut from the same parent and merged cleanly
into a tree that does not compile its test targets. #273 raised the boom
so its bands stack in y from BOOM_Y upward; #274 split
`CrossingKind::spec()` from a 3-tuple into a 2-tuple plus `boom_span()`.
Each was green on its own, `cargo check` does not compile `cfg(test)`
code, and both new tests live in `#[cfg(test)] mod tests` — so nothing
in the pipeline caught it before main.

The two type errors are the smaller half. `across()` picked the boom tip
as the minimum x among vertices between y 0.5 and 1.5, which describes a
bar lying flat across the road; with the bar standing up, every band
sits at a constant x beside the mast and the lowest one's bottom face is
at BOOM_Y itself, so the helper read a point a few centimetres inboard
of the mast rather than a reach. Four assertions were measuring that.

A raised bar's reach is a length, so the helper now reads the boom
bands' top above BOOM_Y. The bands are picked out by the bar's inner
face being the most inboard geometry an installation has — 0.37 m in
from the mast centre against the lamp head's 0.30 m — which also gives
the no-boom answer for free. The four tests restate the numbers #274
pinned as bar lengths: 6.0 m on an 11 m road, 1.6 m on a cycleway,
11.5 m for a full barrier, and 4.0 / 2.1 / 7.5 / none on the fallback.

No production code changes; both landed behaviours are wanted. The note
said a half barrier's boom "reaches back so its tip meets the
centreline" two paragraphs before saying the booms stand raised — it
now states a length instead.

Refs #276