Hyphen-dropped compound modifiers where a component already carries a WordNet sense (#43 remainder)

Part of #43, split off rather than guessed at.

#43's scan found **549 confirmed instances** of this corpus having two
separate words with an ordinary space (\`hard surface\`) where Brown's real
text has one hyphenated compound (\`hard-surface\`), verified via a
context-window word search against \`external/brown_nolines.txt\` (the same
reference \`semcor-compare-brown-nolines\` uses).

Only **1** of those 549 has no existing sense on either component, and is
fixed directly (\`data/learned/br-j04.yaml\`, \`spin spin\` -> \`spin-spin\`).
The remaining **548** turn out to be the dominant shape, not the
exception assumed going in (by analogy with #24's \`NUMBER-word\` case,
where the number's sense is always safely-droppable generic
cardinal/quantity): SemCor densely sense-tags content words, so two
adjacent words forming a compound modifier almost always *both* already
carry their own real, specific WordNet sense.

- **530 instances**: both components individually sensed, e.g.
\`data/belles_lettres/br-g01.yaml\`'s \`Yankee hatred\` -> Brown's
\`Yankee-hatred\` (\`yankee%5:00:00:northern:02\` and \`hatred%1:12:00::\`
both already present); \`entropy-increasing\`, \`old-school\`,
\`rose-tea\`, ...
- **18 instances**: exactly one component sensed, e.g.
\`half-reached\`, \`half-straightened\`, \`hundred-and-eighty-degree\`,
\`smashed-out\`, \`half-past\`, ...

Merging either shape into one token means deciding what happens to two
WordNet senses that can't both survive on a single token index -- a real
content trade-off (which sense, if either, still describes the merged
compound), not a mechanical fix. Needs individual editorial review per
instance (does the head's sense still validly describe the whole
compound? does the modifier's sense have to be dropped entirely, or does
the merged form deserve its own, different sense?) before any automatic
rule could be applied, the same way #24 needed one for its 13
word-prefixed exclusions.

### Suggested next step

Individual review of a sample from both buckets (start with the 18
one-sensed instances -- smaller, and a "drop the unsensed side, keep the
sensed side" rule may turn out safe there) to see whether a defensible
general rule emerges, or whether this needs case-by-case annotation
decisions.

Start with the 18 one-sensed instances, including data/learned/br-j04.yaml, and consult external/brown_nolines.txt with the semcor-compare-brown-nolines reference. Review a sample from both buckets and determine whether a defensible general rule exists or whether each compound needs an editorial decision. Done means the affected instances have documented annotation decisions or a justified rule.

Written by the indexing model from the issue text.