W-432: drop component-group modifiers from the emoji corpus

[wr]

What

Bare Fitzpatrick skin-tone swatches (🏻–🏿, e.g. :medium_skin_tone:) and hair modifiers (🦰🦱🦳🦲) surfaced as standalone search hits. They're emojibase group 2 ("component") — combining modifiers, useless inserted alone. The browser already skips this group, so they only leaked through live search.

Fix

Filter the component group out of the loaded corpus in EmojiDatabase.load, so it's absent from search and exact-match — matching the browser. SkinTone applies modifiers via hardcoded scalars, not the DB, so skin-tone application is unaffected.

Notes

• Stacked on #(W-431) — base will retarget to main once that merges. Review only the top commit.
• Removes the 4 hair components from search too (same rationale); easy to narrow to skin-tones-only if we want hair kept.

Test plan

• New componentModifiersAreExcluded test: no group-2 entries in the corpus, 🏽 gone from byHexcode and from a :medium search.
• Full suite green; verified in the dev app.

Refs W-432

[linear-code]

W-432

[wr-claude-reviewer]

Clean fix. The filter is placed at exactly the right point (right after decode, before self.all is assigned), so the component emojis can never leak into byHexcode, the indexed corpus, or search results. The componentGroup constant being static makes it accessible from the test file without needing a test-only accessor.

Test coverage is solid: the allSatisfy guard checks no group-2 survivors exist, the byHexcode assertion is the reported-case regression test, and the FuzzyMatcher.search assertion closes the loop on the original symptom. Nothing to change.

Note: the ssot CI run was still in progress at review time — no failures observed.