fix: keep jev_classify ids collision-free and tally on a null prototype

[kittimzhe]

What

Two jev_classify defects reported in the "found along the way" section of #34:

1. Fallback id collisions. Generated fallback ids (item0/class0 style) were never checked against supplied ids, so an item omitting its id took the fallback item0 while a later item explicitly named item0 kept it too — duplicated result ids. For classes the same collision also collapsed probabilities keys, silently overwriting one class's distribution with another's. Duplicate supplied ids were already rejected; the gap was fallbacks vs supplied.
2. Plain-object tally. by_class was a plain object literal, so a caller-supplied class id like __proto__ was swallowed by the prototype setter and never counted, while constructor read the inherited function and produced a garbage string.

How

• Apply the two-phase pattern jev_rerank already uses for candidate ids: collect supplied ids first (still rejecting duplicates), then let every generated fallback avoid any supplied or already-used id.
• Tally by_class on Object.create(null), matching the criteria map already built that way in the same tool.

No tool arguments, results, or README examples change.

Verification

• npm run build ✓ · npm run typecheck ✓
• npm test — 227/227 offline (baseline 224 + 3 new regression tests)
• The 3 new tests fail on the unfixed src/ (224 pass / 3 fail), confirming they catch both defects: jev_classify fallback item ids never collide with explicit ones, jev_classify fallback class ids never collide with explicit ones, jev_classify tallies __proto__ and constructor class ids as plain keys.

Both issues were reported in #34; this PR fixes only the jev_classify items and leaves the rest of that list untouched.

[jkudish]

• Thank you for this!
• Both fixes are exactly right — the two-phase id assignment jev_rerank uses, and the null-prototype tally.
• One fix on merge: the third test now actually exercises the constructor tally.
• Shipping in the next patch release.

[kittimzhe]

Thanks for the quick review and merge! And good call strengthening the third test — exercising the constructor tally directly makes it a much better regression guard than what I had. Glad this made it into 0.10.0.