Name keys as terms and signatures, and drop "active" from the threshold

[drametwally]

Second of the two terminology PRs, a sibling of #26 off main. The two touch disjoint lines and can merge in either order.

Revised after review. The first version defined term, element, and key as three parallel words. That was wrong: key is the collective name for what an inverted list can be keyed by, which is a term or a signature, and the code has no element concept at all. The only element identifiers in the repo are BoundedSizeMaxHeap's, the ordinary Java collection sense; every sequence-domain use is prose in comments. A sequence's ordered items live in the same terms array a sparse record's do, so to the code they are terms.

So the section now says two things instead of three: every record holds terms whatever its shape, and a key is a term or a signature and therefore not always something the record itself holds. It also notes that sequence prose says "element" where order is the point, as a reading aid rather than a third kind of thing.

"active similarity threshold" matches nothing

Three occurrences, wrong twice over. The code calls this minSimilarity, and separately this document gives active a specific and different meaning: the one cache accepting writes, as opposed to the graduating caches, which are still searched and still serve results. Reusing it for a threshold spends a word that already has a job.

Two occurrences now name minSimilarity. The third is on a line #26 rewrites, so I left it there rather than stack the branches — after the earlier detached-base incident I am keeping these as siblings. A reviewer grepping this branch will still find that one occurrence; it goes away with #26.

The paper-term item dissolved

Every term I had queued turned out to be one the code already names, so the authority rule — code wins for anything the code names, paper wins only for what it doesn't — rejected all of them:

Queued rename	Why it is not happening
→ comparison threshold	Code says minSimilarity, 224 occurrences
→ popularity cutoff	Code says maxNumRowsPerTerm, plus popularTerms / popularTermDiscardScope
→ record frequency	The quantity max_fraction_ids_per_term is already stated against
→ alphabet	Already in use in DESIGN.md; nothing to do

Verification

• Docs only, no code touched. 547 tests pass, 3 pre-existing OpenBLAS assumption aborts.
• The wrapped occurrence at old line 350/351 was invisible to a line-based grep; found with rg -U --multiline-dotall.