Tighten fuzzy edit-distance threshold for length-5 patterns

[CyrusNajmabadi]

Followup to #82457

Summary

Changes WordSimilarityChecker.GetThreshold from length <= 4 ? 1 : 2 to length <= 5 ? 1 : 2, so that 5-character patterns now require edit distance ≤ 1 instead of ≤ 2 for a fuzzy match.

Motivation

The original threshold tiers (≤4 → 1, ≥5 → 2) were chosen by me 10-15 years ago based on gut feeling and a small amount of manual testing. They work well in practice, but the jump to k=2 at length 5 creates a blind spot in the bigram pre-filter introduced in #82457: the q-gram count lemma gives min_shared = 5 - 1 - 2*2 = 0, meaning the bigram filter cannot eliminate any candidates for 5-character patterns.

The industry standard for this exact problem is Elasticsearch/Lucene's fuzziness: AUTO setting, which uses:

Pattern length	Max edit distance
1–2	0 (no fuzzy matching)
3–5	1
6+	2

This is grounded in Damerau's finding that approximately 80% of human misspellings are edit distance 1, and that allowing 2 edits on short strings produces too many false positives (e.g., two edits can transform a 5-character word into something completely unrelated).

What changes

Only the length-5 case is affected:

Pattern length	Before (k)	After (k)
1–2	N/A (MinFuzzyLength = 3)	N/A (unchanged)
3–4	1	1 (unchanged)
5	2	1
6+	2	2 (unchanged)

MinFuzzyLength = 3 is deliberately kept — below that length, fuzzy matching produces too many spurious hits regardless of threshold.

Effect on pre-filter

With the new threshold, the bigram pre-filter becomes effective at length 5. The min_shared column shows the minimum number of pattern bigrams that must appear in the candidate for the q-gram count lemma to permit a match (|pattern| - 1 - 2k). Higher values mean stronger filtering:

Pattern length	Before: k, min_shared	After: k, min_shared	Filtering strength
3	k=1, 0	k=1, 0	None (always passes)
4	k=1, 1 (of 3 bigrams)	k=1, 1 (of 3 bigrams)	Weak
5	k=2, 0	k=1, 2 (of 4 bigrams)	None -> Moderate
6	k=2, 1 (of 5 bigrams)	k=2, 1 (of 5 bigrams)	Weak
7	k=2, 2 (of 6 bigrams)	k=2, 2 (of 6 bigrams)	Moderate
8	k=2, 3 (of 7 bigrams)	k=2, 3 (of 7 bigrams)	Strong
10+	k=2, increasingly strong	k=2, increasingly strong	Very strong

Risk

Exceptionally low. This is a one-line behavioral change that tightens a fuzzy match heuristic for a single pattern length. Fuzzy matching is already a best-effort "did you mean?" feature — slightly fewer fuzzy matches at length 5 is the expected and desired outcome.