Repeated pattern letters are sometimes free: "aaaaaaaaaab" scores the same as an exact "ab" and outranks "abx"

[matt-edmondson]

What's wrong

In CalculateScoreCore (FuzzySearch/Fuzzy.cs ~223-260), a subject character that re-matches the current best letter (rematch) is charged unmatchedLetterPenalty only when its newScore >= bestLetterScore. When it scores lower, no branch charges it anything. It still sets prevMatched = true, so the character after it also gets the adjacency bonus.

Whether a repeated letter costs anything therefore depends on the bonus earned by the letter it repeats. A first letter with the separator bonus (10) makes every later repeat free, because an adjacent repeat only ever scores 5.

Measured scores for pattern "ab" (current main)

Subject	Score
ab	15
aab	15
aaaaaaaaaab	15
AAAAAAAAb	15
abx	14
abbbbbbbbbb	6
a_b	19
a_bbbbbbbbbb	19

• aaaaaaaaaab ties the exact match and outranks abx, even though it has nine extra characters.
• Repeated bs after ab cost 1 each, but the same repeats after a_b cost nothing.

Why it matters

The score is used for ranking, and the documented rule is that unmatched characters are penalized. Noise made of repeated letters can rank a candidate at or above an exact match, which misorders filter results. Earlier gist-derived scoring defects were fixed in #66 and #82; this one was carried over from the same original gist.

Suggested fix / acceptance criteria

Charge the penalty when a rematch loses as well:

if (newScore >= bestLetterScore)
{
    if (bestLetterIdx is not null) score += unmatchedLetterPenalty;
    ...
}
else
{
    score += unmatchedLetterPenalty;
}

This was tried on a local build. aab scores 14, aaaaaaaaaab scores 6 and a_bbbbbbbbbb scores 10; the other rows are unchanged, and all 51 existing tests still pass.

• Add regression tests asserting Score("ab","ab") > Score("aab","ab"), Score("abx","ab") > Score("aaaaaaaaaab","ab"), and Score("a_b","ab") > Score("a_bbbbbbbbbb","ab").

[matt-edmondson]

Triage

• Category: Bug
• Priority: Medium. Candidates are ranked wrongly (noise can tie or beat an exact match), which affects every consumer's filter ordering, but nothing crashes. The fix is small and has already been validated locally against the existing 51 tests.
• Area / suggested owner: CalculateScoreCore rematch branch in Fuzzy.cs.
• Related: Prefix-penalty scoring is a permanent no-op, and its test can't catch that #66 and Documented prefix-penalty cap doesn't actually cap anything — unbounded score drop for long prefixes #82 fixed earlier scoring defects that came from the same original gist. Case-insensitive matching fails for Greek final sigma and micro sign: Contains("ΟΔΟΣ", "οδος") is false #88 is open in this repo but is a separate issue (case-folding).
• Duplicates / in progress: None found. No open PR covers this.
• Next step: Apply the else { score += unmatchedLetterPenalty; } branch and add the three ordering regression tests from the issue. Downstream consumers (TextFilter, and Frontmatter via #113) may see small ranking changes.

---

Generated by Claude Code