style: drop the explanatory comments added during review

Author: kfaracik

__tests__/hybridRetrieval.test.ts

@@ -587,8 +587,6 @@ describe('hybridRetrieve', () => {
   });
 
   it('covers a de-diacriticised document from a diacriticised query, as FTS would', async () => {
-    // Similarity sits below the semantic threshold and the top-keep floor, and
-    // there is no keyword hit, so this chunk qualifies on term coverage alone.
     const vectorResults = [
       {
         id: '1:0',

database/keywordIndex.ts

@@ -36,9 +36,6 @@ export const ensureKeywordIndex = async (db: DB): Promise<boolean> => {
   return ftsAvailable;
 };
 
-// Indexes every chunk the index is missing, not just the initial empty case: a
-// single swallowed insert (see addChunkToKeywordIndex) would otherwise leave a
-// chunk searchable by vector but never by keyword, with nothing to repair it.
 const backfillKeywordIndex = async (db: DB): Promise<void> => {
   let rows: Record<string, Scalar>[];
   try {

utils/hybridRetrieval.ts

@@ -248,8 +248,6 @@ export const hybridRetrieve = async ({
   }
 
   const terms = extractQueryTerms(prompt);
-  // Folded the same way the keyword index folds, so coverage agrees with what
-  // FTS actually matched instead of missing every de-diacriticised document.
   const coverageTerms = new Set(
     [...terms].map((term) => stemPrefix(foldForMatching(term)))
   );

utils/queryTerms.ts

@@ -131,10 +131,5 @@ const FOLD_MAP: Record<string, string> = {
   ż: 'z',
 };
 
-// The keyword index compares text after two foldings: ł by hand (unicode61's
-// remove_diacritics leaves that stroke letter alone) and the tokenizer's own
-// diacritic stripping. Layers that match in JS rather than through SQLite have
-// to fold the same way, or "płatność" stops matching "platnosc" on one side
-// only. Explicit map instead of NFD normalize, which Hermes does not guarantee.
 export const foldForMatching = (text: string): string =>
   text.toLowerCase().replace(/[ąćęłńóśźż]/g, (char) => FOLD_MAP[char] ?? char);

utils/rankFusion.ts

@@ -31,10 +31,7 @@ export const reciprocalRankFusion = (
 };
 
 // Cosine similarity; normalises internally (LFM2.5 embeddings are non-unit-length).
-// Returns 0 when either vector is empty or zero-length, and when the two differ
-// in length — mismatched dimensions mean the vectors came from different
-// embedding models, and truncating to the shorter one would answer with a
-// plausible-looking number instead of surfacing that.
+// Returns 0 when either vector is empty or zero-length, or when they differ in length.
 export const cosineSimilarity = (a: number[], b: number[]): number => {
   const len = a.length;
   if (len === 0 || b.length !== len) return 0;
@@ -137,9 +134,6 @@ export const maximalMarginalRelevance = (
 };
 
 // Keep leading items until relevance drops below `dropRatio × previous`; at least `minKeep`.
-// Scores must be non-negative — the cut is a ratio, which says nothing about
-// negative scales such as raw BM25. A zero score keeps everything after it,
-// since nothing below zero can follow in a descending non-negative list.
 export const adaptiveKeepCount = (
   sortedScoresDesc: number[],
   minKeep = 1,