// Scoring-text construction: the document fed to the vectorizer.
//
// fluxrec is title-only by design (no body/excerpt ever enters the scoring
// text). On top of the raw title we inject one synthetic feed-identity token
// `feed-<slug>` built from the entry's feed_title, so LR can learn a per-feed
// prior inside the existing TF-IDF + LR model. MinDF shrinks low-data feeds
// toward zero for free (a feed with <2 labeled entries never enters the
// vocabulary), so no separate shrinkage-prior machinery.
//
// Hyphens, not underscores: the tokenizer keeps [a-zA-Z0-9-] and splits on
// everything else, so `feed-marginal-revolution` survives as ONE token while
// `feed_marginal_revolution` would fragment into three. Identical slug
// construction at train (from labels.jsonl feed_title) and serve (from the
// entry's feed.title) keeps the feature consistent across both sides.
package core

import (
	"regexp"
	"strings"
)

var slugSep = regexp.MustCompile(`[^a-z0-9]+`)

// FeedSlug normalizes a feed title into a stable slug of lowercase
// letters/digits joined by single hyphens, trimmed. Empty/whitespace input
// returns "". Idempotent: slug(slug(x)) == slug(x).
func FeedSlug(feedTitle string) string {
	s := strings.ToLower(strings.TrimSpace(feedTitle))
	s = slugSep.ReplaceAllString(s, "-")
	return strings.Trim(s, "-")
}

// ScoringText builds the document fed to the vectorizer: the entry title plus
// one synthetic token "feed-<slug>". If the feed title is empty (no slug), the
// token is omitted and the title alone is returned. No body/excerpt is ever
// added — title-only is the design.
func ScoringText(title, feedTitle string) string {
	slug := FeedSlug(feedTitle)
	if slug == "" {
		return title
	}
	return title + " feed-" + slug
}