/**
 * v0.4 Site Classifier — runs BEFORE rules to gate which rule set applies.
 *
 * Today the engine runs every rule against every audited site regardless of
 * whether the site is actually programmatic-SEO. A 23-page marketing site
 * gets pSEO-targeted findings that don't apply; a 50,000-page directory
 * gets the same audit shape as a small blog. This pre-flight classifier
 * decides which "kind" of site we're auditing and emits a list of ruleIds
 * that the dispatcher will skip when site type is small-marketing/blog.
 *
 * Heuristics (v1, ships in v0.4):
 *   - Sitemap URL count: <50 = bias small/blog; ≥1000 = bias programmatic.
 *   - URL-pattern clustering: normalize numeric/slug segments; if top-3
 *     templates cover ≥60% of URLs → strong programmatic signal.
 *   - Framework signal: pass-through from dev-server detection.
 *
 * Deferred to v0.4.1+ (per spec §4.11):
 *   - DOM-skeleton hashing across pages.
 *   - Per-cluster classification (mixed pSEO + marketing).
 *   - Per-page applicability tagging on findings.
 */
export type SiteType = "programmatic-directory" | "small-marketing" | "blog" | "ecommerce" | "docs" | "unclear";
export type ClassificationSignal = {
    kind: "sitemap-url-count";
    value: number;
} | {
    kind: "url-pattern-cluster-coverage";
    topTemplate: string;
    pages: number;
    ratio: number;
} | {
    kind: "framework-detected";
    value: "nextjs" | "vite" | "astro" | "unknown";
}
/**
 * v0.5.3 — emitted when `applyDegenerationGuard` downgrades a
 * `small-marketing` or `blog` classification to `unclear` because the
 * corpus is degenerate (mostly thin / mostly identical titles). Surfacing
 * this in `signals` lets the UI explain why severity demotions didn't
 * apply.
 */
 | {
    kind: "degeneration-guard-tripped";
    reason: "median-thin" | "title-duplicate-heavy";
    value: number;
};
export interface SiteClassification {
    type: SiteType;
    /** 0–1 confidence. Below 0.6 → treat as `unclear` and run all rules. */
    confidence: number;
    /** Ordered observations that fed the classification. UI surfaces these. */
    signals: ClassificationSignal[];
    /**
     * RuleIds suppressed because of this classification. The dispatcher
     * checks this list before invoking each rule.
     *
     * Empty array when type is `programmatic-directory`, `ecommerce`, or
     * `unclear` — those run all rules.
     */
    suppressedRules: string[];
}
/** Rules suppressed for non-pSEO sites (small-marketing / blog). */
export declare const PSEO_ONLY_RULE_IDS: readonly string[];
/**
 * Normalize a pathname into a hashable template by replacing path segments
 * that look like values with type placeholders.
 *
 * Examples:
 *   /california/los-angeles/plumbers   → /:slug/:slug/:slug
 *   /blog/hello-world                  → /blog/:slug
 *   /post/12345                        → /post/:n
 *   /                                  → /
 */
export declare function normalizePathToTemplate(pathname: string): string;
/** Compute template-cluster ratios from a URL list. Returns top entries first. */
export declare function clusterUrlTemplates(urls: string[]): Array<{
    template: string;
    count: number;
    ratio: number;
}>;
export interface ClassifySiteInput {
    /** All discovered URLs (sitemap + crawl). Unfiltered. */
    urls: string[];
    /** Framework detected via dev-server response headers, if any. */
    framework?: "nextjs" | "vite" | "astro" | "unknown";
}
/**
 * Classify a site from its URL list + framework signal. Pure function.
 *
 * Contract: callers must pass the FULL discovered URL list (sitemap +
 * crawl), not the post-sample list. The classifier needs the raw size
 * signal to distinguish a 5000-page directory from a 25-page sample of one.
 */
export declare function classifySite(input: ClassifySiteInput): SiteClassification;
/**
 * v0.5.3 — corpus-quality guard against "small-marketing" / "blog"
 * classification masking degenerate sites. The `small-marketing` profile
 * demotes `spam/thin-content`, `aeo/citable-facts`, `aeo/freshness-signals`,
 * `spam/doorway-pattern` etc. to `info` to avoid false-positives on legit
 * 6-page marketing sites (linear.app etc). But a 6-page site with 0 unique
 * content per page (e.g. an un-translated language switcher pretending to be
 * a directory) trips the same shape and inherits the demotions, escaping
 * with grade B.
 *
 * This guard runs AFTER classification, with parsed-page stats. If the
 * corpus is degenerate (median word count < 50 OR ≥50% of pages share an
 * identical title), the classification is downgraded to `unclear` so the
 * demotion table doesn't apply — the natural rule severities then fire.
 *
 * Only `small-marketing` and `blog` are guarded. The other types either
 * already run all rules (`unclear`, `programmatic-directory`, `ecommerce`,
 * `docs`) or aren't reached by the small-corpus path.
 */
export declare function applyDegenerationGuard(classification: SiteClassification, corpusStats: {
    medianWordCount: number;
    identicalTitleRatio: number;
    pageCount: number;
}): SiteClassification;
/**
 * Compute the corpus stats `applyDegenerationGuard` consumes. Pulled out so
 * tests can pass a fixture stat block directly without constructing
 * `ParsedPage` instances.
 */
export declare function corpusStatsFromPages(pages: ReadonlyArray<{
    title: string;
    contentText: string;
}>): {
    medianWordCount: number;
    identicalTitleRatio: number;
    pageCount: number;
};
//# sourceMappingURL=site-classifier.d.ts.map