fix(ms-ai-architect): registry-herding Fase 1a — +5 sitemap-prefiks + skjemaløs URL-ekstraksjon

Update-mekanismen lot siterte URL-er stå som not_in_sitemap fordi deres docset
aldri ble pollet, og skjemaløse citater (tabellceller: bare learn.microsoft.com/...
uten https://) ble aldri ekstrahert.

- Taxonomy: +5 docsets (graph, ai-builder, power-apps, power-automate,
  microsoftsearch) — hver ett child-sitemap, verifisert live mot indeksen.
  Bevist: 21/24 not_in_sitemap-URL-er i disse docsetene blir nå tracked.
- url-normalize: extractUrls fanger skjemaløse citater-med-sti (krever /path
  etter domenet → avviser bare-domene-prosa og JSON-eksempler); normalizeUrl
  kanonikaliserer scheme til https. Bevist: +19 nye URL-er ekstraherbare.
- Bugfix: backtick fra inline-kode-citat (`learn.microsoft.com/...`) lekket inn
  i URL-en — ekskludert i regex + trailing-strip.

TDD: tests/kb-update/test-url-normalize.test.mjs (ny, 14) + test-taxonomy
prefiks-count 18→23 + Fase 1a-assertion. Full suite 338/338, 0 regresjon.
Registry-refresh (build-registry --merge + poll) bevisst utsatt — unngår
552KB re-order-churn; effekten er empirisk verifisert read-only.

scripts/kb-update/data/domain-taxonomy.json

@@ -3,7 +3,7 @@
   "description": "Single source of truth (lag 0) for 'in-domain' classification. Consolidates four previously-divergent taxonomies into one: (a) poll-sitemaps TARGET_PREFIXES, (b) discover-new-urls INCLUDE/EXCLUDE, (c) category->skill ownership, (d) report-changes getFilePriority. Scripts READ this file; they no longer embed their own copies.",
   "provenance": {
     "category_skill": "PHYSICAL DISK is canonical (skills/<skill>/references/<category>/). category-skill-map.json drifted on 4 entries (copilot-extensibility, monitoring-observability, performance-scalability, prompt-engineering) and is no longer consumed by code.",
-    "sitemap_prefixes": "poll-sitemaps superset (18). discover-new-urls previously had only 12 — now reads this list for poll parity.",
+    "sitemap_prefixes": "poll-sitemaps superset (23). discover-new-urls previously had only 12 — now reads this list for poll parity. Fase 1a (2026-06-26) added 5 stack-relevant docsets — graph, ai-builder, power-apps, power-automate, microsoftsearch — each a single child sitemap verified live against the index, reclaiming cited URLs previously stranded as not_in_sitemap.",
     "notes": "dotnet_en-us_ deliberately excluded: 75 sitemaps, only ~12 matches — not worth weekly polling.",
     "course_products": "Product slugs enumerated LIVE from the Learn Platform API (C3.3, server-side /api/v1/modules?products=<slug> probe). Doc slug 'azure-ai-foundry' is DEAD (0 hits) → Foundry/AI-services courses surface under 'azure-openai'. Slug gotchas confirmed live: 'fabric' (not 'microsoft-fabric'), 'entra' (not 'microsoft-entra-id'), 'azure-cosmos-db' (not 'cosmos-db'), 'azure-cognitive-search' (not 'azure-ai-search'). 'azure-kubernetes-service' is live but excluded (no AI category → noise). Maps slug->category; the owning skill is DERIVED from category_skill via makeCourseClassifier (never stored, cannot diverge)."
   },
@@ -25,7 +25,12 @@
     "microsoft-365_en-us_",
     "training_en-us_",
     "cloud-computing_en-us_",
-    "privacy_en-us_"
+    "privacy_en-us_",
+    "graph_en-us_",
+    "ai-builder_en-us_",
+    "power-apps_en-us_",
+    "power-automate_en-us_",
+    "microsoftsearch_en-us_"
   ],
   "category_skill": {
     "agent-orchestration": "ms-ai-engineering",

scripts/kb-update/lib/url-normalize.mjs

@@ -4,6 +4,7 @@
 /**
  * Normalize a learn.microsoft.com URL to a canonical form.
  * Rules applied in order:
+ *   0. Canonicalise scheme — accept schema-less + http(s) citations, store https
  *   1. Strip trailing punctuation leaked from markdown
  *   2. Strip fragment (#anchor)
  *   3. Strip ?view= and other query params
@@ -18,8 +19,16 @@ export function normalizeUrl(raw) {
 
   let url = raw;
 
+  // 0. Canonicalise scheme — citations appear schema-less (table cells: bare
+  //    learn.microsoft.com/...) or as http; strip any leading scheme and force
+  //    https so a schema-less citation matches the same key as its https form
+  //    (and the https sitemap entries) instead of creating a duplicate entry.
+  url = url.replace(/^https?:\/\//i, '');
+  url = 'https://' + url;
+
   // 1. Strip trailing punctuation that leaked from markdown context
-  url = url.replace(/[.,;:!?'")}\]]+$/, '');
+  //    (incl. backtick from inline-code citations: `learn.microsoft.com/...`)
+  url = url.replace(/[.,;:!?'")}\]`]+$/, '');
 
   // 2. Strip fragment
   const hashIdx = url.indexOf('#');
@@ -43,18 +52,23 @@ export function normalizeUrl(raw) {
 
 /**
  * Extract all learn.microsoft.com URLs from a text string.
- * Handles all 5 citation formats found in reference files:
+ * Handles all citation formats found in reference files:
  *   - Markdown links: [text](https://learn.microsoft.com/...)
  *   - Bare URLs on their own line
  *   - URL: prefix format
  *   - Dash-bullet format
  *   - Table cell format
+ *   - Schema-less citations: bare learn.microsoft.com/<path> (no https://)
+ * The scheme is optional but a path (`/...`) after the domain is REQUIRED — this
+ * captures schema-less citations-with-path while rejecting bare-domain prose
+ * mentions ("(learn.microsoft.com, blogs...)") and JSON example strings
+ * ({"url": "learn.microsoft.com"}), which carry no trackable source.
  * @param {string} text
  * @returns {string[]} array of normalized unique URLs
  */
 export function extractUrls(text) {
   if (!text) return [];
-  const regex = /https:\/\/learn\.microsoft\.com[^\s)"'<>\]|]+/g;
+  const regex = /(?:https?:\/\/)?learn\.microsoft\.com\/[^\s)"'<>\]|`]+/g;
   const seen = new Set();
   const results = [];
   let match;