feat(phase3): learning helpers — summary, glossary, flashcards (SM-2), quizzes

Deterministic, on-device, no model:
- src/lib/learn pure modules (tokenize, summary [TextRank-ish], glossary
  [definition-pattern + frequency], flashcards [cloze/Q-A], srs [SM-2], quiz
  [MCQ with distractors]) — 37 unit tests.
- Flashcard persistence: Dexie v4 + native v4 `flashcards` table; create/list/
  listDue/updateSrs/delete/counts; cascades (transcript delete, course->Unsorted).
- UI: transcript "Study aids" (generate summary+glossary, click-to-seek; create
  flashcards), Study screen (SM-2 review + Anki CSV export), per-lecture Quiz,
  library Study link with due-count badge.

215 tests green, 0 tsc errors, web export builds.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/lib/learn/tokenize.ts

@@ -0,0 +1,127 @@
+// Text tokenization for the deterministic study helpers (summary, glossary,
+// flashcards, quiz). No model, no I/O, no platform deps — pure & testable.
+//
+// IMPORTANT: relative imports only inside src/lib (vitest has no '@/*' alias).
+
+/**
+ * A small, language-agnostic-leaning English stopword set. Kept intentionally
+ * compact: enough to stop "the/and/of" from dominating term frequency, without
+ * stripping domain words. Used by both summary scoring and glossary candidate
+ * selection.
+ */
+export const STOPWORDS: ReadonlySet<string> = new Set([
+  'the',
+  'and',
+  'for',
+  'are',
+  'but',
+  'not',
+  'you',
+  'all',
+  'any',
+  'can',
+  'had',
+  'her',
+  'was',
+  'one',
+  'our',
+  'out',
+  'has',
+  'his',
+  'how',
+  'its',
+  'may',
+  'new',
+  'now',
+  'old',
+  'see',
+  'two',
+  'way',
+  'who',
+  'did',
+  'get',
+  'him',
+  'use',
+  'this',
+  'that',
+  'with',
+  'from',
+  'they',
+  'will',
+  'what',
+  'when',
+  'were',
+  'been',
+  'have',
+  'into',
+  'than',
+  'them',
+  'then',
+  'some',
+  'such',
+  'also',
+  'just',
+  'like',
+  'over',
+  'only',
+  'most',
+  'much',
+  'very',
+  'each',
+  'because',
+  'about',
+  'which',
+  'their',
+  'there',
+  'these',
+  'those',
+  'would',
+  'could',
+  'should',
+  'where',
+  'while',
+  'being',
+  'between',
+  'here',
+  'does',
+  'doing',
+  'done',
+  'made',
+  'make',
+  'used',
+  'using',
+  'both',
+  'more',
+  'less',
+]);
+
+/**
+ * Split text into sentences on sentence-ending punctuation (`.`, `?`, `!`) and
+ * newlines. Each result is trimmed; empties are dropped.
+ *
+ * This is intentionally simple (no abbreviation handling) — lecture transcripts
+ * are conversational and rarely contain "Dr." / "e.g." style edge cases, and a
+ * mis-split sentence only marginally affects summary/glossary scoring.
+ */
+export function splitSentences(text: string): string[] {
+  return text
+    // Break after any run of .?! (keep the run with the sentence it ends), and
+    // also break on newlines so transcript line breaks become boundaries.
+    .split(/(?<=[.?!])\s+|[\r\n]+/)
+    .map((s) => s.trim())
+    .filter((s) => s.length > 0);
+}
+
+/**
+ * Tokenize text into "content words": lowercase, split on non-word runs,
+ * dropping empties, stopwords, and very short tokens (length < 3).
+ *
+ * Used for term-frequency scoring; short tokens and stopwords carry little
+ * topical signal and would otherwise dominate the frequency counts.
+ */
+export function tokenizeWords(text: string): string[] {
+  return text
+    .toLowerCase()
+    .split(/\W+/)
+    .filter((t) => t.length >= 3 && !STOPWORDS.has(t));
+}