feat: improve stats dashboard and annotation settings

changes/2026-03-15-known-words-config-section.md

@@ -0,0 +1,5 @@
+type: changed
+area: anki
+
+- Changed known-word cache settings to live under `ankiConnect.knownWords` instead of mixing them into `ankiConnect.nPlusOne`.
+- Kept legacy `ankiConnect.nPlusOne` known-word keys and older `ankiConnect.behavior.nPlusOne*` keys as deprecated compatibility fallbacks.

changes/2026-03-15-session-delete-from-sessions-tab.md

@@ -0,0 +1,4 @@
+type: changed
+area: stats
+
+- Added session deletion to the Sessions tab with the same confirmation prompt used by anime episode/session deletes, and removed all associated session rows from the stats database.

changes/2026-03-15-stats-overlay-port-fix.md

@@ -0,0 +1,4 @@
+type: fixed
+area: stats
+
+- Fixed the in-app stats overlay so it connects to the configured `stats.serverPort` instead of falling back to the default port.

changes/2026-03-15-surface-frequency-fallback.md

@@ -0,0 +1,9 @@
+type: fixed
+area: overlay
+
+- Fixed subtitle frequency tagging for merged lookup-backed tokens like `陰に` by falling back to exact surface-form Yomitan frequencies when the normalized headword lookup misses.
+- Fixed MeCab merged-token position mapping across line breaks so merged content-plus-particle tokens like `陰に` keep their matched Yomitan frequency instead of inheriting shifted POS tags.
+- Fixed grouped frequency parsing in both Yomitan and fallback frequency-dictionary lookups so display values like `118,121` use the leading rank instead of collapsing the rank and occurrence count into `118121`.
+- Fixed frequency-rank ingestion to ignore Yomitan dictionaries explicitly marked `occurrence-based`, so raw occurrence counts are no longer treated as subtitle rank values.
+- Fixed inflected headword frequency tagging to prefer ranks from the selected Yomitan `termsFind` popup entry itself, ordered by configured dictionary priority, so forms like `潜み` use primary-dictionary ranks like `4073` before falling back to lower-priority raw lemma metadata such as `CC100`.
+- Fixed annotation-stage frequency filtering so exact kanji noun tokens like `者` keep their matched rank even when MeCab labels them `名詞/非自立`, instead of dropping the highlight after scan-time frequency lookup succeeds.