• Pre-compiled phrase and obfuscation regexes in ProfanityDetector constructor
  (eliminates per-call RegExp construction in detect())
• Pre-compiled single SUBSTITUTION_REGEX at module level for normalizeText()
• Added O(1) cue-by-ID Map in CueIndex (replaces per-frame O(n) Array.find)
• Added binary search to CueIndex.getNextCues() (O(log n) vs O(n))
• Added CueIndex.getProfanityCueCount(), replaced cues.filter() at hot paths
• Skip RAF loop work when video is paused (keep loop alive for resume)
• Replace unconditional cues.sort() with isSorted check (merge-aware)
• Removed dead normalizedText variable in detect()
• Removed redundant inline sort in PlutoTV accumulation path

• Type-annotate regex match variables as RegExpExecArray | null
• Re-export ProfanityWindow type for content script import
• Add tabId guards in background message handler
• Type-annotate DNR rules as browser.declarativeNetRequest.Rule[]
• Cast TextTrackCue to VTTCue for .text property access
• Cast nativeOpen.apply() through unknown to XMLHttpRequest
• Type-annotate .text() callback param in fandango
• Cast querySelector('track') to HTMLTrackElement
• Wrap querySelectorAll results in Array.from() in all page-scripts

• 191 tests across 5 test files (up from 187)
• 4 new tests for word detection (standalone, split, hyphenated, compound)

Religious compound profanity support and false positive fixes:
- Add explicit combinations (blocked at all levels)
- Update phrase regex to match hyphenated variants
- Whitelist "damnable" at Low sensitivity (theological term)
- Remove "naked" from wordlist (not profanity)
- Verified zero false positives on two religious sermon SRTs

Fix disable button: clear overlay, unmute, persist state, reach all frames
Fix pre-existing TypeScript errors: undeclared vars, dead code, missing Cue fields