audiobook_creator/gui_proper_noun_player.py

"""
proper_noun_player.py
──────────────────────
GUI for auditing proper noun pronunciations.

Three columns (all persisted as JSON, original manifest never modified):
  • Review   – words not yet audited
  • Correct  – words that already pronounce fine
  • Fixes    – linked list: original word → phonetic replacement
               e.g.  "Nephi" → "Kneephi"

Hotkeys (always active):
  Space          – replay current word
  s              – stop audio
  Escape         – reset fix entry to original word, refocus review list

On the Review list:
  ↑ / ↓          – navigate
  Click / Enter  – play word AND focus fix entry

On the fix entry (bottom bar, right of the word label):
  Start typing to overwrite the pre-filled word.
  Enter  →  if text == original word  →  mark Correct, advance to next
            if text differs           →  add as Fix, advance to next
  Escape →  reset text to original word, return focus to review list

On the Correct list:
  Delete / BackSpace – move selected word back to Review

On the Fixes list:
  Delete / BackSpace – move selected fix back to Review

"Apply Fixes to Text" writes a TTS-ready copy of the source file with all
substitutions applied (case-sensitive whole-word replace).

Data files (auto-created in output_proper_nouns/):
  correct_words.json       – list of correct words
  pronunciation_fixes.json – { "Nephi": "Kneephi", … }

Run:
    .venv/bin/python proper_noun_player.py
"""

import json
import os
import re
import threading
import tkinter as tk
from tkinter import ttk, messagebox
from pathlib import Path

# Model is already cached locally — skip all HuggingFace Hub network calls
os.environ.setdefault("HF_HUB_OFFLINE", "1")

import sounddevice as sd
import soundfile as sf

MANIFEST_FILE        = Path("output_proper_nouns/manifest.json")
OUTPUT_DIR           = Path("output_proper_nouns")
REPLACEMENTS_DIR     = OUTPUT_DIR / "replacements_cache"
CORRECT_FILE         = OUTPUT_DIR / "correct_words.json"
FIXES_FILE           = OUTPUT_DIR / "pronunciation_fixes.json"
SOURCE_TEXT          = Path("Audio Master Nem Full.txt")
FIXED_TEXT_OUT       = Path("Audio Master Nem Full (TTS Fixed).txt")

VOICE                = "am_michael"
SAMPLE_RATE          = 24000

# ── Colours ────────────────────────────────────────────────────────────────────
BG      = "#1e1e2e"
BG2     = "#181825"
BG3     = "#313244"
FG      = "#cdd6f4"
FG_DIM  = "#6c7086"
GREEN   = "#a6e3a1"
BLUE    = "#89b4fa"
RED     = "#f38ba8"
YELLOW  = "#f9e2af"
MAUVE   = "#cba6f7"

# ── Audio ──────────────────────────────────────────────────────────────────────

def play_async(path: Path) -> None:
    sd.stop()
    def _play():
        data, sr = sf.read(str(path), dtype="float32")
        sd.play(data, sr)
    threading.Thread(target=_play, daemon=True).start()


def _slug(text: str) -> str:
    """Safe filename from arbitrary text."""
    return re.sub(r"[^a-zA-Z0-9_-]", "_", text).strip("_")[:80]


# Lazy KPipeline singleton — only imported+loaded on first synthesis request
_pipeline = None
_pipeline_lock = threading.Lock()

def _get_pipeline():
    global _pipeline
    if _pipeline is None:
        with _pipeline_lock:
            if _pipeline is None:
                import warnings
                from kokoro import KPipeline  # type: ignore
                with warnings.catch_warnings():
                    warnings.filterwarnings("ignore", category=UserWarning)
                    warnings.filterwarnings("ignore", category=FutureWarning)
                    warnings.filterwarnings("ignore", message=".*unauthenticated.*")
                    _pipeline = KPipeline(lang_code="a", repo_id="hexgrad/Kokoro-82M")
    return _pipeline


def synth_and_play(text: str, on_ready=None) -> None:
    """Synthesise *text* with Kokoro (cached) and play it.
    Runs entirely on a daemon thread so the GUI never blocks.
    *on_ready(path)* is called on the same thread once the file is written.
    """
    def _run():
        path = _synth_to_cache(text)
        if path:
            if on_ready:
                on_ready(path)
            play_async(path)

    threading.Thread(target=_run, daemon=True).start()


def _synth_to_cache(text: str) -> "Path | None":
    """Synthesise *text* to a cached WAV and return its path (or None on failure).
    Skips synthesis if the file already exists.  Safe to call from any thread.
    """
    REPLACEMENTS_DIR.mkdir(parents=True, exist_ok=True)
    cache_path = REPLACEMENTS_DIR / f"{_slug(text)}.wav"
    if not cache_path.exists():
        import warnings
        import numpy as np
        pipeline = _get_pipeline()
        chunks = []
        with warnings.catch_warnings():
            warnings.filterwarnings("ignore", category=UserWarning)
            for _, _, audio in pipeline(text, voice=VOICE):
                if audio is not None:
                    chunks.append(audio)
        if chunks:
            combined = np.concatenate(chunks)
            sf.write(str(cache_path), combined, SAMPLE_RATE)
    return cache_path if cache_path.exists() else None


# ── Persistence helpers ────────────────────────────────────────────────────────

def load_json(path: Path, default):
    if path.exists():
        return json.loads(path.read_text(encoding="utf-8"))
    return default

def save_json(path: Path, obj) -> None:
    path.write_text(json.dumps(obj, ensure_ascii=False, indent=2), encoding="utf-8")


# ── Styled widget helpers ──────────────────────────────────────────────────────

def make_listbox(parent) -> tuple[tk.Listbox, tk.Frame]:
    frame = tk.Frame(parent, bg=BG2, bd=0)
    sb = ttk.Scrollbar(frame, orient="vertical")
    sb.pack(side="right", fill="y")
    lb = tk.Listbox(
        frame,
        yscrollcommand=sb.set,
        font=("Helvetica", 11),
        bg=BG2, fg=FG,
        selectbackground=BLUE, selectforeground=BG,
        activestyle="none", bd=0, highlightthickness=0, relief="flat",
        exportselection=False,
    )
    lb.pack(side="left", fill="both", expand=True)
    sb.config(command=lb.yview)
    return lb, frame

def styled_btn(parent, text, command, color=FG, bg=BG3, **kw):
    return tk.Button(
        parent, text=text, command=command,
        bg=bg, fg=color, activebackground=BG2, activeforeground=color,
        font=("Helvetica", 10, "bold"), relief="flat", bd=0,
        padx=10, pady=5, cursor="hand2", **kw
    )

def section_label(parent, text):
    return tk.Label(parent, text=text, bg=BG, fg=FG_DIM,
                    font=("Helvetica", 9, "bold"), anchor="w")


# ── Main app ───────────────────────────────────────────────────────────────────

class ProperNounAuditor(tk.Tk):

    # tracks which word is currently loaded into the fix entry
    _fix_entry_word: str = ""

    def __init__(self, manifest: dict[str, str]) -> None:
        super().__init__()
        self.title("Proper Noun Pronunciation Auditor")
        self.geometry("1020x700")
        self.minsize(800, 500)
        self.configure(bg=BG)

        self.manifest: dict[str, str] = manifest
        self.all_words: list[str] = sorted(manifest.keys(), key=str.casefold)

        # Persistent data — correct is newest-first; fixes dict preserves insertion order
        self.correct: list[str]     = load_json(CORRECT_FILE, [])
        self.fixes: dict[str, str]  = load_json(FIXES_FILE, {})

        self._build_ui()
        self._refresh_all()

        # Window-level hotkeys (work even when a listbox has keyboard focus)
        self.bind("<space>",  lambda e: self._replay())
        self.bind("s",        lambda e: sd.stop())
        self.bind("r",        lambda e: self._regen_current()
                  if self.focus_get() is not self._fix_entry else None)
        self.bind("<Escape>", lambda e: self._reset_fix_entry())

    # ── UI construction ────────────────────────────────────────────────────────

    def _build_ui(self) -> None:
        PAD = 8

        # Title bar
        title_bar = tk.Frame(self, bg=BG, pady=6)
        title_bar.pack(fill="x", padx=PAD)
        tk.Label(title_bar, text="Proper Noun Pronunciation Auditor",
                 font=("Helvetica", 15, "bold"), bg=BG, fg=FG).pack(side="left")
        hint = "Space=replay  r=regen  s=stop  Esc=reset fix  Del=remove from list  Enter=correct|fix"
        tk.Label(title_bar, text=hint,
                 font=("Helvetica", 8), bg=BG, fg=FG_DIM).pack(side="left", padx=14)

        # Three-column body
        body = tk.Frame(self, bg=BG)
        body.pack(fill="both", expand=True, padx=PAD, pady=(0, PAD))
        body.columnconfigure(0, weight=3)
        body.columnconfigure(1, weight=2)
        body.columnconfigure(2, weight=2)
        body.rowconfigure(0, weight=1)

        # ── Column 0: Review list ──────────────────────────────────────────────
        col0 = tk.Frame(body, bg=BG)
        col0.grid(row=0, column=0, sticky="nsew", padx=(0, PAD))

        filter_row = tk.Frame(col0, bg=BG)
        filter_row.pack(fill="x", pady=(0, 4))
        tk.Label(filter_row, text="Filter:", bg=BG, fg=FG,
                 font=("Helvetica", 10)).pack(side="left", padx=(0, 4))
        self.search_var = tk.StringVar()
        self.search_var.trace_add("write", lambda *_: self._refresh_review())
        self._filter_entry = tk.Entry(
            filter_row, textvariable=self.search_var,
            font=("Helvetica", 11), bg=BG3, fg=FG,
            insertbackground=FG, relief="flat", bd=4)
        self._filter_entry.pack(side="left", fill="x", expand=True)
        self._filter_entry.focus_set()
        styled_btn(filter_row, "✕", lambda: self.search_var.set(""),
                   color=RED, bg=BG3).pack(side="left", padx=(3, 0))

        hdr0 = tk.Frame(col0, bg=BG)
        hdr0.pack(fill="x")
        section_label(hdr0, "TO REVIEW").pack(side="left")
        self.review_count_var = tk.StringVar()
        tk.Label(hdr0, textvariable=self.review_count_var, bg=BG, fg=FG_DIM,
                 font=("Helvetica", 9)).pack(side="right")

        self.review_lb, review_frame = make_listbox(col0)
        review_frame.pack(fill="both", expand=True)
        self.review_lb.bind("<<ListboxSelect>>", self._on_review_select)
        self.review_lb.bind("<Return>", self._on_review_select)

        # ── Column 1: Correct list ─────────────────────────────────────────────
        col1 = tk.Frame(body, bg=BG)
        col1.grid(row=0, column=1, sticky="nsew", padx=(0, PAD))

        hdr1 = tk.Frame(col1, bg=BG)
        hdr1.pack(fill="x")
        section_label(hdr1, "✓ CORRECT  [Del=remove]").pack(side="left")
        self.correct_count_var = tk.StringVar()
        tk.Label(hdr1, textvariable=self.correct_count_var, bg=BG, fg=FG_DIM,
                 font=("Helvetica", 9)).pack(side="right")

        self.correct_lb, correct_frame = make_listbox(col1)
        correct_frame.pack(fill="both", expand=True)
        self.correct_lb.bind("<<ListboxSelect>>",
                             lambda e: self._on_side_select(self.correct_lb))
        self.correct_lb.bind("<Delete>",
                             lambda e: self._move_back(self.correct_lb, is_dict=False))
        self.correct_lb.bind("<BackSpace>",
                             lambda e: self._move_back(self.correct_lb, is_dict=False))

        styled_btn(col1, "← Back to Review  [Del]",
                   lambda: self._move_back(self.correct_lb, is_dict=False),
                   color=YELLOW).pack(fill="x", pady=(4, 0))

        # ── Column 2: Fixes list ───────────────────────────────────────────────
        col2 = tk.Frame(body, bg=BG)
        col2.grid(row=0, column=2, sticky="nsew")

        hdr2 = tk.Frame(col2, bg=BG)
        hdr2.pack(fill="x")
        section_label(hdr2, "⇄ FIXES  (original → phonetic)").pack(side="left")
        self.fixes_count_var = tk.StringVar()
        tk.Label(hdr2, textvariable=self.fixes_count_var, bg=BG, fg=FG_DIM,
                 font=("Helvetica", 9)).pack(side="right")

        self.fixes_lb, fixes_frame = make_listbox(col2)
        fixes_frame.pack(fill="both", expand=True)
        self.fixes_lb.bind("<<ListboxSelect>>",
                           lambda e: self._on_side_select(self.fixes_lb))
        self.fixes_lb.bind("<Delete>",
                           lambda e: self._move_back(self.fixes_lb, is_dict=True))
        self.fixes_lb.bind("<BackSpace>",
                           lambda e: self._move_back(self.fixes_lb, is_dict=True))

        styled_btn(col2, "← Back to Review  [Del]",
                   lambda: self._move_back(self.fixes_lb, is_dict=True),
                   color=YELLOW).pack(fill="x", pady=(4, 0))

        # ── Bottom action bar ──────────────────────────────────────────────────
        action_bar = tk.Frame(self, bg=BG3, pady=8)
        action_bar.pack(fill="x")

        # Now-playing word label
        tk.Label(action_bar, text="▶", bg=BG3, fg=GREEN,
                 font=("Helvetica", 11)).pack(side="left", padx=(10, 2))
        self.now_playing_var = tk.StringVar(value="—")
        tk.Label(action_bar, textvariable=self.now_playing_var,
                 bg=BG3, fg=GREEN, font=("Helvetica", 11, "bold"),
                 width=20, anchor="w").pack(side="left")

        # Inline fix entry — right next to the word, auto-focused on word click
        tk.Label(action_bar, text="→", bg=BG3, fg=MAUVE,
                 font=("Helvetica", 13, "bold")).pack(side="left", padx=(6, 3))
        self.fix_var = tk.StringVar()
        self._fix_entry = tk.Entry(
            action_bar, textvariable=self.fix_var,
            font=("Helvetica", 11), bg=BG2, fg=MAUVE,
            insertbackground=MAUVE, relief="flat", bd=4, width=22)
        self._fix_entry.pack(side="left")
        self._fix_entry.bind("<Return>", lambda e: self._enter_action())
        self._fix_entry.bind("<Escape>", lambda e: self._reset_fix_entry())
        self._fix_entry.bind("<Up>",   lambda e: (self._navigate_review(-1), "break")[1])
        self._fix_entry.bind("<Down>", lambda e: (self._navigate_review(+1), "break")[1])

        tk.Label(action_bar, text="Enter=correct  (edit first for fix)  Esc=reset",
                 bg=BG3, fg=FG_DIM, font=("Helvetica", 8)).pack(side="left", padx=(5, 10))

        tk.Label(action_bar, text="│", bg=BG3, fg=FG_DIM).pack(side="left", padx=4)
        styled_btn(action_bar, "■ Stop  [s]", sd.stop,
                   color=RED).pack(side="left", padx=4)
        styled_btn(action_bar, "↺ Replay  [Space]", self._replay,
                   color=BLUE).pack(side="left", padx=2)
        styled_btn(action_bar, "↻ Regen  [r]", self._regen_current,
                   color=GREEN).pack(side="left", padx=2)

        tk.Label(action_bar, text="│", bg=BG3, fg=FG_DIM).pack(side="left", padx=4)
        styled_btn(action_bar, "⇄ Apply Fixes to Text",
                   self._apply_fixes, color=YELLOW, bg=BG2).pack(side="left", padx=4)
        styled_btn(action_bar, "⬇ Export Remaining",
                   self._export_remaining, color=BLUE, bg=BG2).pack(side="left", padx=4)

        tk.Label(action_bar, text="│", bg=BG3, fg=FG_DIM).pack(side="left", padx=4)
        self._pregen_btn = styled_btn(
            action_bar, "↻ Pre-gen Fix Audio",
            self._pregen_all_fix_audio, color=MAUVE, bg=BG2)
        self._pregen_btn.pack(side="left", padx=4)
        self._pregen_status_var = tk.StringVar(value="")
        tk.Label(action_bar, textvariable=self._pregen_status_var,
                 bg=BG3, fg=FG_DIM, font=("Helvetica", 8),
                 width=28, anchor="w").pack(side="left", padx=(4, 10))

    # ── Refresh helpers ────────────────────────────────────────────────────────

    def _review_words(self) -> list[str]:
        excluded = set(self.correct) | set(self.fixes.keys())
        q = self.search_var.get().strip().casefold()
        words = [w for w in self.all_words if w not in excluded]
        if q:
            words = [w for w in words if q in w.casefold()]
        return words

    def _refresh_review(self) -> None:
        words = self._review_words()
        self.review_lb.delete(0, "end")
        for w in words:
            self.review_lb.insert("end", f"  {w}")
        self.review_count_var.set(f"{len(words)}")

    def _refresh_correct(self) -> None:
        self.correct_lb.delete(0, "end")
        for w in self.correct:  # already newest-first
            self.correct_lb.insert("end", f"  {w}")
        self.correct_count_var.set(f"{len(self.correct)}")

    def _refresh_fixes(self) -> None:
        self.fixes_lb.delete(0, "end")
        for orig, rep in reversed(list(self.fixes.items())):  # newest-first
            self.fixes_lb.insert("end", f"  {orig}  →  {rep}")
        self.fixes_count_var.set(f"{len(self.fixes)}")

    def _refresh_all(self) -> None:
        self._refresh_review()
        self._refresh_correct()
        self._refresh_fixes()

    # ── Playback ───────────────────────────────────────────────────────────────

    def _play_word(self, word: str) -> None:
        wav_name = self.manifest.get(word)
        if not wav_name:
            return
        wav_path = OUTPUT_DIR / wav_name
        if not wav_path.exists():
            messagebox.showwarning("Missing audio",
                                   f"No audio file for '{word}'.\n"
                                   "Run generate_proper_noun_audio.py first.")
            return
        self.now_playing_var.set(word)
        play_async(wav_path)

    # ── Selection callbacks ────────────────────────────────────────────────────

    def _on_review_select(self, event=None) -> None:
        sel = self.review_lb.curselection()
        if not sel:
            return
        word = self.review_lb.get(sel[0]).strip()
        self._fix_entry_word = word
        self.fix_var.set(word)              # pre-fill fix entry with the word
        self._fix_entry.selection_range(0, "end")
        self._fix_entry.icursor("end")
        # Defer focus so the listbox doesn't reclaim it after the click event settles
        self.after(0, self._fix_entry.focus_set)
        self._play_word(word)

    def _on_side_select(self, listbox: tk.Listbox) -> None:
        sel = listbox.curselection()
        if not sel:
            return
        row = listbox.get(sel[0]).strip()
        parts = row.split("  →  ")
        original = parts[0].strip()

        if listbox is self.fixes_lb and len(parts) == 2:
            # Show original → replacement in the fix entry, play the replacement
            replacement = parts[1].strip()
            self._fix_entry_word = original
            self.fix_var.set(replacement)
            self.now_playing_var.set(f"… {replacement}")
            def _on_ready(_path):
                self.after(0, lambda: self.now_playing_var.set(replacement))
            synth_and_play(replacement, on_ready=_on_ready)
        else:
            # Correct list — show word in fix entry, play it
            self._fix_entry_word = original
            self.fix_var.set(original)
            self._play_word(original)

    # ── Actions ────────────────────────────────────────────────────────────────

    def _selected_review_word(self) -> str | None:
        sel = self.review_lb.curselection()
        if not sel:
            return None
        return self.review_lb.get(sel[0]).strip()

    def _enter_action(self) -> None:
        """Smart Enter handler for the fix entry.

        If the entry text matches the original word  → mark Correct.
        If the entry text differs from the original  → add as Fix.
        """
        word = self._fix_entry_word or self._selected_review_word()
        if not word:
            return
        text = self.fix_var.get().strip()
        if not text or text == word:
            self._mark_correct_word(word)
        else:
            self._add_fix_for_word(word, text)

    def _reset_fix_entry(self) -> None:
        """Escape: reset fix entry to the original word, refocus the review list."""
        self.fix_var.set(self._fix_entry_word)
        self.review_lb.focus_set()

    def _replay(self) -> None:
        if self._fix_entry_word:
            self._play_word(self._fix_entry_word)

    def _regen_current(self) -> None:
        """Delete the cached WAV for the current word/replacement and re-synthesise."""
        word = self._fix_entry_word
        if not word:
            return

        # Determine which file to delete based on context
        fix_text = self.fix_var.get().strip()
        # If the fix box contains something different from the word, regen that text
        is_fix_replacement = bool(fix_text and fix_text != word)

        if is_fix_replacement:
            # Re-gen the cached replacement audio
            target = REPLACEMENTS_DIR / f"{_slug(fix_text)}.wav"
            if target.exists():
                target.unlink()
            self.now_playing_var.set(f"… regen {fix_text}")
            def _on_ready(_p):
                self.after(0, lambda: self.now_playing_var.set(fix_text))
            synth_and_play(fix_text, on_ready=_on_ready)
        else:
            # Re-gen the manifest audio for the review word
            wav_name = self.manifest.get(word)
            if not wav_name:
                return
            wav_path = OUTPUT_DIR / wav_name
            if wav_path.exists():
                wav_path.unlink()
            self.now_playing_var.set(f"… regen {word}")

            def _regen():
                import warnings, numpy as np
                pipeline = _get_pipeline()
                chunks = []
                with warnings.catch_warnings():
                    warnings.filterwarnings("ignore", category=UserWarning)
                    for _, _, audio in pipeline(word, voice=VOICE):
                        if audio is not None:
                            chunks.append(audio)
                if chunks:
                    sf.write(str(wav_path), np.concatenate(chunks), SAMPLE_RATE)
                    self.after(0, lambda: self.now_playing_var.set(word))
                    play_async(wav_path)

            threading.Thread(target=_regen, daemon=True).start()

    def _navigate_review(self, delta: int) -> None:
        """Move the review list selection up (delta=-1) or down (delta=+1)."""
        size = self.review_lb.size()
        if size == 0:
            return
        sel = self.review_lb.curselection()
        current = sel[0] if sel else -1
        new_idx = max(0, min(size - 1, current + delta))
        if new_idx == current:
            return
        self.review_lb.selection_clear(0, "end")
        self.review_lb.selection_set(new_idx)
        self.review_lb.see(new_idx)
        self.review_lb.event_generate("<<ListboxSelect>>")

    def _advance_review(self, from_idx: int = 0) -> None:
        """Select the item at from_idx (clamped), positioned in the upper portion
        of the viewport so the word doesn't end up in the bottom half unless
        the list can't scroll any further down."""
        size = self.review_lb.size()
        if size == 0:
            return
        target = min(from_idx, size - 1)
        self.review_lb.selection_clear(0, "end")
        self.review_lb.selection_set(target)

        # First call see() to let tk calculate the viewport, then reposition.
        self.review_lb.see(target)
        self.review_lb.update_idletasks()

        first, last = self.review_lb.yview()
        visible_count = max(1, round((last - first) * size))

        # Ideal top-of-viewport: put target ~1/4 down from the top
        ideal_top = target - visible_count // 4
        ideal_top = max(0, ideal_top)

        self.review_lb.yview_moveto(ideal_top / size)
        self.review_lb.event_generate("<<ListboxSelect>>")

    def _mark_correct_word(self, word: str) -> None:
        idx = self.review_lb.curselection()
        from_idx = idx[0] if idx else 0
        if word not in self.correct:
            self.correct.insert(0, word)
        save_json(CORRECT_FILE, self.correct)
        self._fix_entry_word = ""
        self.fix_var.set("")
        self.now_playing_var.set("—")
        self._refresh_all()
        self._advance_review(from_idx)

    def _mark_correct(self) -> None:
        word = self._selected_review_word()
        if not word:
            messagebox.showinfo("Nothing selected",
                                "Select a word from the Review list first.")
            return
        self._mark_correct_word(word)

    def _add_fix_for_word(self, word: str, replacement: str) -> None:
        idx = self.review_lb.curselection()
        from_idx = idx[0] if idx else 0
        # Remove and re-add so updated entries bubble to the top
        self.fixes.pop(word, None)
        self.fixes[word] = replacement
        save_json(FIXES_FILE, self.fixes)
        self._fix_entry_word = ""
        self.fix_var.set("")
        self.now_playing_var.set("—")
        self._refresh_all()
        self._advance_review(from_idx)

    def _add_fix(self) -> None:
        word = self._selected_review_word()
        replacement = self.fix_var.get().strip()
        if not word:
            messagebox.showinfo("Nothing selected",
                                "Select a word from the Review list first.")
            return
        if not replacement or replacement == word:
            messagebox.showinfo("No replacement",
                                "Type the phonetic replacement in the Fix box.")
            return
        self._add_fix_for_word(word, replacement)

    def _move_back(self, listbox: tk.Listbox, is_dict: bool) -> None:
        sel = listbox.curselection()
        if not sel:
            return
        raw = listbox.get(sel[0]).strip().split("  →  ")[0].strip()
        if is_dict:
            self.fixes.pop(raw, None)
            save_json(FIXES_FILE, self.fixes)
        else:
            if raw in self.correct:
                self.correct.remove(raw)
            save_json(CORRECT_FILE, self.correct)
        self._refresh_all()

    # ── Apply fixes to source text ─────────────────────────────────────────────

    def _pregen_all_fix_audio(self) -> None:
        """Synthesise and cache audio for every replacement phonetic string."""
        if not self.fixes:
            messagebox.showinfo("No fixes", "The Fixes list is empty.")
            return

        replacements = list(self.fixes.values())
        total = len(replacements)
        already = sum(
            1 for r in replacements
            if (REPLACEMENTS_DIR / f"{_slug(r)}.wav").exists()
        )

        # Confirm if it'll take a while
        new_count = total - already
        if new_count == 0:
            messagebox.showinfo("Already done",
                                f"All {total} replacement clips already exist.")
            return

        self._pregen_btn.config(state="disabled")
        self._pregen_status_var.set(f"0 / {new_count} new  ({already} cached)")

        def _run():
            done = 0
            for rep in replacements:
                cache_path = REPLACEMENTS_DIR / f"{_slug(rep)}.wav"
                if not cache_path.exists():
                    _synth_to_cache(rep)
                    done += 1
                    self.after(0, lambda d=done, t=new_count:
                               self._pregen_status_var.set(f"{d} / {t} synthesised…"))
            self.after(0, lambda: self._pregen_status_var.set(
                f"Done — {total} clips ready"))
            self.after(0, lambda: self._pregen_btn.config(state="normal"))

        threading.Thread(target=_run, daemon=True).start()

    def _export_remaining(self) -> None:
        words = self._review_words()
        if not words:
            messagebox.showinfo("Nothing to export", "No words left to review.")
            return
        out = OUTPUT_DIR / "remaining_review.txt"
        out.write_text("\n".join(words), encoding="utf-8")
        messagebox.showinfo("Exported",
                            f"{len(words)} words written to:\n{out}")

    def _apply_fixes(self) -> None:
        if not self.fixes:
            messagebox.showinfo("No fixes", "The Fixes list is empty.")
            return
        if not SOURCE_TEXT.exists():
            messagebox.showerror("Source not found", f"Cannot find:\n{SOURCE_TEXT}")
            return
        text = SOURCE_TEXT.read_text(encoding="utf-8")
        count_total = 0
        for original, replacement in self.fixes.items():
            pattern = r'\b' + re.escape(original) + r'\b'
            new_text, n = re.subn(pattern, replacement, text)
            if n:
                text = new_text
                count_total += n
        FIXED_TEXT_OUT.write_text(text, encoding="utf-8")
        messagebox.showinfo(
            "Done",
            f"Applied {len(self.fixes)} fix rules ({count_total} replacements).\n\n"
            f"Saved to:\n{FIXED_TEXT_OUT}"
        )


# ── Entry point ────────────────────────────────────────────────────────────────

def main() -> None:
    if not MANIFEST_FILE.exists():
        print(f"Manifest not found: '{MANIFEST_FILE}'")
        print("Run generate_proper_noun_audio.py first.")
        raise SystemExit(1)

    manifest: dict[str, str] = json.loads(MANIFEST_FILE.read_text(encoding="utf-8"))
    print(f"Loaded {len(manifest)} entries from manifest.")

    app = ProperNounAuditor(manifest)
    app.mainloop()


if __name__ == "__main__":
    main()