Full correctness check against the docs: safer writes, reverts and texts

Safety
- Writes are serialized (Workspace.lock): two background jobs rebuilding one
  package (a sound and its subtitle applied side by side) lost an edit.
- A store copies a bank file or package as "original" only if it matches the
  shipped catalog (Workspace.unoriginal_files: in-memory sounds, bank copies of
  voice lines, subtitles); the editor refuses to switch stores while changes
  are applied. A second store could otherwise resurrect a reverted sound.
- Mod import in the editor asks before overwriting files changed outside the
  tool (it forced silently).
- Revert all in the editor reverts texts too; the uninstaller counts text
  edits before offering to delete the store.
- ackit restore-backup: the full-backup restore decisions.md described; it
  keeps the sounds of WADs without a full copy recorded.
- build_catalog also stops when a subtitle differs from the catalog.

Languages and texts
- Game.language() returns languages without voices (JPN, RUS, ...) and
  Game.voice_archive() their INT voices; ackit text --lang JPN works.
- Menu lookups parse one cached file (find & replace: minutes -> ms).
- Subtitles decode cp1252 as Windows does: the UTF-8-in-cp1252 line
  Speech_TAPE_J_Answer_1_00 (KOR/POL/POR/RUS) is written back byte-for-byte;
  SUB_KOR/SUB_RUS rebuilt from the original game (only those rows changed).
- Mods exported while another language is loaded keep name and line.

Other fixes
- Down-mix folds extra channels in at -3 dB (a 5.1 centre voice was dropped).
- e2e_editor no longer overwrites the user's settings.json; kill_game waits
  for the game to exit before reverts.
- Recorder handles Int32/UInt8 input and split frames; package block cache
  locked; conflict messages, CLI help examples, stale docstrings and doc
  references fixed; tests no longer read this PC's game language.

Docs (decisions, user guide, README, subtitles.md, banks.md, catalog README,
REPORT, NOTES, CLAUDE.md) updated; check_docs: 102 claims match. 148 tests;
e2e_features 9/9, e2e_editor and verify PASS; game original afterwards.

ackit/audio.py

@@ -53,21 +53,29 @@ def resample(x, rate, new_rate):
 
 
 def to_channels(x, channels):
-    """Up/down-mix. Mono is copied to every channel; extra channels get the average."""
+    """Up/down-mix. Up: mono is copied to every channel, other sources keep their channels and the new ones get
+    the average. Down: nothing is dropped - every channel goes into the mono sum, or the channels beyond the
+    kept ones (e.g. the centre and surrounds of 5.1) are added to every kept one at -3 dB (the standard
+    fold-down); a result louder than the source's peak is scaled back to it, so it does not clip."""
     x = _as2d(x)
     cur = x.shape[1]
     if cur == channels:
         return x
-    if channels == 1:
-        return x.mean(axis=1, keepdims=True)
     if cur == 1:
         return np.repeat(x, channels, axis=1)
-    out = np.empty((len(x), channels), dtype=np.float32)
-    keep = min(cur, channels)
-    out[:, :keep] = x[:, :keep]
-    if channels > keep:
-        out[:, keep:] = x.mean(axis=1, keepdims=True)
-    return out
+    if channels > cur:
+        out = np.empty((len(x), channels), dtype=np.float32)
+        out[:, :cur] = x
+        out[:, cur:] = x.mean(axis=1, keepdims=True)
+        return out
+    if channels == 1:
+        out = x.sum(axis=1, keepdims=True)
+    else:
+        out = x[:, :channels] + 0.7071 * x[:, channels:].sum(axis=1, keepdims=True)
+    peak_in, peak_out = float(np.max(np.abs(x), initial=0.0)), float(np.max(np.abs(out), initial=0.0))
+    if peak_out > peak_in > 0:
+        out = out * (peak_in / peak_out)
+    return out.astype(np.float32)
 
 
 def trim(x, rate, start=None, end=None):