latency: cache profile writes; docs: data2 resolved, latency and fade measurements

docs/NOTES.md

@@ -47,8 +47,25 @@ From `msttsl/mstts.inf` (`[TTSReg]`):
   `.COD` codebook, plus shared `DICT_CI.*`, `phondef`, `AMPRULES`, `ENDINGS`.
 
 Extraction: `unshield -O -d extracted x installer/flextalk/data1.cab`
-(`-O` = old compression; the archive.org package has no `data2.cab`, so the
-final file `WatsonUNInstall.dll` is incomplete).
+(`-O` = old compression). The package is single-volume: the original 1997 upload log
+(`WS_FTP.LOG`) lists `data1.cab` only, and `-O` extracts all 450 entries with exact
+sizes. Without `-O`, unshield's new decompressor reports a phantom "volume 2" and fails
+on `MSVCRT.DLL` / `WatsonUNInstall.dll`; with `-O` both are complete, valid PE files.
+
+The `.ENV` files the engine probes (`ae1m11lc.ENV`) were never shipped; synthesis works
+without them.
+
+## Latency and output shape (measured 2026-09-16)
+
+- Warm wrapper time: ~410-425 ms per short line (engine call 190-220 ms, ffmpeg measure
+  ~110 ms, apply ~90 ms, profile write ~57 ms when it runs). First (cold) call ~730 ms.
+- The Flextalk profile write is cached per voice in the Wine prefix
+  (`~/.wine-microslopsam/.microslopsam-profile<GUID>.reg`): unchanged profiles skip
+  `wine reg import`. Recreating the prefix invalidates the cache.
+- Fades: the 8 ms fade-in/out is applied to the trimmed file whose 25-30 ms padding
+  silence is retained, so it shapes silence, not speech. Measured head: first sample
+  >30 at 22.6 ms, speech (>500) at 29.6 ms, natural ~15 ms attack; tail decays over
+  ~18 ms. No click, no attenuation of the first word.
 
 ## SAPI4 runtime engine discovery (`speech.dll`)