speech.py --engines: default to F5 only; 'all' opts into every engine

Previous default ('--engines unset' == all engines enabled) made it too
easy for stray tts-1-qwen requests to silently load 10 GiB of weights
onto a GPU shared with an LLM. Flipped: the lean F5-only allowlist is
now the default, and operators explicitly opt into more.

  python speech.py                       # default: {tts-1-f5}
  python speech.py --engines f5,piper    # F5 + Piper
  python speech.py --engines all         # every importable engine (old default)
  python speech.py --engines qwen,f5     # back to Qwen + F5 for GPU servers

Implementation: _parse_engines_env() now returns set(DEFAULT_ENGINES)
when SPEECH_ENABLED_ENGINES is unset, and treats the literal 'all' as a
None sentinel (no allowlist applied). DEFAULT_ENGINES = frozenset({'tts-1-f5'}).

Workers still pick this up via SPEECH_ENABLED_ENGINES env var; __main__
only sets the var when --engines was passed, so default-path workers
re-parse the empty env -> DEFAULT_ENGINES path.

speech.py

@@ -82,10 +82,18 @@ ENGINE_SHORT_TO_MODEL = {
 }
 ALL_MODEL_IDS = set(ENGINE_SHORT_TO_MODEL.values())
 
+# Lean default — F5 alone fits comfortably on a GPU shared with an LLM.
+# Override with --engines (or SPEECH_ENABLED_ENGINES env var):
+#   --engines f5,piper   -> {tts-1-f5, tts-1}
+#   --engines all        -> no allowlist (every importable engine enabled)
+DEFAULT_ENGINES = frozenset({'tts-1-f5'})
+
 def _parse_engines_env():
     raw = os.environ.get('SPEECH_ENABLED_ENGINES', '').strip()
     if not raw:
-        return None
+        return set(DEFAULT_ENGINES)
+    if raw.lower() == 'all':
+        return None  # explicit opt-in to every engine
     allowed = set()
     for e in (s.strip() for s in raw.split(',')):
         if not e:
@@ -98,11 +106,11 @@ def _parse_engines_env():
             raise ValueError(
                 f"SPEECH_ENABLED_ENGINES: unknown engine '{e}' "
                 f"(valid short names: {', '.join(sorted(ENGINE_SHORT_TO_MODEL))}, "
-                f"or full tts-1-* ids)"
+                f"full tts-1-* ids, or the keyword 'all')"
             )
     return allowed
 
-ENABLED_ENGINES = _parse_engines_env()  # None = all enabled; set = allowlist
+ENABLED_ENGINES = _parse_engines_env()  # None = unrestricted; set = allowlist
 
 def is_engine_available(model_id):
     """Whether the TTS engine for a given model_id is loadable in this process.
@@ -1539,9 +1547,8 @@ if __name__ == "__main__":
         help="Comma-separated allowlist of TTS engines to enable. "
              "Short names: f5, qwen, piper, xtts, silero, kokoro "
              "(or full model IDs like tts-1-f5). "
-             "Default: all engines enabled. "
-             "Example: --engines f5,piper keeps speech lean enough to "
-             "share GPU with an LLM server.")
+             "Pass --engines all to enable every importable engine. "
+             "Default: f5 only (lean, fits on a GPU shared with an LLM).")
 
     args = parser.parse_args()
 
@@ -1549,7 +1556,9 @@ if __name__ == "__main__":
     logger.add(sink=sys.stderr, level=args.log_level)
 
     # Propagate --engines to uvicorn workers via env var (workers re-import
-    # this module + read SPEECH_ENABLED_ENGINES into ENABLED_ENGINES).
+    # this module + read SPEECH_ENABLED_ENGINES into ENABLED_ENGINES at
+    # module load). When --engines is omitted, the env var stays unset
+    # and _parse_engines_env() falls through to DEFAULT_ENGINES (f5 only).
     if args.engines:
         os.environ['SPEECH_ENABLED_ENGINES'] = args.engines
         try:
@@ -1557,9 +1566,13 @@ if __name__ == "__main__":
         except ValueError as e:
             logger.error(str(e))
             sys.exit(2)
-        logger.info(f"--engines: restricting to {sorted(ENABLED_ENGINES)}")
+        if ENABLED_ENGINES is None:
+            logger.info("--engines all: every importable engine enabled")
+        else:
+            logger.info(f"--engines: restricting to {sorted(ENABLED_ENGINES)}")
     else:
-        logger.info("All engines enabled (no --engines restriction)")
+        logger.info(f"Default engine allowlist: {sorted(ENABLED_ENGINES)} "
+                    f"(pass --engines all or --engines f5,qwen,... to override)")
 
     if args.preload and not XTTS_AVAILABLE:
         logger.error("Cannot preload XTTS model - XTTS dependencies not available")