Give a wedged ComfyUI ten restarts per clip instead of three

Three was not enough in practice: a backend that dies during model load can
burn all three before it ever comes up healthy, and the clip is then skipped —
or, when the backend is simply gone, the run stalls waiting for it.

Both caps move together. COMFY_AUTOSTART_MAX is what actually gates a relaunch,
so raising CLIP_CRASH_LIMIT alone would still have stopped at three restarts.
Counting is unchanged: both are per clip, not per run.

The cost of a truly dead backend goes up accordingly — up to ten rounds of a
90s cooldown plus the 180s health timeout before a clip is given up on.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NkWDx9dVB7xBpqntvjvm6A

lektor_ai_comfy.py

@@ -1542,7 +1542,7 @@ BACKEND_RETRY_SECONDS = 3.0     # how often to look for ComfyUI while it is down
 
 # Launching the backend is guarded: a ComfyUI that dies on startup would
 # otherwise spawn a new terminal window every few seconds.
-COMFY_AUTOSTART_MAX      = 3      # restarts allowed while one clip is in flight
+COMFY_AUTOSTART_MAX      = 10     # restarts allowed while one clip is in flight
 COMFY_AUTOSTART_COOLDOWN = 90.0   # seconds between attempts
 
 # A clip that kills the backend this many times is skipped rather than retried
@@ -1550,7 +1550,7 @@ COMFY_AUTOSTART_COOLDOWN = 90.0   # seconds between attempts
 # tripped a CUDA device-side assert inside the audio tokenizer, which cannot be
 # caught in Python — and without a cap an overnight queue stops there and does
 # nothing until morning.
-CLIP_CRASH_LIMIT = 3
+CLIP_CRASH_LIMIT = 10
 
 # How long a restarted backend gets to answer before we conclude it is not the
 # clip's fault but the backend's absence.
@@ -1752,7 +1752,7 @@ def trim_comfy_log(path: Path = None, max_lines: int = COMFY_LOG_MAX_LINES) -> i
 def wait_for_backend(timeout: float = 300.0) -> bool:
     """Block until the backend answers again. After a restart the port is not
     listening for a while, and submitting into it would look like another
-    crash and burn one of the clip's three attempts."""
+    crash and burn one of the clip's attempts."""
     deadline = time.monotonic() + timeout
     while time.monotonic() < deadline:
         try:
@@ -1813,7 +1813,7 @@ def run_workflow(workflow: dict, poll_interval: float = 0.25,
     prompt_id, offline = None, False
     start, interval, next_dot = time.monotonic(), poll_interval, 1.0
     # Restarts are counted per clip, not per run. A crash in block 4000 of an
-    # overnight queue deserves the same three attempts the first one got.
+    # overnight queue deserves the same number of attempts the first one got.
     # One death is enough to hand control back: the caller retries with a
     # different seed, which is the only thing that changes the outcome.
     crashes = 0
@@ -2002,7 +2002,7 @@ def generate_clips(blocks: list, voice_filename: str, voice_ref_text: str,
             # A crash is seed-dependent, not purely text-dependent: the block
             # that killed ComfyUI three times in a 2073-clip run generated
             # perfectly on its own with a different seed. So each attempt gets a
-            # fresh seed — three tries at the identical computation would just
+            # fresh seed — repeated tries at the identical computation would just
             # reproduce the same abort.
             seed = base_seed + idx + attempt * 100_003
             try:
@@ -2015,7 +2015,7 @@ def generate_clips(blocks: list, voice_filename: str, voice_ref_text: str,
                 break
             except BackendCrashed as exc:
                 attempt += 1
-                # The backend died three times on this clip. That means one of
+                # The backend died on this clip. That means one of
                 # two very different things, and the difference decides whether
                 # skipping is right: either this clip kills ComfyUI — some
                 # inputs abort it from inside CUDA — or ComfyUI is simply not