fix(prompt): strip web prompt-helper @ markers before diffusion

The web prompt-helper glues an @ marker onto character/series tokens
(e.g. @hatsune_miku). Diffusion models don't recognise it, so strip a
leading @-before-word at the top of generate_image — covering both the
direct image path and the video first-frame render. Emails / mid-word @
are preserved.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -148,6 +148,21 @@ MINOR_NEG = (
 # from biting into ``girlish`` etc.
 _GIRL_RE = re.compile(r"(?<![a-z0-9])(\d*)\s*girls?(?![a-z])", re.IGNORECASE)
 
+# The web prompt-helper glues an ``@`` marker onto character/series tokens it
+# inserts (e.g. ``@hatsune_miku``). That ``@`` is purely a UI affordance — the
+# diffusion models don't recognise it and it pollutes the conditioning. Strip a
+# leading ``@`` wherever it sits directly in front of a word so the bare booru
+# token (``hatsune_miku``) reaches the sampler. Mid-word ``@`` (none expected in
+# tag-space) is left alone.
+_AT_MARK_RE = re.compile(r"(?<![\w@])@(?=\w)")
+
+
+def _strip_trigger_marks(prompt):
+    """Remove the prompt-helper ``@`` markers from a user prompt (no-op if none)."""
+    if not prompt:
+        return prompt
+    return _AT_MARK_RE.sub("", prompt)
+
 
 def _adultify_prompt(prompt):
     """Replace youthful female subject tokens with adult equivalents.
@@ -952,6 +967,11 @@ def generate_image(
     _gpu_start = time.time()
     total_steps = int(num_inference_steps)
 
+    # Strip the web prompt-helper's ``@`` markers (e.g. ``@hatsune_miku``) up
+    # front — the models don't recognise them — so the improve/adultify passes
+    # and the sampler all see the clean tokens.
+    prompt = _strip_trigger_marks(prompt)
+
     # Optional prompt-improvement pass: rewrite the user's prompt with the VLM
     # before sampling. Runs inside this same @spaces.GPU window (single ZeroGPU
     # allocation), mirroring how the moderation check rides along below.