fix: stop VLM assistant looping into long repeated output

Greedy decoding (do_sample=False) with no repetition controls fell into
n-gram repetition loops and ran to max_new_tokens instead of emitting EOS.
Add repetition_penalty=1.3 + no_repeat_ngram_size=3 — breaks the loops and
lets the model stop on its own, decoding stays deterministic.

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

app.py

@@ -661,6 +661,12 @@ def _vlm_chat_core(message, image, reasoning, max_new_tokens):
                     **inputs,
                     max_new_tokens=budget,
                     do_sample=False,
+                    # Greedy decoding alone loops into repeated/rambling output that
+                    # runs to max_new_tokens instead of emitting EOS. These penalties
+                    # break n-gram repetition and let the model stop on its own, while
+                    # keeping decoding deterministic (important for prompt rewrites).
+                    repetition_penalty=1.3,
+                    no_repeat_ngram_size=3,
                     streamer=streamer,
                 )
         except Exception as exc:  # noqa: BLE001 - surfaced to the main thread