feat(commentary): refine rafters-critic persona and improve commentary prompt for humor

config/agents/rafters-critic.yaml

@@ -3,10 +3,12 @@ role: observer
 handler: commentator
 persona: >
   You are the Heckler in the Rafters — a washed-up theater critic who wandered into
-  the show one season, never left, and now sums up whatever the cast conjures with
-  affectionate snark from the cheap seats. You are funny, never cruel, and you never
-  break the spell of the show. You watch the whole stage, not the plot summary: in one
-  punchy line you capture what just happened, specific to it, and nothing else.
+  the show one season, never left, and now roasts whatever the cast conjures from the
+  cheap seats. Your voice is quick and deadpan, half snarky sports announcer, half
+  exhausted regular who has seen this exact nonsense before: you find the one absurd
+  detail in a scene and skewer it with a comparison nobody saw coming. Funny first,
+  affectionate always, never cruel, and you never break the spell of the show — you
+  don't narrate the plot or describe "the scene," you land one perfect line and sit back down.
 # Scenario-agnostic: this critic summarises only the public ledger, so it drops into any
 # cast (a scene, a debate, a mystery, a guessing game) with no engine edits.
 # Never event-woken (subscribes_to: []) so it can never re-trigger on its own remark;

modal/catalogue.py

@@ -283,8 +283,14 @@ GOOGLE_MODELS: tuple[ModelConfig, ...] = (
         # Standard gemma4 MoE arch (NOT the unified 12B path): served by a native
         # vLLM class on the pinned stable release (0.19.1+), so NO nightly, no
         # transformers pin, and CUDA graphs + async scheduling work — defaults stand.
-        # Text-only in the cast: zero the auto-detected multimodal caps.
-        mm_limits={"image": 0},
+        # Text-only in the cast, but image is this model's ONLY modality: zeroing it
+        # (as the 12B does for image+audio) empties the active-modality set, and vLLM
+        # 0.21.0's MultiModalBudget then calls max() on an empty sequence and crashes
+        # on boot (compute_mm_encoder_budget). The 12B escapes this only because its
+        # nightly wheel carries the defensive fix; on the stable pin we can't zero the
+        # last modality. So keep one image slot: the vision encoder warmup is tiny and
+        # the cast never sends images. Don't drop this to 0 without bumping vLLM.
+        mm_limits={"image": 1},
     ),
 )
 

src/agents/commentator.py

@@ -124,12 +124,23 @@ class Commentator(ManifestAgent):
     # ── prompt steering ─────────────────────────────────────────────────────────
 
     def _build_extra_prompt(self, projection: StageProjection, recent_events: tuple[Event, ...]) -> str:
-        """Steer the model toward a short, funny summary of the beat (not narration)."""
+        """Steer the model toward a genuinely funny one-line heckle of the beat.
+
+        Small models can't be funny on the word "funny" alone — they default to
+        cheerful narration. So we hand them a comedian's recipe: latch onto one
+        concrete detail, then break it with a twist (absurd comparison, deadpan
+        undercut, or mock-serious overreaction). Specific + surprising = the laugh."""
         return (
             "YOUR JOB\n"
-            "Sum up the beat above in ONE punchy, funny line — capture what the cast just "
-            "did, with affectionate wit, like a heckle from the cheap seats. Be specific. "
-            "No stage directions, no quotation marks, no lists. Just the line."
+            "Heckle the beat above with ONE short, funny line — the kind that gets a laugh, "
+            "not a polite nod. Work the bit like this:\n"
+            "- Grab ONE specific thing the cast just did — a prop, a word, a choice — and make "
+            "THAT the target. Never a vague 'well, that happened'.\n"
+            "- Then break it: an absurd comparison, a deadpan undercut, or a mock-serious "
+            "overreaction. The twist is where the laugh lives — surprise beats cleverness.\n"
+            "- Punch up at the drama, never down at a person. Affectionate, never cruel.\n"
+            "- ONE sentence. No narration, no stage directions, no quotation marks, no lists, "
+            "no emoji, no setup-then-punchline. Just the line, like you shouted it from the rafters."
         )
 
     # ── turn ──────────────────────────────────────────────────────────────────