fix(backend): first pass respects refined env flag; add A/B override

The first pass was hardcoded to baseline, so CHALCHITRA_REFINED_PROMPT was
ignored except on retry. Now the first pass uses the configured default
(refined=None reads the env). Adds an optional per-request `refined` override
(API + interpret) for a single clean pass — used to A/B baseline vs refined on
the same image against the live model. The frontend never sends it.

app.py

@@ -106,6 +106,7 @@ class Moment(BaseModel):
     images: list[str] = Field(default_factory=list)  # data URLs or raw base64 JPEG
     fragment: str = ""
     exclude: list[str] = Field(default_factory=list)  # titles already shown this session
+    refined: bool | None = None  # optional A/B override of the prompt variant
 
 
 @app.post("/api/interpret")
@@ -114,7 +115,11 @@ async def api_interpret(moment: Moment):
         return JSONResponse({"error": "Bring at least one image."}, status_code=400)
     try:
         return interpret(
-            moment.images, moment.fragment, provider=_provider(), exclude=moment.exclude
+            moment.images,
+            moment.fragment,
+            provider=_provider(),
+            exclude=moment.exclude,
+            refined=moment.refined,
         )
     except InputError as exc:
         # Bad/oversized/unsupported input — rejected before reaching the model.

backend/oracle.py

@@ -148,6 +148,7 @@ def interpret(
     fragment: str = "",
     provider: Provider | None = None,
     exclude: list[str] | None = None,
+    refined: bool | None = None,
 ) -> dict[str, Any]:
     """Run the full flow. `images` are data URLs or raw base64 JPEG strings.
 
@@ -155,9 +156,10 @@ def interpret(
     different films that still fit the moment (the "other films" re-roll).
     Stateless — that list lives on the client and rides in the request.
 
-    If the first interpretation reads flat or breaks the "never describe the
-    photo" rule, retry once with the refined prompt and a hotter temperature,
-    then keep whichever pass is stronger.
+    `refined` overrides the prompt variant for a single clean pass (no retry) —
+    used for A/B testing the baseline vs refined prompt on the same image. When
+    None (the normal path), the first pass uses the configured default and one
+    retry with the refined prompt covers a flat or unparseable reply.
     """
     # Guardrails first: decode/validate/re-encode images and clean the inputs
     # before anything reaches the model. InputError propagates to a clean 4xx.
@@ -168,10 +170,15 @@ def interpret(
     provider = provider or get_provider()
     content = _build_content(images, fragment, exclude)
 
-    # First pass with the configured prompt. If it parses and reads well, ship it.
+    # Explicit override (A/B / debugging): one clean pass with the chosen variant.
+    if refined is not None:
+        return _run(provider, content, refined=refined)
+
+    # First pass with the configured default (refined=None reads the env flag).
+    # If it parses and reads well, ship it.
     first: dict[str, Any] | None = None
     try:
-        first = _run(provider, content, refined=False)
+        first = _run(provider, content, refined=None)
         if not _is_weak(first["interpretation"]):
             return first
     except OracleError: