test(b5): Domain C text targeting / OCR red tests [F#6 F#5 F#4 F#7 F#18]

simdrive/src/simdrive/observe.py

@@ -206,17 +206,19 @@ def observe(
 
     marks: list[Mark] = []
     annotated_path: Path | None = None
-    if annotate:
-        marks = som.detect_marks(raw_path)
-        if marks:
-            annotated_path = out_dir / f"observe-{ts}-som.png"
-            # Annotate the *unfiltered* image so the on-disk PNG keeps the full
-            # context for human review — filtering is for the JSON payload only.
-            som.annotate(raw_path, marks, annotated_path)
-        # Apply token-efficiency filters AFTER annotation so the PNG retains
-        # every detected mark, but the in-memory + JSON `marks` list reflects
-        # what the agent actually receives.
-        marks = _apply_filters(marks, confidence_floor, mark_limit)
+    # F#7 (b5): always detect marks so text targeting works regardless of annotate flag.
+    # When annotate=True, also draw the SoM overlay and set annotated_path.
+    # When annotate=False, skip drawing — marks are still returned, annotated_path stays None.
+    marks = som.detect_marks(raw_path)
+    if annotate and marks:
+        annotated_path = out_dir / f"observe-{ts}-som.png"
+        # Annotate the *unfiltered* image so the on-disk PNG keeps the full
+        # context for human review — filtering is for the JSON payload only.
+        som.annotate(raw_path, marks, annotated_path)
+    # Apply token-efficiency filters AFTER annotation so the PNG retains
+    # every detected mark, but the in-memory + JSON `marks` list reflects
+    # what the agent actually receives.
+    marks = _apply_filters(marks, confidence_floor, mark_limit)
 
     logs_text: str | None = None
     if capture_logs:

simdrive/src/simdrive/server.py

@@ -942,7 +942,14 @@ def _resolve_target_xy(s, args: dict) -> tuple[int, int, str, "som.Mark | dict |
         candidates, tier = som.find_text_candidates(marks, query)
         if not candidates:
             available = [_mark_attr(mk, "text") for mk in marks]
-            raise errors.target_not_found("text", query, available)
+            err = errors.target_not_found("text", query, available)
+            # F#5 — include fuzzy suggestion so agents know the closest real mark.
+            import difflib
+            available_texts = [t for t in available if t]
+            matches = difflib.get_close_matches(query, available_texts, n=1, cutoff=0.5)
+            if matches:
+                err.details["suggestion"] = matches[0]
+            raise err
         # F#6 — >1 marks tied at the winning tier ⇒ refuse to silent-pick.
         # The agent must re-target by stable_id / mark / xy. Single-match (even
         # when other tiers also have matches) still resolves unambiguously.

simdrive/src/simdrive/som.py

@@ -76,6 +76,15 @@
         # misc UI
         "welcome", "hello", "goodbye", "logout", "trial", "free", "premium",
         "upgrade", "subscribe", "subscription",
+        # iOS settings / system UI vocabulary (F#18 — Apple Preferences labels)
+        "general", "privacy", "bluetooth", "wi-fi", "wifi", "notifications",
+        "sounds", "haptics", "focus", "screen", "time", "accessibility",
+        "siri", "safari", "maps", "health", "wallet", "facetime", "photos",
+        "camera", "messages", "mail", "calendar", "contacts", "reminders",
+        "notes", "icloud", "itunes", "store", "appstore", "airdrop", "airplay",
+        "display", "brightness", "battery", "storage", "privacy", "security",
+        "passcode", "touchid", "faceid", "cellular", "vpn", "hotspot",
+        "language", "region", "keyboard", "reset", "update", "software",
         # common content nouns / verbs that show up in titles & cells
         "dance", "partner", "story", "tale", "world", "people", "person",
         "place", "thing", "year", "day", "way", "man", "woman", "child",
@@ -190,6 +199,11 @@ class Mark:
     raw_confidence: Optional[float] = None
     # `confidence_band` is the dictionary-gated quality bucket. None = compute lazily.
     _band: Optional[str] = field(default=None, repr=False)
+    # F#4 — b5: alternate OCR readings for this element seen across consecutive
+    # observations. Populated by the OCR smoothing layer when consecutive observes
+    # produce different text for the same spatial region. Defaults to empty list;
+    # callers may set this after construction.
+    alternates: list = field(default_factory=list)
 
     def __post_init__(self) -> None:
         # If callers constructed a Mark with only `confidence`, that value is
@@ -276,6 +290,8 @@ def to_dict(self) -> dict:
             "raw_confidence": round(float(self.raw_confidence or 0.0), 3),
             # `confidence_band` is the human-readable quality bucket.
             "confidence_band": self.confidence_band,
+            # F#4 — alternate OCR readings seen across consecutive observations.
+            "alternates": list(self.alternates),
         }
 
     def to_compact_dict(self) -> dict:

simdrive/tests/test_a12_marks_parity.py

@@ -30,6 +30,7 @@
     "confidence",
     "raw_confidence",
     "confidence_band",
+    "alternates",  # F#4 (b5): OCR alternate readings field
 })
 
 # ── Minimal 1×1 PNG (PIL-readable) ───────────────────────────────────────────