Strengthen hidden overlap detection with anti-autocomplete and proximity checks

LLM was autocompleting partial words and reading them as full text, missing
the hidden overlap. New approach: explicit "DO NOT AUTOCOMPLETE" instruction,
character-level boundary check (what background is each character on),
spatial proximity check (text touching product = fail regardless), and
concrete example using the actual test case ("para" where "p" is hidden
on dark purple product appears as "ara").

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/visual_qc_apps/text_product_overlap/app.py

@@ -71,17 +71,32 @@ CALLOUT TEXT — lenient check:
 - However, callout text that directly OVERLAPS or TOUCHES the product hero zone (text sitting on top of the product imagery, text squeezed into tight space where the product edge is immediately adjacent with no clear separation) is still a FAIL. Having a connector line does not excuse the text physically overlapping the product.
 - The key test: is the callout text surrounded by clear/clean space, or is it crammed against the product imagery?
 
-=== STEP 4: HIDDEN OVERLAP DETECTION — INCOMPLETE WORDS ===
+=== STEP 4: HIDDEN OVERLAP DETECTION — TEXT AT THE PRODUCT EDGE ===
 
-This step catches overlaps where the text colour matches the product colour, making the overlapping letters INVISIBLE. The LLM cannot see a black letter on a dark product, but it CAN detect that a word is incomplete.
+CRITICAL: This step catches a specific failure mode that is easy to miss. When dark text is positioned next to a dark product (e.g., black text adjacent to a dark purple bottle), any letters that cross onto the product surface become INVISIBLE — they blend into the product colour. You will NOT be able to see these hidden letters. Instead, you must use INDIRECT EVIDENCE to detect the overlap.
 
-For every text block positioned adjacent to or near the product hero zone:
-1. Read each word carefully. Does every word appear COMPLETE — starting and ending with expected letters that form a real word?
-2. If a word appears TRUNCATED, MISSING ITS FIRST OR LAST LETTER(S), or has a gap where a character should be, AND the missing portion would be positioned over/against the product hero zone — this is strong evidence of hidden text-product overlap.
-3. Example: Reading "ragrância" near the product edge instead of "Fragrância" — the missing "F" is hidden behind the dark product because the black text blends into the dark purple surface. The incomplete word PROVES the overlap exists.
-4. Also check for words that look like they START mid-word (e.g., "ech" instead of "tech", "ara" instead of "para") — these suggest the beginning of the word is hidden behind the product.
+WARNING — DO NOT AUTOCOMPLETE: When you read text near the product edge, your natural language ability will try to "fill in" missing letters and read complete words. You MUST resist this. Look at what is ACTUALLY VISIBLE as rendered pixels, not what the word "should" say.
 
-This counts as a FAIL even though the overlapping letter cannot be visually detected — the incomplete word is definitive proof that text extends into the product zone. Score 3-4 for this type of hidden overlap.
+DETECTION METHOD — For every text block adjacent to the product hero zone, perform these checks:
+
+A) CHARACTER-LEVEL BOUNDARY CHECK:
+   - Find the text characters that are CLOSEST to the product hero zone edge.
+   - Look at the ACTUAL BACKGROUND behind those characters. Are they on the white/light background, or are they at the transition point where the background becomes the dark product surface?
+   - If the closest characters to the product are positioned AT or BEYOND the boundary where the background changes from light to dark (i.e., where the product surface begins), those characters may be invisible.
+
+B) WORD COMPLETENESS CHECK:
+   - For each word in a text block near the product, check: does the word START with a visible, clear character on a contrasting background? Does it END with one?
+   - Check EVERY LINE of multi-line text blocks, not just the first line. A word on the SECOND or THIRD line may be the one overlapping the product.
+   - A word that appears to begin with a lowercase letter when it should be capitalised (e.g., "ragrância" instead of "Fragrância") suggests the capital first letter is hidden on the dark product.
+   - A word that doesn't form a recognisable word but would if you added a letter on the product side (e.g., "ara" where the word should be "para" — the "p" is hidden on the dark product surface) is evidence of hidden overlap.
+   - CONCRETE EXAMPLE: A text block reads "Fragrância tech / para maximizar a / eficácia." positioned to the right of a dark purple product. The "p" in "para" on the second line sits on the dark purple surface and is invisible — the visible text appears to read "ara maximizar a". This is a FAIL.
+   - Spell out each word in the text block closest to the product. Does every character appear visually clear and on a contrasting background?
+
+C) SPATIAL PROXIMITY CHECK:
+   - If a text block is positioned so its edge is RIGHT AT the product hero zone boundary (touching or within a few pixels), it is highly likely that at least one character crosses onto the product surface. Even if you think you can read the full word, the physical proximity means overlap is almost certain.
+   - A text block whose edge TOUCHES the product hero zone = FAIL, even if you believe you can read every word.
+
+If ANY of these checks indicate hidden overlap, score 3-4 and report the specific evidence in incomplete_words_detected.
 
 === STEP 5: HEADLINE WIDTH CHECK (CRITICAL) ===