More accurate variant selection in search results, especially when the product itself already matches the query.

Variant matching now compares each variant's relevance against the product as a whole, instead of comparing variants against each other. The system only selects a specific variant when it genuinely adds relevance beyond what the product already provides.

• Smarter selection: When a product is already a strong match for the search query, the default variant is preserved instead of swapping to an arbitrary one. Previously, small differences between variant embeddings could cause unnecessary variant switches.
• Fewer false positives: Interchangeable variants like "8oz" and "16oz" are no longer swapped unnecessarily when the query doesn't express a variant preference. The default variant is preserved unless a specific variant is meaningfully more relevant to the search.
• Intent-driven overrides: When a search query does target a specific variant, for example searching for a theme or style that only one variant's imagery represents, that variant is still surfaced as expected.

The full variant selection priority remains the same: explicit filters first, then keyword matches, then pin defaults, then confident vector matches, then the default variant by position. This change improves how "confident" is determined in the semantic matching step.