Scraper: Filter out non-coffee products (cajik, tea, equipment, gift cards from discovery)

[peterhadac]

Problem

The scraper currently discovers and extracts every link it finds on a roaster's listing page. Unfortunately, many roasters sell non-coffee items alongside their coffee:

• cajik / čaj - herbal teas, fruit infusions, black/green tea
• Coffee equipment - brewers, grinders, scales, towels
• Gift cards / voucher codes
• Chocolate / confectionery - some roasters sell chocolate bars
• Flavorings / syrups - vanilla, caramel syrups for coffee

These clutter the table and make the data harder to navigate. 'cajik' entries appear in scrape results because the roaster's product grid mixes them with coffee - they have a valid origin and packaging but aren't coffee.

Proposed

1. Tighten product-link detection

Add a negative keyword list to the product-link discovery filter to block obviously non-coffee pages from being sent to the LLM:

• Czech/Slovak: cajík, cajíky, čaj, doplnok, darceká karta, zdravotencky, pytel, mlynček
• English: tea, gift card, voucher, syrup, equipment, brewer, towel, chocolate

Match against the URL slug/path segment AND the listing page link text when available.

2. LLM decline fallback (already partially working)

The extract_product tool is optional - the model can decline. But this wastes an LLM call.

3. Post-extraction sanity check

If the extracted product name strongly suggests non-coffee (e.g. 'green tea matcha', 'ceramic pour-over'), mark as not_a_product.

Implementation notes

• Start with a simple blocklist of keywords in a new filter function near the top of the discovery loop.
• Track how many items are filtered - should reduce unnecessary LLM calls and clean up product listings.