[liquidsec]

The protection classifier's vendor rules were hard-coded in src/antibot.rs, so fixing a false positive meant a rebuild and a release. They now live in antibot_rules.json, and the code only runs them.

Rules as data

• antibot_rules.json holds an ordered list of rules: a vendor, an outcome (challenge, blocked, present), an optional on_refusal outcome, and a when condition built from status, header_present, header_contains, cookie_set, cookie_prefix and body_contains, combined with all, any and not. The first match wins, as before. Rules keep their reasoning in an optional why field, since JSON has no comments.
• The file is compiled in as the default. blasthttp.set_antibot_rules(path) (antibot::set_rules in Rust) replaces the rules for the process at runtime, and set_antibot_rules() restores the built-in ones. A malformed file is refused when it loads.
• Vendor is now a name rather than an enum, so a rules file can add a product. Python callers already received vendors as strings and are unaffected.
• Behaviour is unchanged: every existing classifier test passes against the JSON rules.

Imperva
_Incapsula_Resource anywhere in a body counted as a challenge. Imperva appends its loader script (_Incapsula_Resource?SWJIYLWA=…&ns=1&cb=…) to pages it let through, including the origin's own error pages, so ordinary pages were judged needs_browser and sent to a browser that had nothing to solve. A challenge now needs Imperva's incident page (_Incapsula_Resource?CWUDNSAI=), or a reference to the resource that isn't the loader.