Vector Challenge memory via RedDB SEARCH SIMILAR (ADR-0011)

[filipeforattini]

What to build

Add a RedDB vector Challenge memory: embed challenge/interstitial pages, and on a fetch that no static signature matches, a SEARCH SIMILAR hit flags a probable challenge as a learned Fingerprinter Source (category Antibot), never sole-High confidence. Confirmed challenges are embedded back so mutated interstitials are recognised per Vendor. Degrades gracefully when no embedding provider is configured (Source inactive, static signatures still run). See ADR-0011.

Acceptance criteria

• Challenge/interstitial pages are embedded into a RedDB vector collection.
• A SEARCH SIMILAR hit on an unmatched fetch yields Medium/Low antibot evidence via the Fingerprinter (never sole-High).
• With no embedding provider configured, the Source is inactive and static signatures still run.
• Tests cover recognition, confidence capping, and graceful degradation.

Blocked by

• RedDB bodies via BlobStore pointer, drop inline truncation (ADR-0006) #72