AI: TypeSafe judges ad and tracker hosts; smart capture asks about DNS-only hosts

[Neoo-Blue]

What

TypeSafe domain judge. Smart capture and the domain tester's "Ask the assistant" can ask TypeSafe's Jev model (System One API) two typed questions per host: is it ad or tracking infrastructure (noul), and how likely is a DNS block to break something (choice: low / medium / high). It works with no chat model configured. When a chat model is configured, it takes over if TypeSafe fails. Chat, briefs and reviews are unchanged.

• internal/ai/typesafe.go: client, questions, verdict mapping. Retries 429/5xx with backoff; a bad key stops the whole batch after one request.
• Config ai.typesafe.{enabled,api_key}: masked in the API, kept on backup restore, and settable from Settings > Assistant > TypeSafe domain judge.
• A 40% or higher probability of "high" breakage is reported as high, so the host goes to a person instead of being blocked automatically.

Smart capture on DNS-only nodes. The heuristics measure referrers, response sizes and paths, so a host seen only as DNS lookups scored 0 and never reached the model. On the reference Pi node, 11,794 of 11,839 candidates were at 0.0 and nothing had ever been judged.

• DNS-only candidates are escalated once, while their status is candidate.
• The verdict stands on its own instead of being blended with that 0, and it is capped below auto_block_score. A verdict on the name alone can only send a host to review, never block it.
• Review rows with an AI verdict and nothing new to measure keep that verdict for a person.
• PendingCandidates puts unjudged rows ahead of review rows, so a review backlog cannot starve the 500-row window.
• A quiet interval no longer erases a host's earlier HTTP evidence (withStoredHTTP).
• classify() uses errors.As, so wrapped provider errors are recognised.

Behaviour change for nodes without TypeSafe

With smart_capture.use_ai on, DNS-only candidates now reach the chat model too: one fast-model call per pass (40 hosts) while the backlog lasts. The judge prompt now says that a third_party_ratio of 0 with no HTTP evidence means "not measured", not "first-party".

Testing

• go vet ./... and go test -race ./... pass locally.
• New internal/ai/typesafe_test.go (httptest) covers verdict mapping, a DNS-only state carrying no zeros, the heuristic score staying out of the state, and a bad key aborting after one request.
• New internal/adblock/smart_test.go covers DNS-only escalation, the capped final score, review verdicts kept across passes, and HTTP history surviving a quiet pass. Disabling withStoredHTTP makes it fail.
• Live run over 120 real hostnames from the Pi: AppsFlyer 0.91, Crashlytics 0.85, AdBlock telemetry 0.81; first-party services and status pages below 0.5.
• Deployed to the Pi as v1.32.1-typesafe. The first pass logged judge: 40/40 TypeSafe verdicts (2 to review, 38 dismissed, 0 blocked).

🤖 Generated with Claude Code