atperson explicit ingestion policy

atperson

#20: explicit public-data ingestion policy

The network path learned text from the authenticated account's public home timeline, but nothing defined which fetched records are eligible. Without an explicit policy, a future daemon could quietly broaden the learning surface through runtime code rather than a reviewable design decision.

src/app/ingestion_policy.{hpp,cpp} is now that decision point. It consumes already-extracted fields — record type, author DID, text, embed type, viewer state, feed reason — and produces a decision: eligible, or skipped with a machine-readable reason. Protocol mechanics stay in Wolfram and the client; the policy never touches the wire.

Rule order is deliberate:

1. Record type: only app.bsky.feed.post records are supported; anything else is unsupported-record rather than mis-parsed.
2. Self-observation: the account's own output is never learned from (self-authored). Explicit exclusion, not accident — learning from self-authored records would create a feedback loop between the entity's output and its experience.
3. Moderation and relationship state: viewer-blocked, viewer-blocked-by, viewer-muted, and moderation-filtered posts are skipped. The account chose not to see that content; atperson respects that choice.
4. Text presence: empty text is empty-text; image/video-only posts are non-text-only.

Replies and reposts remain eligible and are tagged with their reason. Text is text.

The ledger semantics matter as much as the rules: a policy-skipped item is not silently dropped. It still flows through the ledger as an observation with outcome SKIPPED, so observed-but-not-learned stays distinguishable from never-fetched. Replays are idempotent for skipped items too — the existing (source id + digest) dedup index counts a replayed skip as a duplicate, not a fresh skip. One consequence worth noting: a future policy change that reclassifies a skip class as eligible can't retroactively learn already-committed SKIPPED entries without an outcome patch or rebuild-from-ledger — the same open problem as unlearning a committed contribution.

Tests: 15 policy assertions covering every skip class plus precedence (self-authored beats viewer state; record type beats everything; a degenerate empty account DID never self-skips), and three sync-engine integration tests (skipped items ledgered not trained, skipped items deduplicate across replays, reposts/replies learned with reason tags). 13/13 tests in both core-only and network builds. Issue #20 closed; roadmap #29 ticked.