YantrikDB

Two additive engine follow-ups to the v0.7.21 link model and the trader orphan postmortem. No schema change.

record_with_links_partial()

Per-link-outcome variant of record_with_links(), for callers (the MCP remember-with-links surface) that need to know which links landed without re-querying. The record commits first (durable via the oplog); each link is then attempted independently — a failure is captured, not propagated, so one bad link doesn't abort the rest or fail the call.

Returns (rid, [LinkResult]) where each LinkResult is:

• Inserted — new link row written
• AlreadyExists — the idempotent UNIQUE(source,target,type) hit (distinguished for telemetry; retry-safe)
• Failed { error } — this link failed; the record is still durable

Python: db.record_with_links_partial(...) returns (rid, [{result, target_rid, link_type, error?}]).

audit_leak_candidates()

Replaces the compaction-confused "orphan" metric. The trader postmortem proved the old definition (memories lacking oplog presence) is a benign oplog-compaction artifact — locally-originated memories whose oplog rows aged out of the retention window look orphan-shaped while being perfectly healthy.

The windowed check only flags memories created within the surviving oplog window (created_at >= MIN(oplog.timestamp)) that still lack both an oplog op and a replication_apply_log entry — where the oplog row should still exist, so its absence is a genuine signal (a live write-path bug, or a direct-SQL insert bypassing the engine). Outside the window, absence is expected compaction and is not counted. Zero new storage.

Returns LeakAuditReport { window_floor, candidate_count, candidate_rids }. Python: db.audit_leak_candidates(max_rids=100).

Stats

1534/1534 lib tests pass. cargo fmt --check --all + cargo clippy --all-targets clean. Python bindings verified locally against a real interpreter.

Upgrade notes

Purely additive — no schema migration, no behavior change to existing paths. Safe to adopt directly.

🤖 Generated with Claude Code