[ebigunso]

Completion cleanup of the scene plan (stacked on #121). The plan's standing condition is a value audit of every part at completion, review-driven additions included.

Removed or made smaller

• A test that built its state by writing beneath the facade: no public caller can make a source object go missing. The Missing report itself stays, since silently dropping a source would be a false claim.
• A description's list of matches. It was the raw nearest-neighbour list taken before any verification or floor: any words reach something in a non-empty store, so the list implied familiarity the library cannot assert, and after a failed vector delete it could name a forgotten memory. A description is now reported as used for a content cue; per-cue reach returns with measured floors.
• The rejection of a blank participant. A blank entry is the easiest value to build, and rejecting it cost the character the whole experience; it is dropped, on writes and retrievals alike, and everything else is stored as given.
• Positional pairing of precomputed embeddings with their objects. Pairing is by object type, id and surface; an id alone is not an identity, because an episode and an observation may share one (found by the internal review, with a regression through the public facade).
• Schema text that repeated the vector guide.

Added: the one behavior the slice was missing

• A participant named by key in a caller-built plan was stored but never linked, so meeting that person later did not bring the memory back; only remember derived the link. Commit now guarantees it, in the same graph batch, and derives nothing where a link already reaches that person, so remember's link set and the measured baselines do not move.

Evidence: internal Tier D review approved after one finding (the pairing key), confirmed with its own probe on the real stores. fmt, check, clippy with warnings denied and the full suite pass for the worker and independently for the reviewer.

🤖 Generated with Claude Code