Tests the retrieval and context guarantees from spec/retrieval.md and spec/context.md. Requires the core capability.

R1. Pagination is deterministic

Given: 25 stored records in one scope When: retrieval runs with limit=10, offset=0, then limit=10, offset=10 Then: the two pages are disjoint and together cover 20 distinct records.

R2. Total counts before limit

Given: 25 stored records When: retrieval runs with limit=5 Then: total is 25 and truncated is true.

R3. Offset beyond total is empty, not an error

Given: 5 stored records When: retrieval runs with offset=100 Then: records is empty, total is 5, and no error is returned.

R4. Type filter returns only matching types

Given: a mix of fact and decision records When: retrieval runs with types:["decision"] Then: only decision records are returned.

R5. Context assembly carries scope

Given: records retrieved from scope {project: "x"} When: a context is assembled Then: context.scope equals {project: "x"}.

R6. Context assembly carries provenance

Given: retrieved records with provenance When: a context is assembled Then: each record in context.records retains its provenance, and the context itself has assembly provenance.agent.

R7. Empty context is valid

Given: a query matching nothing When: a context is assembled from it Then: a context with records: [] is returned — not an error.

R8. Metadata round-trips through retrieval

Given: a record with a custom metadata object When: stored then retrieved Then: metadata is returned unchanged.

R9. Tags filter

Given: records tagged infra and ui When: retrieval runs with tags:["infra"] Then: only infra-tagged records are returned (per the implementation’s documented AND/OR semantics).

R10. Limit cap is discoverable

Given: an implementation with a max retrieve limit When: capability discovery is invoked Then: limits.max_records_per_retrieve is reported (or the limit is unbounded and documented).