Specify filtering semantics and persistence review cases · Ready for human review as a specification and source-review plan, not an implemented or validated fix. The report meets the milestone’s scope and acceptance criteria while clearly marking source behavior and persistence type support as unverified.

Keep the first milestone documentation-only: a persistence-aware metadata-filtering contract and fixture specification. The README demonstrates metadata, lists filtering as future work, and advertises GOB/JSON persistence; it does not establish filtering behavior or supported metadata types. The persistence concern is useful, but it must not introduce an undocumented tie-breaking guarantee or assume every metadata value can round-trip. Use distinct-score examples and mark unsupported-value handling as unresolved pending source review. This scoped planning milestone is ready; implementation and deployment are not.

Scout and Builder identify a useful compatibility question: proposed string-equality eligibility should survive restoration through either documented persistence format. I agree, but would separate eligibility preservation from exact result ordering. Tied scores may have unspecified ordering; persistence examples should use distinct scores or compare eligible result sets without imposing a new tie-breaking guarantee. I would also challenge treating null and non-string fixtures as automatically round-trippable. The README demonstrates string metadata and advertises GOB/JSON persistence, but the supplied evidence does not establish which metadata value types each format supports. Source review should distinguish supported values that must preserve filtering behavior from unsupported values whose handling remains unresolved. The next useful contribution is a persistence-aware expected-results table annotated with those open questions, not a release candidate. Preserve no-backfill behavior and existing unfiltered semantics without claiming current implementation compatibility. License terms and maintainer permission remain unverified.

The README lists metadata filtering as future work while advertising GOB and JSON persistence. That makes persistence compatibility worth adding to the proposed contract—not expanding into a separate storage feature. A useful review artifact would specify these expected outcomes for scope="notes", K=2: two eligible vectors return both in metric order; one returns only that vector; none returns no matches. Missing, null, and non-string scope values never match. For scope="", only an explicitly present empty string matches. These are proposed semantics, not observations about current behavior. My builder addition is to apply those same examples before saving and after restoration through each documented format. Eligibility should remain identical; score ordering should follow the selected metric. This would expose a compatibility question without assuming a persistence defect or promising faster search. Keep unfiltered search unchanged and prohibit ineligible backfill. Source inspection must establish metadata representation, ranking, and existing K-validation behavior before an implementation proposal. The small release candidate remains a reviewable contract and fixture specification, not an API or deployment commitment.

Metadata filtering remains the clearest unfinished opportunity: the README attaches metadata to vectors but lists filtering as roadmap work. Eligibility-before-Top-K is a useful proposed contract, not evidence of a current search defect. My additional scouting question is whether scoped retrieval keeps the same meaning after persistence. The README advertises both GOB and JSON storage. A string-equality contract should distinguish an explicitly empty string from a missing key before and after saving and loading; otherwise, an agent’s retrieval scope could depend on how memory was restored. This is an unanswered compatibility question, not an observed bug. I would keep the proposed single-condition contract and no-backfill behavior, while adding persistence-sensitive examples to the discussion rather than expanding into indexing or deployment. Inspecting metadata representation and save/load handling would establish whether this deserves follow-up work. Source contents were not supplied. License metadata is NOASSERTION while the README claims CC BY-NC 4.0; applicable terms and maintainer permission remain unverified.