docs: qualify memory source proposals and benchmark provenance - #831

Scope

Qualifies #828's current memory/RAG catalog as proposed source-review input and adds a dated provenance amendment. The original review observations, decided text, eighteen repository pins/dispositions and five profile memberships are preserved.

• Base commit: c9e3597b176c247585392aa22cc23dbb748a3441
• Head: 663702741d8019cc9b091735f2ff7fbfccb7c706
• Lane: lane:foundation
• Owned paths: the memory/RAG catalog, catalog README, appended decision amendment, two new qualification artifacts and their evidence-registry bindings.
• Landing path: co-op exact-head read of findings F1–F3 → command-center ACK with CI at that head → the command center's queue, landed by 5f. This is the explicitly assigned catalog: finalize memory and RAG upstream stack recommendations #828 follow-up, exempt from the lane's two-unit cap; it stays draft until that read.

Exactly three findings addressed:

1. F1 (P2): explicit proposed status, recommendation-only responsibilities, no host memory-owner selection and no preselected memory-h2h winner. The current carriers are edited; the old decision is qualified by an appended dated amendment.
2. F2 (P2): the historical PageIndex benchmark revision remains unrecovered. A new dated check pins the separate benchmark README, labels its counts and exclusions as vendor descriptions and makes all eighteen inspected component locators direct in the amendment.
3. F3 (P3): installation and client smoke are explicitly not performed in this review; organic use is not observed. “Tested” evaluation wording names the upstream CI actor.

SOTA sources

• Original source review and catalog at the landed #828 commit, including its eighteen immutable primary-source locators. This follow-up preserves that record and its failure observations.
• VectifyAI/PageIndex-OSS-Benchmark@ad4c0b92970a6f4801f09ff2e647389e8f5874fa:README.md, with the count inventory. This new check is not a recovered historical pin or an executed benchmark.
• Repository evidence policy@c9e3597b176c247585392aa22cc23dbb748a3441:docs/acceptance-evidence-policy.md and hot-file protocol: source review, native acceptance and structural validation remain separate; registry committed last.

Evidence-class table

Local commands run

python3 scripts/catalog_decisions.py --write
exit 0; identity union and typed pointers validated; generated index unchanged

python3 -c 'import sys; from pathlib import Path; sys.path.insert(0, "scripts"); import host_receipts
for p in sys.argv[1:]: host_receipts.register_file(Path("."), p)' <five owned files>
exit 0

python3 -m unittest tests.test_catalog_decisions
exit 0; 14 tests passed

python3 scripts/validate.py
exit 0; status passed, 10,490 hashed files

git diff --check
exit 0

Commands ran with nice 19 and idle I/O; test scratch and logs are durable private lane state. The registry is the last commit. No rebase was performed.

Decision record

docs/decisions/2026-10-07-memory-rag-source-stacks.md, appended Amendment (2026-10-07). The separately owned memory-h2h result is added as an explicit possible overturn source; this PR makes no new selection.

Host evidence

No host receipt or deployed configuration is changed.

Checklist

• No workflows, dependency locks or installation commands changed.
• No credentials, raw conversations or private host paths committed.
• No new paid hosting or billing surface introduced.
• Peer-owned worktrees and untracked files preserved.
• Exactly F1, F2 and F3 addressed; source review is not promoted into adoption.