Answers the question "can memory-bank/ be dropped?" with: yes, but not by deleting it, because one thing in it was both unique and more current than its public counterpart.

What was in there

memory-bank/ was a second documentation tree that nothing depended on: no mkdocs.yml nav entry, no sdist entry, no tox.ini or CI reference, no .claude/ reference. Only CLAUDE.md and AGENTS.md pointed at it. It had already shrunk from ten files to two, and that pass also had to delete six stale files and correct a skip count that had drifted. #800 closed on the same premise: stale context in these files misdirects work.

Auditing what was left, everything was covered elsewhere:

The part worth keeping

in_depth/property-mapping.md calls itself the single source of truth for the option model, but it still said materialization happens "centrally at the compute boundary" and that it "never changes matching or FeatureSet splitting". The compute boundary is now the idempotent safety net, not the primary site, and the splitting half is false: Engine.add_feature_to_collection rebinds options as a feature enters the plan, precisely so default-equivalent twins canonicalize before splitting. The memory bank described this correctly, so the public page was the stale one.

That section is corrected, the declared-vs-effective stage table moves in beside it, and two tests that already pin this behavior join the invariant map. The corrected text also spells out three things the page never stated: that the engine warns on a default-equivalent merge and that the survivor's dependency declaration comes from the first-listed request, that the direct FeatureSet path raises rather than merging, and that filter matching calls match_feature_group_criteria again after intake and therefore sees effective options where resolution saw declared ones.

Then the directory goes. CLAUDE.md and AGENTS.md trade their Memory Bank section for a pointer to the documentation that is actually maintained, including the in-tree READMEs, and the transient phased-work todo.md moves to the repository root and is gitignored.

Validation

tox green: 7374 passed, 170 skipped, plus ruff format, ruff check, licenses, mypy --strict, bandit. The python314 env, which is the only one that executes documentation tests, was run separately. The documentation additions are prose and a table, so no new executable code block enters the mktestdocs suite.