Local cache hygiene

Large local caches are expected during development and benchmark work. They are ignored by Git, but they should be reviewed before removal because some contain expensive or non-reproducible intermediate results.

Inventory

The repository audit found these local surfaces:

Path	Observed size	Regeneration and risk
target/	51 GB	Cargo build output; reproducible from the source and lockfile
site/	5.8 MB	MkDocs output; reproducible with the strict documentation build
../memd-bench/fable-work/	64 GB	External review workspace; remove only after its useful outputs are frozen elsewhere
../memd-bench/run-output/	1.1 GB	Benchmark runs; preserve until manifests and bundles have been verified
../memd-bench/benchmark-data/	116 MB	Datasets, source checkouts, and model/build metadata; some downloads are costly
../memd-bench/.ruff_cache/	40 KB	Linter cache; disposable

Sizes are a point-in-time audit, not limits. Recheck before cleanup:

du -sh target site \
  ../memd-bench/fable-work \
  ../memd-bench/run-output \
  ../memd-bench/benchmark-data \
  ../memd-bench/.ruff_cache 2>/dev/null

Reviewed cleanup commands

Run these commands manually from the memd repository root. Inspect the paths first; none is run automatically by the project.

# Reproducible Rust and documentation output.
cargo clean
rm -rf -- site

# Disposable benchmark linter cache.
rm -rf -- ../memd-bench/.ruff_cache

The following caches require an evidence check first:

# List immutable manifests and bundles before considering benchmark cleanup.
find ../memd-bench/run-output -type f \
  \( -name 'seed.*.json' -o -name 'retrieve.*.json' \
     -o -name 'qa.*.json' -o -name 'judge.*.json' \) -print
find ../memd-bench/benchmark-artifacts -maxdepth 2 -type f -print 2>/dev/null

Do not remove run-output/, benchmark-data/, or fable-work/ until every needed result is reproducible from a validated bundle. Never use git clean as a cache-management shortcut: it cannot distinguish disposable output from valuable ignored benchmark data.