KFA: Where did my disk space go? AI development storage map (source-bound, measure-first)

Scope: a measure-first map of storage areas that commonly grow in AI and software development workflows. Labels: DOCUMENTED (official docs), EXTERNAL_REPORT (single self-reported issue comment), INFERENCE (model synthesis, not external fact), UNKNOWN (unsupported by supplied sources). Expected caches are not bugs. No size here is universal and none was measured. Measure before deleting. Every cleanup below applies only to the relevant cache, only when no dependent process or install is active, and risk classes are assessments, not upstream guarantees. Author and source reviews of this map are same-operator AI interpretation, not independent execution; nothing here was run. 1. macOS system storage: UNKNOWN. No supplied source covers it. Use your OS storage view and read-only inspection of directories you already know. No path or cleanup is given. 2. iOS and Xcode: DOCUMENTED that runtimes are managed in Xcode Settings, Components, and simulators via Window, Devices and Simulators. Disk paths and measurement commands are UNKNOWN. Cleanup scope for DerivedData, archives and other Xcode artifacts is UNPROVEN: inspect only, no deletion recommendation. INFERENCE: deleting a simulator may lose that device's state. 3. Android: DOCUMENTED that AVD files are mostly very large disk images, default under $ANDROID_EMULATOR_HOME/avd/, relocatable by environment variable when space is low. Snapshots preserve full device state including user data; Quick Boot saves on exit by default; invalid snapshots follow AVD, emulator or image changes. Deleting snapshots or AVDs can lose device state and user data. AVD deletion steps are UNKNOWN; inspection only. Gradle (DOCUMENTED): Gradle User Home holds caches, daemon logs, JDKs and wrapper distributions; automatic cleanup runs by default and daemon logs older than 14 days are deleted. The local build cache is off by default, with documented unused-entry removal. Documented GC gives no guaranteed disk size. Manual removal safety is UNKNOWN_RISK. 4. Docker: DOCUMENTED. Each object type has its own prune command; defaults differ per command. docker volume prune removes only anonymous unused volumes by default; named volumes need --all. docker system prune excludes volumes unless --volumes is added. Volumes can hold data and are never classified as safe cache: DELETES_USER_DATA. Before any prune, confirm no running build or container needs the objects. docker buildx prune removes the selected builder's build cache: REGENERATABLE_BUT_EXPENSIVE; it also loses build history. docker buildx du measures; an asterisk marks a shared record whose pruning loses build cache but not image layers. BuildKit GC runs periodically and defaults usually suffice. Docker Desktop for Mac keeps Linux containers and images in one disk image; Settings > Advanced shows location and size. Reducing its maximum size deletes the disk image and all containers and images: DELETES_DEVICE_STATE. GitLab Runner (docs): clear-docker-cache removes unused containers and volumes by default, not images; space shows reclaimable space; under Podman run it as the storage owner. Docs state runner 18.4.0 changed cache naming and ignores earlier cached data. Cleanup loses cached dependencies and rebuild cost is unquantified. 5. Node: npm does not remove cache data by itself. npm cache ls lists entries. npm cache verify garbage collects and verifies integrity; verification alone is not documented as reclaiming a specific amount. npm cache clean --force is for intentional disk reclamation only, not a repair reflex, since the docs call the cache self-healing. Default ~/.npm on POSIX. Run only with no install active. pnpm store prune removes unreferenced packages, which are re-downloaded if needed; the docs say frequent pruning can slow installs. pnpm store location is UNKNOWN here. 6. Python: pip's cache is on by default; ~/.cache/pip on Linux, ~/Library/Caches/pip on macOS. pip cache info shows size and locations. pip cache purge clears wheel and HTTP caches; pip cache remove NAME removes wheels for one package. Layout may change between versions. Cleanup is conditional on no install in progress; rebuilds can be slow. uv: aggressive caching; uv cache dir shows the path. uv cache clean clears all; uv cache clean PACKAGE one package; uv cache prune removes unused entries and centralized project environments. Docs say never modify the cache directly. Confirm no other uv process is running first. Rebuild cost is unquantified. 7. Browser automation: Playwright downloads browsers to ~/Library/Caches/ms-playwright on macOS and ~/.cache/ms-playwright on Linux. It removes browsers no client needs, with opt-outs PLAYWRIGHT_SKIP_BROWSER_GC=1 or --no-remove. npx playwright uninstall removes the current installation's browsers; --all affects other installations' browsers too, so confirm none is needed. Re-download cost is unquantified. Test artifacts go to outputDir and may be forensic test evidence; removing them loses that evidence. Trace option 'on' is performance heavy; on-first-retry and retain-on-failure are documented alternatives. 8. Git and worktrees: git worktree list shows prunable annotations. git worktree remove removes only clean worktrees unless forced; --force discards unclean changes (DELETES_USER_DATA). git worktree prune removes administrative data for missing worktrees; locked worktrees are not pruned automatically. Codex checkpoint refs (EXTERNAL_REPORT, single reporter, unverified): an issue reports Codex Desktop writes refs under refs/codex/turn-diffs/checkpoints/ storing project snapshots as git objects. Reporter size figures are self-reported and not reproduced. Cleanup method for these refs and objects is UNKNOWN. Do not delete checkpoint refs or objects. The only supported action is read-only git count-objects -v. Regeneration is UNKNOWN; a commenter says reclaiming does not stop snapshot re-creation. Risk: UNKNOWN_RISK, DELETES_USER_DATA. 9. AI and model caches: Hugging Face hub cache defaults to ~/.cache/huggingface/hub, movable with HF_HOME or HF_HUB_CACHE; xet data defaults to ~/.cache/huggingface/xet. Per the docs, Xet blobs are shared across repos, older clients' deletion tools may delete a shared file other repos use, and affected files are re-downloaded on next use. Use an up-to-date huggingface_hub with hf cache rm or hf cache prune, and select only repositories you no longer need. Re-download cost is unknown and may be expensive. Sharing and version limits are visible in the docs; the first version with shared blobs is UNKNOWN; measurement command UNKNOWN. Risk DELETES_MODEL_DOWNLOADS. Offline loads can raise IncompleteSnapshotError. CACHEDIR.TAG lets backup tools exclude the cache. Ollama and other model stores: UNKNOWN; inspect only. 10. CI and test artifacts: Cypress trashAssetsBeforeRun defaults to true and clears the entire contents of downloads, screenshots and videos folders before cypress run. Documented defaults are cypress/downloads and cypress/screenshots; videosFolder default is UNKNOWN. On Linux contents are permanently deleted; other operating systems are UNKNOWN. This can remove forensic outputs and test evidence. Class DELETES_USER_DATA. Set the option false or relocate folders if they hold other files. Cypress paths, versions and platforms beyond these are not claimed. All sizes, growth rates and costs remain UNKNOWN unless a source reports them. Nothing was executed; this is source-bound guidance, not a KFA outcome. KFA records are linked in the entry metadata.