[jstar0]

Codex now compresses every rollout that has been idle for seven days into a sibling rollout-*.jsonl.zst and deletes the .jsonl. That pulled sessions out of the store one by one. On a real machine, the saved figure fell from 69% to 38%, and 3,623 .zst files held 24.2 GB that CodexFold never touched. This branch has the service fold those files too, keeps folding working when the disk is nearly full, and fixes three production faults found while running it.

Folding Codex .zst files

Each settled .zst is folded in place:

• Byte-exact proof first. libzstd 1.5.7 (the version Codex links, via cgo github.com/DataDog/zstd) must re-encode the decoded content to the file's exact bytes.
• Space is released without changing the file. The data blocks are hole-punched and a marker attribute is set. Path, inode, size, times and attributes stay.
• The mount serves the original bytes. It re-encodes them through one shared encoder per file, refuses writes, and returns EIO rather than empty blocks if it cannot.
• Only when it pays. A dry run prices each file first, and a file is folded only when that costs under 80% of its .zst.
• fs native-zst status and fs native-zst restore --all --apply are added. Run the restore before installing a build without this support.

Folding on a tight disk

• Streamed .zst folds. A .zst is folded straight from its decoder instead of staging its decoded content.
• Exact dry-run pricing. When space is short, the planner and fold --apply measure the real cost instead of assuming three times the raw size. Prices are cached in enrollment/prices.json.
• Selective pack compaction. A pack build rewrites only the old pack files worth rewriting, within free space, and hard-links the rest.
• Better pack compression. Pack builds entropy-code small blocks, which saves about 2.1 GB on a real 14.8 GB pack.

Production fixes

• "Too many open files" on every mount read.
  • Cause: the FSKit extension runs at launchd's 256-descriptor limit. A folded file was handed a shared read window that primed nine prefetch channels, about 14 descriptors per open file.
  • Fix: folded files no longer get that window. The extension raises its descriptor limit at start and caps prefetch channels at 48 across all files. App build is bumped to 117.
• The same .zst re-punched every 30 s. APFS kept a few extents of one punched file allocated, so it was proved and punched again on every pass. It is now left alone until it changes.
• Automatic enrollment stalled, and service start rolled back.
  • Cause: a deleted managed session's native snapshot is hidden by the mount on purpose, but the readiness probe waited for it. Enrollment reported mount-warming for 1,626 sessions and enrolled nothing.
  • Fix: the probe skips deletion-tombstone routes.
  • A failed native FSKit start also left the mount attached with no backend. It now reclaims the mount the way an explicit stop does.

Validation

• go vet, Go tests for cli, mountfs, nativezst, vfs, enroll, pack and codexzstd, and Windows and Linux cross-builds.
• Swift ReadCacheTests, and a type check of the extension.
• Deployed on the maintainer's Mac and checked through the real mount:
  • every folded file reads back byte-identical;
  • 1,200 files held open at once with no error (1,438 extension descriptors);
  • 20 concurrent folded reads peaked at 64 descriptors.

Windows

• The new .zst code compiles for Windows, but Windows does not fold .zst files yet: in-place folding needs hole punching and marker attributes.
• Windows CI covers compilation and adapter tests on this branch. It has not been run against a real WinFsp drive.