UT GOTY retained overlay: identical-payload reuse and disk admission blocker

Storage improvement (later task, 2026-11-18)

Evidence: scratch/runs/20261118-overlay-identical-write-reuse-ut-install/. nodeDirStore now reuses an immutable blob only after an exact byte comparison. Candidates are grouped by stored payload length, with the current path checked first; cross-file and same-batch aliases work without changing the version-1 index format. Comparisons use bounded reads. Metadata remains per record, including attributes, timestamps, sparse layout, directories and whiteouts. Publication still writes new blobs before atomically replacing the index; collection and rollback retain all referenced aliases. Historical SHA-1 and UUID blob names need no migration. This improves storage capacity; it is not evidence of a guest or API defect.

A real 4 MiB filesystem control measured the old equal-overwrite-plus-alias publication peak at 12,582,912 allocated payload bytes (three blobs), settling at 8,388,608. The new store's peak and final allocation both stayed at 4,194,304 bytes (one blob). These are stat.blocks * 512 measurements, not logical file lengths or a deduplication estimate.

An authentic recovery fork copied the genuine 415-record index and hardlinked its immutable blobs. Two rounds of normal store re-persistence, in batches of 32 records with fresh store instances between rounds, preserved every payload and all metadata without allocating new payload blocks. The retained payload total is 448,673,130 bytes. Every resulting file still referenced its original base inode. The disposable fork was removed after verification. Both genuine partial overlays were SHA-256 verified before and after; neither index nor payload bytes changed. The unique CD1 ISO also remains unchanged and uncopied. This is filesystem recovery/storage evidence, not an installer execution or an invented installed tree.

All 33 overlay tests, the persistence-restore test, 12 publication tests and the locked build passed. The OPFS browser test skipped because Chrome was not found. The genuine Winamp installer regression passed installation, restart hydration and SIGTERM final flush using published Git sources in memory plus the owned store change and the clean a6002d5d… module. An earlier invocation refused before guest launch because the shared-tree region mirror differed from that module; its failed receipt is retained separately. Shared foreign runtime changes were not adopted into the successful regression graph.

UT was not launched. Admission measured 2,040,721,408 bytes free. Reserving the genuine default component page's 702,545,920 bytes leaves 1,338,175,488, which is 161,824,512 below the 1,500,000,000-byte floor before transient flushes, logs and drains. Equal-byte re-persistence proves reuse but does not bound Setup's different or intermediate writes; crediting all retained bytes as guaranteed installer savings would be unsupported. Safe capacity therefore remains unproved. Install 1200 / total 2400 / final 300 seconds were declared before admission; no UT clock started, old scope replay, endpoint request or extension occurred. Installation completion, game startup and ordinary Quit remain unverified. No new installer images exist. Box ID, expiry and stop receipt are null; state is NOT_CREATED.

Earlier read-only investigation (before payload reuse)

Evidence: scratch/runs/20261118-ut-goty-retained-overlay-install-game-start/. No new installer or original game process was launched. Installation completion, startup, gameplay and ordinary Quit remain unproved. Both predecessor runtime scopes remain closed UNKNOWN, with no endpoint request or replay.

Read-only verification found the larger partial overlay unchanged: 415 records, 395 files, 20 directories, 448,673,130 blob bytes, index SHA-256 c60cdcbe4f3e5bfbeec2b89dc6f254a38d2ad369d6de130dee39a97f12448811. Every referenced blob matches the prior retained size and SHA-256. AS-Frigate is a zero-byte persisted file despite the prior live 60% expansion observation. The older 16 retained index/blob artifacts were also hashed and checked against their original artifact manifest. Neither partial tree is an installed game.

Normal recovery is nodeDirStore plus VfsOverlay.hydrate before DLL discovery. The CLI accepts one writable --overlay-dir; it has no read-only overlay base or content-addressed blob reuse option. Each file write creates a UUID blob; the index is replaced atomically, then superseded blob names are unlinked. A tiny real-store probe established that an exact index copy with hardlinked immutable blobs can fork a partial filesystem without changing the base's index or bytes when the fork is subsequently written. The genuine partial filesystem was not forked or written. This probe proves filesystem recovery mechanics, not installer resume behavior or a reduction in Setup's output allocation. Per-file --vfs-mount was rejected as an exact recovery route because it synthesizes attributes and does not preserve directory/time/whiteout semantics.

A fork saves initial duplication, but ordinary Setup may replace existing files. The store does not deduplicate equal output; preserved base links keep old bytes allocated. All new batch blobs precede index publication and old-blob collection. A synchronous flush can therefore cross the floor before a separate poller can intervene. Crediting the existing 448 MB against new writable capacity would not establish safe expansion and flush.

At admission, actual free space was 1,974,247,424 bytes. The genuine prior default component page's conservative 670 MiB requirement is 702,545,920 bytes, leaving 1,271,701,504 bytes. This is 228,298,496 bytes below the required 1,500,000,000-byte floor, before any allowance for transient flush blobs, logs or concurrent writers. Fresh Setup and an overwrite-capable retained fork both fail this conservative admission. No media copy, foreign deletion, manual extraction into the guest, guest/configuration/registry/state patch, file omission or fabricated installation state was used. The bounded blocker is insufficient proven disk capacity, not a new installer failure.

Published Git 0b55b3c2f1be569f8799ee1390ebfd8c620cb39b compiled directly from Git objects to module SHA-256 a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8 (2,418,710 bytes). A memory-only Git source inventory and loopback HTTP byte checks exclude held shared-tree WIP. This was preparation, not a browser or native execution. Current authored wizard/progress helpers, native scene/progress guards and supervisor match published Git. Prior native/CPUWorker controls are inherited evidence; they were not rerun here. The prior original's 204a3300… module included the unarmed foreign timer observer and is not claimed equivalent to the clean module. Relevant worktree patches remained byte-identical during preparation. CD1 remained at its unique existing path, 649,633,792 bytes, SHA-256 e184984ca88f001c5ddd52035d76cd64e266e26c74975161b5ed72366c74704f.

The predeclared future run limits are install 1200 seconds, total 2400 seconds, and final Quit/cleanup reserve 300 seconds. No clock began and no deadline was extended. Before any admitted run, its independent disk guard must cover final overlay flush and all actual native/child/supervisor/enclosing waits and both EOFs; its terminal markers must differ from action receipts. A new admission must account for transient flush capacity as well as expansion. No runtime repair was authored, so no new old-fail/new-pass runtime claim or code build gate applies. Only this note is published.

The board watcher has an actual child-close wait and both EOFs; the loopback server closed normally. There are no owned runtime children or boxes. Box ID, expiry and stop receipt are null; state is NOT_CREATED. No agent, worker, Python, Wine, public deployment or held-title action was used.