Skip to main content

destructive paths

Reviewed source: 5002438059bd747146f61e3f3763461fbc1968ca (2026-10-18). This inventory supports #234. It records source gates, failure behavior, tests, and evidence limits. It is not the execution review for #237. That issue still needs an RC SHA and a full-tier run.

Review method and status​

The review walked app/ and payload/, enumerated the relevant mutation primitives, and reconciled them with the tests and field-report entry points. The pinned fisherman submodule has a separate row. This checkout does not initialize submodules. Its selected-data path follows the pinned commit, not a branch.

For each mutation, the inventory records:

  • the precondition and the second confirmation or observation before mutation;
  • what remains reversible, retained, or retryable after a failure; and
  • the test or live evidence and the boundary that evidence does not cross.

Open marks an unresolved item, not clearance. A row turns green only after its gate and reversal pass on the release candidate. Temporary files, mocked commands, or an empty issue search do not prove user-data safety.

The table includes product data deletion inside the storage directory because it can destroy user work even when no external path changes.

Operation and sourceCurrent gatesReversal or failure behaviorProof and remaining gap
Allocate root.disk: app/root_disk_creation.go, app/disk_windows.go, app/app.goBackend refuses any existing path before pipeline operations; exclusive file creation; positive size and final regular-file/size checkExisting image preserved; newly allocated partial image retained on failure; recovery is a separate action#444 merged. Actual temporary-file existing/equal-size/symlink/racing/concurrent controls pass on Linux and Windows. Native initializer is mocked; actual VDL/full install is not proved by these tests.
Shrink C:, create/format data volume: app/disk_windows.go:CreateDataPartitionDurable .pending authority is reserved before mutation; source/disk identity is captured and rechecked after formatting; PowerShell stops on error; BitLocker suspension is limited to one rebootA failed creation retains the pending marker and does not authorize deletion; a created partition stays preserved if receipt persistence failsapp/disk_windows.go, app/partition-policy.ps1, and app/partition-policy.tests.ps1 cover the source and receipt contracts. Open #234/#197: no measured restoration of C: after a later shrink/format/ACL failure, and UseMaximumSize still needs exact extent/size proof.
Remove data partition and extend C:: app/disk_windows.go:removePartitionAndExtendCRequires the durable creation receipt and explicit removal; Get-WootcPartitionRemovalBinding re-enumerates the disk and partition and matches GUID, drive, source, and disk; deletion and extension each require a positive observationA .deleted receipt records observed deletion; extension can be retried from the receipt; mismatches and failed observations retain ownership data and refuseapp/partition-policy.ps1 and app/partition-policy.tests.ps1 cover identity swaps, failed removal, failed resize, and false success. Open #237: the actual Windows deletion/extension run is still required.
Native Manage/Recovery destructive action: app/native_maintenance.go, app/native_install_session.goStrict action shape; session, brand, and single-use intent binding; fresh engine observations; original-user revalidation; the same facts are rechecked immediately before UninstallWith or boot repairMalformed, replayed, foreign, or changed plans refuse before the action; uninstall keeps root.disk by default and reports incomplete cleanupapp/native_maintenance_test.go covers plan shape, replay, changed observations, and refusal. The test records a fake action; no real Windows mutation is proved (#237).
Native install confirmation: app/native_install_session.go, app/native_install_authorization_windows.goTyped immutable install plan; image/storage validation; captured-user revalidation; session/brand/one-shot intent binding; fresh observation immediately before starting the existing install pipelineA failed or mismatched confirmation consumes the intent; the existing install pipeline owns rollback and disarm behaviorapp/native_install_session_test.go, app/native_install_session_windows_test.go, and the authorization path cover protocol and replay refusal. The native shell has no full destructive hardware run yet (#237).
Change hibernation/Fast Startup: app/installer_windows.go, app/power_state.goPrior-state marker; install/recovery/uninstall pathsRestoration attempts recorded prior settingsParser tests in app/power_state_test.go and tests/unit/uninstall-restoration.bats do not prove actual Windows restoration on real hardware. RC verification remains #237/#238.
Create/arm firmware BCD entry: app/installer_esp.go:configureBCD, armBootChain, app/boot_txn.goExport prior BCD; hash check of the staged ESP chain before entry creation and again before the arm; journal install/boot-txn.json records each step; the one-shot is the last change before rebootEach step compares bcdedit /enum firmware with the intended chain. A failure or cancel removes the entry and reads the store again. Residue keeps the journal in rollback-unconfirmed, and the startup task tries again#286: tests in app/boot_txn_test.go cut power at each boundary and fail each step against a simulated store. Open #234: a Windows VM run with real interruptions is still necessary.
Cancel/recover/uninstall BCD cleanup: app/installer_esp.go:disarmOneShot, deleteWootcBCDEntries, app/recovery_windows.goJournal GUID or saved GUID (full GUID check), or entry description; the sweep removes only the wootc elements of the one-shotRemoves wootc entries, then reads the store. Cleanup that it cannot see keeps the journal, armed.json and the recovery tasks for a retry#286 tests cover a foreign one-shot, an entry that does not delete, and a store that does not read. Open #234: failed BCD enumeration still becomes false absence in hasWootcBCDEntry. Actual cleanup on Windows still needs counterexamples.
Initial ESP kernel/loader/config staging: app/installer_esp.go, payload/deployer/deploy.shProduct ownership manifest and selected boot chainExisting boot backup/cleanup paths; multiple files changeOpen #286/#234: immutable whole-chain rollback and process-cut firmware acceptance required. Cross-vendor transition cannot use the upgrade's mixed-trio assumption.
Uninstall ESP files: app/esp_cleanup.goRaw relative-path validation; modern manifest authority; regular-file/no-symlink checks; complete preflightRemoves exact claims; observes whole-plan absence; keeps manifest on partial failure for retry; preserves foreign neighbors#442 merged. Actual temporary filesystem tests reject traversal, ambiguous case, directory claims and lying removers. Nine Windows tests pass, one case-ambiguity test skips on NTFS and passes on Linux. No real ESP mutation in that proof.
Delete installed disk/staged data: app/installer_windows.go:uninstallWithDeleteRootDisk/RemovePartition options; default keeps existing diskRecursive cleanup removes installer subtrees; verification followsOpen #234: namespace-based recursive removal and discarded removal errors need ownership/error review; partition receipt must be validated before deleting its metadata. Default preservation is distinct from explicit root-disk deletion consent.
Discard an unfinished attempt's root.disk before a resume: app/install_recovery.go:prepareResumeAtRe-observes and re-classifies on every call. The journal starts only after the gate that refuses an existing root.disk. The attempt must have reached "Making room for Linux" and stopped before the boot steps. armed.json and deployer-started.json must be absent, and the file must not be older than the attemptVerified download pieces are kept. Any doubt refuses the resume and offers removal insteadapp/install_recovery_test.go refuses an older file, a touched boot chain, and a started deployer, and confirms that a refused resume keeps the file. Uses temporary files only. No interrupted Windows VM run proves it yet (#287).
Read host volumes and mount selected host NTFS read-write: payload/deployer/deploy.sh:scan_for_root_diskRead-only scan, root image discovery; selected volume then mounted read-writeUnmount/retry/diagnostic pathstests/unit/scan-root-disk.bats, NTFS state-writer and clean-unmount checks cover components. Open #234/#370: bind actual host identity at writes and prove dirty/hibernated/failed-observation behavior; source scan alone is not data-loss clearance.
Advance VDL, attach loop and provision image: payload/deployer/deploy.shSelected root image; loop layout and deployment verificationWrites inside raw image; failure diagnostics and Windows returnExisting successful native-cycle runs prove their exact cells. Open #234/#285/#286: interruption/retry must preserve an existing installed image; a boot-ready size is not installation consent.
Format helper target/scratch: payload/builder/wootc-builder.sh:blank_disk, prepare_storageBoth block devices, dedicated serials, whole-disk type, minimum sizes, no partitions/signatures/mount before first formatDisposable dedicated disks; installation failure reportedHelper storage preflight evidence exists; it does not prove target desktop or restart. #178 remains open. Identity changes between preflight and mutation still need final audit.
Installed ESP kernel/initramfs/config refresh: payload/migration/wootc-esp-syncResolves the host ESP from the recorded UUID; requires the owned EFI/wootc namespace; the direct route writes .new then renames; the controller route observes boot identity and ownership before each writeDirect kernel/config replacement is per-file, not a whole-set transaction; the controller retains journal state for recovery and refuses an unprepared boottests/unit/esp-chain-refresh.bats, tests/unit/test_wootc_esp_control.py, and tests/unit/test_wootc_esp_transaction.py cover the split paths. Open #237/#333: direct pair interruption and real firmware acceptance remain unproved.
Signed shim/GRUB/MokManager refresh: payload/migration/wootc-esp-sync:sync_signed_chainwootc-esp-control:prepare binds current boot identity, ownership, source package, firmware trust, and signatures; publication accepts only receipt-listed destinations and the prepared boot IDThe transaction archives and recovers the signed trio; an ambiguous source or changed boot refuses before publicationtests/unit/test_wootc_esp_control.py, tests/unit/test_wootc_esp_transaction.py, and tests/unit/test_esp_chain_refresh.py cover the controller. Open #237/#333: production integration, process-cut recovery, and firmware-upgrade boot acceptance remain unproved.
Native graduation to blank whole disk: payload/migration/wootc-go-native:graduate_to_disk_execute--to-disk requires a block device different from the host disk; execution requires --execute plus WOOTC_GN_ALLOW_DESTRUCTIVE=1; fisherman owns target partition/format checks; the image is already localSource Windows/root image remains untouched; only the explicitly selected target changes; failure leaves the source available for retrytests/unit/go-native.bats covers the dry-run and destructive flag boundary. Historical GUI run 36265248670 proves a blank-disk cell, not the native shell. User-work persistence and native boot remain open (#237).
Same-disk shrink/reclaim plan: payload/migration/wootc-go-native:graduate_plan, cmd_migratentfsresize --info is read-only; --reclaim requires native boot and at least one converted folder; in-place execution without --to-disk refuses after printing the planCurrent main performs no same-disk shrink, partition delete, or Windows-entry removal; the plan is not rollback evidencetests/unit/go-native.bats proves the dry-run and refusal boundary. The printed ntfsresize/sfdisk/efibootmgr commands are not implemented acceptance; do not claim same-disk safety.
Write/move imported user files: payload/migration/wootc-mount-user-dirs, wootc-convert-dir, plugin importsProfile map, category operation and bridge pathsCategory-specific rollback and ledger; contracts differ per importerMigration component suite proves seeded cases only. #427 actual BitLocker editor save/reopen/second-boot proof is running on draft434; no result yet. Audit all importer overwrite/collision paths before #234 closure.
Import from an external Windows or WSL volume: payload/migration/wootc-import, wootc-wsl-bridge, wootc-detect-apps, wootc-steam-bridgeRunning-disk exclusion; BitLocker unlock via stdin; read-only source mount; explicit category selection; --update/no-clobber checks for standard filesEject unmounts the bind and source and closes the BitLocker mapper; partial destination files remain for inspection and retrytests/unit/import.bats and tests/unit/test_plugin_import_results.py cover the orchestration and result shape. Open #234: bridge failures are tolerated, Steam rewrites its VDF, Firefox/Thunderbird write profile metadata, and no common per-file receipt or rollback proves every existing native file survives an interrupted import.
Shred staged vault, Wi-Fi exports and recovery key: payload/deployer/deploy.sh, payload/migration/wootc-wifi-bridge, wootc-convert-dirProduct staging paths and import/convert flow. The deployer moves the recovery key to /etc/wootc/bitlocker-key and shreds the NTFS copy only after it verifies the move. The bridge seals it to bitlocker-key.cred with systemd-creds --with-key=auto and shreds the plaintext only after the sealed copy decrypts to the same bytes. wootc-convert-dir shreds both when no folder is bridged from C:Credentials deliberately removed; rm fallback does not promise physical erasure. Shutdown keeps the key, because the bridge unlocks C: on each boottests/unit/bitlocker-key-lifecycle.bats covers the move, the fallback, shutdown and conversion. The BitLocker E2E cell reads Documents on a second Phase-2 boot. The sealed path is not yet observed in a VM. #281 staged session envelope lifecycle remains open. Never retain secret values as audit evidence.

| Add/Remove Programs registry and prior-power mirror: app/installer_windows.go:registerUninstallEntry, unregisterUninstallEntry, app/sysprobe_windows.go | Fixed HKLM product key; escaped branding values | Registry deletion attempted after power restoration | Open #234/#238: writes/removal discard errors, no ownership receipt for a pre-existing key, and key removal can destroy the last power-state mirror after unsuccessful restoration. Need actual failure/foreign-key controls and observed retained retry data. | | Boot repair and Windows-only restore (#290): app/boot_repair.go, app/recovery_windows.go:repairBootFromArmed | Reads bcdedit /enum firmware first. Acts only on entries with the wootc description. Refuses when the recorded GUID has a different description, when armed.json and bcd-guid.txt disagree, when the next boot goes to a foreign entry, or when the ESP manifest is unreadable | Restore clears the one-shot and removes wootc entries from the permanent order. It keeps the entry, the ESP files and armed.json. Repair passes only when a new observation shows the recorded entry as the next boot and the recorded ESP hashes. If that check fails, repair takes wootc out of the boot order again | Decision-table and command-list tests in app/boot_repair_test.go use bcdedit dumps, not a real BCD store. A Windows VM run of each action is still necessary. | | Recovery scheduled tasks: app/recovery_windows.go:registerRecoveryTasks, unregisterRecoveryTasks | Fixed task names; SYSTEM startup and elevated logon principals; the existing task's executable, arguments, principal, run level, trigger and action/trigger counts must match before reuse or removal | Registration refuses a foreign same-name task before the first BCD write. Cleanup is idempotent for absent tasks, verifies removal, and returns command or observation errors | The task identity is bound to the staged executable and its recorded hash is checked before startup recovery. The broader privileged-state and TOCTOU audit remains open under #234/#370. | | Trusted state directory ACL: app/state_trust_windows.go:prepareTrustedStateTree | Trusted parent volume; restricted creation descriptor; recursive owner/DACL/no-reparse inspection before root ACL update; rejects unsafe existing trees instead of repairing them | Fails before trusting planted artifacts; does not restore earlier ACLs | Component native trust tests exist. #370 remains open for the full reader/writer/transition surface and TOCTOU review; this source gate does not prove every elevated artifact consumer. Dedicated-volume root ACL and OEM equivalent need separate ownership proof. | | BitLocker recovery-password capture/write: app/bitlocker_windows.go | Selected volume; existing recovery protector; state directory and file ACL attempt | Plaintext recovery password written to install staging; ACL failure emits warning and returns success | Open #279/#370/#427: current main strips separators, incompatible with Linux password parsing. Canonical-format repair is in frozen draft431/434 awaiting runtime acceptance; never describe it as shipped. Machine-bound key storage and failure-closed secret lifecycle remain required. | | OEM carve/decrypt/key/BCD equivalents: tests/e2e/setup-wootc.ps1 | Measured C: shrink; freshly formatted storage; conversion wait; positive BCD path/entry/sequence checks | Per-step cleanup attempts and one-shot rollback | Open #234/#286: UseMaximumSize still needs exact extent proof; two absent BitLocker objects count as plaintext; BCD rollback drops errors. Harness success cannot clear Go product variants. Frozen draft434 adds a separate complete-C:-encryption readiness gate; its run is not terminal. | | Native folder conversion swap/marker: payload/migration/wootc-convert-dir | Allowed folder, nonempty source, second successful rsync; native destination must be empty; actual empty-directory removal after unmount; final move forbids nesting | Failed swap preserves native files and staging; no new marker; retry preserves Linux edits | #447 merged after all applicable hosted checks. Five actual-script/private-file controls prove successful final placement, repeat-edit preservation, hidden native files, move-time race refusal and the faulty-move mutant. Mount/identity/ownership calls are mocked. Open #234/#285: full staging ownership, locking and crash recovery remain unproved. | | Browser profile import: payload/migration/wootc-import-browser | Selection flags and source profile; Chromium files copied only if missing | Firefox uses fixed windows-import.wootc destination and updates profile list; Windows originals remain | Open #234: repeat Firefox rsync/cp can overwrite edits in a prior import, and raw profiles.ini relative path needs traversal/no-reparse validation. The Chromium existence check/copy is not exclusive creation. Seeded happy-path import does not prove collision safety. | | Office dictionary/templates/fonts import: payload/migration/wootc-office-bridge | Existing source; dictionary words deduplicated; templates/fonts use no-clobber copy | Appends dictionary; recursive ownership changes and outcome record | Open #234: no-clobber command success is not proof a file arrived; symlink destinations, partial write recovery and actual ownership before recursive chown still need refusal controls. |

Dependency review: selected Windows data before partitioning​

On 2026-10-18, review of current main 5002438059bd747146f61e3f3763461fbc1968ca and its pinned fisherman e2b316600fc68ba397730b39d1eb205dc1f1f240 found another open path. This review does not clear the other inventory rows.

Operation and sourceCurrent gatesReversal or failure behaviorProof and remaining gap
Copy explicitly selected Windows data before partitioning: fisherman fisherman/cmd/fisherman/main.go, fisherman/internal/slurp/data.goNon-manual installation with a supplied slurp configuration; source mounted read-only; byte budget reserves 2 GiB on /runExtraction errors become a non-fatal “Data migration skipped” message. Budget exhaustion can return partial or empty results without an error. The caller then continues toward partitioning.Open #234/#237: source review found no complete-copy gate before the destructive boundary. The stream also skips unreadable files, copy failures, hidden/system names, and files over 500 MB. A reproduced test on the exact pinned source returned a zero budget and Found=false; it did not execute partitioning. Required proof: incomplete requested migration stops the actual caller before partition or format, with owned source data unchanged.

The source test's scratch directory differs from the hard-coded /run capacity observation. A fixed budget can make the test repeatable. It does not prove that all selected files arrived.

A repair must check each selected file. It must report each explicit exclusion. It must stop destructive work if the copy is incomplete. Tests must observe that the actual caller stopped. A file count or a warning message cannot prove this.

A complete copy in RAM can be lost after shutdown. Before partition or format, the caller must verify a durable copy on storage outside the target. That copy and its receipt must survive interrupted writes and failed imports.

Field-report corpus​

Review date: 2026-10-18. GitHub's all-state search and independent cross-checks returned no submitted field report:

SearchResultInterpretation
label:field-report0 issuesNo issue carries the template label.
in:title "[hardware]"10 issues/PRsAll are planning issues or PRs; none is a filed run report.
in:body "Build or tag tested" or "Overall result"0 issuesNo submitted body contains the form's required result fields.
in:body "Uninstall proof"5 issues/PRsAll five are documentation or planning changes, not reports.
ReportCause and fixData impactDisposition
None foundNot applicableNo open filed report describes lost or corrupted user data. An empty corpus does not prove zero loss.The corpus is empty; no report can be triaged to green, fixed, or filed.

Issue #216's current closure evidence is #525, which adds the three-machine protocol but says the real runs still need people with the machines. Issue #210 has no individual report IDs in its body or comments. The open #302 cohort and

237 RC verification remain the evidence path. This audit does not turn the

absence of reports into a no-loss claim.

The report form itself remains the authoritative intake path: .github/ISSUE_TEMPLATE/manual-test-report.yml.

Remaining verification​

This revision completes the current-main source walk and the issue-corpus review. It does not claim release-candidate clearance. Rows marked Open remain actionable findings: their gates, reversal behavior, or evidence have not crossed the boundary stated in the row.

The inventory covers every current-head mutation site in app/ and payload/. The native shell's install and maintenance consent boundaries are explicit above. The ESP direct and controlled refresh paths stay separate. The selected-data path for pinned fisherman remains a separate row because an in-memory partial copy must not come before the partition step.

The next required evidence includes an RC-dated run for every named unit-only gate. It also includes the real Windows partition and boot interruption counterexamples, importer collision controls, and durable selected-data migration. Issues #197, #237, #279, #285, #286, #333, #370, #427, and the related hardware/cohort gates track these items. A later RC review must diff its source SHA against this inventory before it marks the rows green.