Tracking: one agent-mode patch engine for apply and rollback

[mikolalysenko]

[agent] Filed by the scheduled architecture audit routine (CLI and core). Register: discussion #560 register.

Kind: tracking. Source: review 7.4 and 7.6 #4; register C24.

Problem

Verified on main @ 045d7ec. Agent-mode apply (patch/apply.rs, 3,665 lines) and rollback (patch/rollback.rs, 2,585 lines) implement one transformation in two directions, written twice:

Concern	apply	rollback	Drift
Verify types	VerifyStatus/VerifyResult	VerifyRollbackStatus/VerifyRollbackResult	identical fields; rollback adds MissingBlob
Per-file verify	verify_file_patch	verify_file_rollback	same guard, hash, and before/after comparison with the roles swapped; exact == in both (#707)
pnpm/vlt store-copy fan-out	apply_package_patch	rollback_package_patch	copy-pasted loop
Copy result fold	fold_copy_result	fold_copy_result	drifted: apply carries only the ownership advisory, rollback any advisory; both drop the copy's per-file records (#756)
Sidecar boundary	dispatch_fixup	dispatch_rollback_fixup	same best-effort wrapper

Symptoms

• Agent-mode apply reports already_patched / applied: 0 when it actually patched an unpatched pnpm peer-variant copy (the store-copy pass's writes are never reported) #756: apply (and, per its comment, rollback) report already_patched/already_original after writing an unpatched store copy, because each fold drops the copy's per-file records.
• Agent-mode apply and rollback reject a manifest hash in uppercase hex that blob download accepts as valid #707: hash case policy decided in each verifier separately.

Target design

One engine in patch/engine.rs, parameterized by Direction { Apply { sources, policy }, Rollback { blobs } }:

• one FileVerify { file, status, message, current_hash, expected_hash, target_hash } with one FileStatus (Ready, AlreadyDone, HashMismatch, NotFound, MissingSource);
• one PackageOutcome with files_changed (today's files_patched and files_rolled_back);
• one store-copy fan-out and fold, and one sidecar boundary.

apply_package_patch/rollback_package_patch stay as thin public wrappers, so the CLI does not change.

Checklist (ordered; each one is a PR)

• Share one pnpm/vlt store-copy fan-out between agent apply and rollback, folding each copy's per-file records #772: one store-copy fan-out and fold shared by apply and rollback, which merges per-file records (fixes the root cause of Agent-mode apply reports already_patched / applied: 0 when it actually patched an unpatched pnpm peer-variant copy (the store-copy pass's writes are never reported) #756).
• Unify the verify result types (FileVerify, FileStatus) behind type aliases; mechanical, no behavior change.
• One per-file verifier with direction-specific source checks (rollback's before-blob probe); folds in Agent-mode apply and rollback reject a manifest hash in uppercase hex that blob download accepts as valid #707's case-insensitive compare.
• One sidecar boundary (dispatch_fixup / dispatch_rollback_fixup behind one Direction match).
• Collapse apply_package_patch_at / rollback_package_patch_at into the engine and delete the old aliases.

Acceptance criteria

• Every child has landed, and patch/rollback.rs holds only the rollback-specific source logic (before-blob probe, new-file delete).
• The core apply/rollback unit tests and the CLI apply, rollback, remove and repair integration tests stay green throughout.

Dependencies

• Coordinate with Delete PatchSources::mem_blobs, the single-variant VendorSource predicates, the redirect-state group-commit capture and the group_commit switch-off oracle #746 (C23, PatchSources::mem_blobs delete) and Agent-mode apply and rollback reject a manifest hash in uppercase hex that blob download accepts as valid #707 (C41); with open PRs Fix agent mode patching linked first-party source (#626) #634 and Full Gradle support in agent, hosted and vendored modes #646, which touch patch/apply.rs; and with whichever fix lands for Agent-mode apply reports already_patched / applied: 0 when it actually patched an unpatched pnpm peer-variant copy (the store-copy pass's writes are never reported) #756.