[cb1kenobi]

Moves Harper off rocksdb-js's deprecated registryStatus() alias and calls getRegistryStatus() directly. No runtime behavior changes.

⊙ Problem

rocksdb-js 2.11.0 renamed registryStatus() to getRegistryStatus() (HarperFast/rocksdb-js#887 — Add getRegistryStatus() and deprecate the registryStatus() alias). The old name is now a deprecated alias that rocksdb-js v3.0.0 removes. Harper pins 2.12.0 but still calls the old name everywhere. Closes #3359.

❓ Your call: is a standalone rename worth a PR now?

Yes: it is mechanical, and doing it now means the rocksdb-js v3 bump will not break on a removed export. Folding it into that bump instead would mix a rename into a riskier change.

💡 Solution

• Every import and call now uses getRegistryStatus(). It is the same function, so behavior does not change.
• Long-lived transaction reporting calls it directly. The typeof guard and its early returns covered rocksdb-js versions the exact 2.12.0 pin rules out.
• Tests inject a fake registry through a module-local reader, which works under the compiled build and typestrip alike.

⚖️ Alternatives

• Also migrate harper-pro: nothing to do. Its only uses live in its core submodule and follow the next pointer bump.
• Patch the rocksdb-js export from testUtils: tried in an earlier commit. Under typestrip the module binds the separate ESM build, so the fake never applies.
• Node registerHooks fake in a child process: works under both loaders, but moves about 40 tests into a child.

❓ Your call: overrule the planning review's forwarding helpers?

It suggested keeping fakeRegistryStatus/restoreRegistryStatus and forwarding them to the setter. Overruled on size: against main, that is 124 changed test lines plus a helper, versus 26 rename-only lines here.

❓ Your call: also drop the transactionDetails degraded branch?

Kept, with its warning reworded. The 2.12.0 pin makes it unreachable, and removing it also deletes its test. A small follow-up if you want it gone.

🔧 Changes

• resources/longLivedTransactions.ts: import; module-local readRegistryStatus replaces the typeof-guarded registryStatusFn, used by the sweep and holder candidates; guards removed from startup and holder reports; setRegistryStatusForTests defaults to the real reader; degraded-mode warning reworded.
• resources/databases.ts: import, incomplete-drop resume check, detached-root handle check in dropDatabase.
• resources/branchDatabase.ts: import, branchReferenceCount.
• dataLayer/rocksdbBackup.ts: import, verifyDatabaseClosed poll.
• dataLayer/DESIGN.md: three prose references in the restore notes.
• resources/DESIGN.md: brings it back under its 1000-line budget, which main exceeds and the required runLinter check enforces. It carries Prune resources/DESIGN.md back under its 1000-line budget (fixes main's Lint Source Code) #3430's two prune commits (1, 2), which alone leave it at 1004 lines after main's newer notes. It also condenses two audit-clock notes (1, 2), dropping only a stage-history clause.

✅ Verification

End-to-end route: the existing integration suite, which drives drop, branch and backup/restore through these call sites.

• Targeted integration: restore-backup-online, restore-backup-preflight and drop-table-restart-recreate pass 11/11. That covers the online-restore closure poll and the drop paths end to end.
• test:integration:all: green in CI on this head, 24 jobs (6 shards × Node v24, Windows, Bun, uWS) (run). Locally it stalls at sql-engine-differential.test.mjs, which needs two loopback addresses; this machine has one.
• LMDB engine leg for the five affected resource test files: 59 passing, 71 RocksDB-only tests skipped.
• longLivedTransactions.test.js + commitConflictDeadline.test.js: 60 passing. With dist patched to call rocksdb-js directly instead of the reader, 24 fail, so the seam is load-bearing. Typestrip cannot load the module from source today (4 pre-existing breaks, typestrip dev runtime is broken on main (unattributed JSON import) and no CI job runs test:unit:typestrip #2623), so the dual-loader property rests on the seam being module-local.
• No whole-word registryStatus remains (git grep -w). The test edits are the same rename, so each still asserts against the real registry: rocksdbBackup.test.js, perFileDatabaseTeardown.js, checksDatabases.js, closeLoadedDatabases.test.js, commitConflictDeadline.test.js, databaseAliasIdentity.test.js, databases.test.js, evictionBusy.test.js, longLivedTransactions.test.js, txn-tracking.test.js, testUtils.js. evictionBusy.test.js landed on main during this work. After the rebase, it and longLivedTransactions.test.js pass 57/57.

Full unit gates and lint (run before the seam change)

• test:unit:main (excluding applicationSpawn.test.js, which hangs locally): 7351 passing, 16 failing. All 16 fail the same way on base (module resolution, preload, process-group and bundle-pack suites).
• test:unit:resources: 4494 passing, 1 failing (transaction.test.js audit refs). That file passes 49/49 on its own, on both branch and base, so the failure is order-dependent.
• Build, typecheck and prettier pass. oxlint's only warning is in authCredentialDeferral.test.js, which this PR does not touch.

_{🤖 Generated by Claude Code (Claude Opus 5.5); posted via @cb1kenobi.}

Release target: main — dependency-API rename, no backport needed

Related PRs: #3323 overlaps (merged; added the evictionBusy.test.js call site migrated here), #3430 overlaps (its resources/DESIGN.md prune commits are carried here; identical hunks)
Complexity: easy
Framing-Verdict: better-alternative-exists (3815b60cf479)

_{Review-Coverage: authored=claude; ran=codex,gemini,cursor-grok,cursor-muse; adjudicated=domain; declined=cursor-composer,cursor-kimi; rounds=3; full=2 @ 5bf547d}

_{Review-Attention: study ~15m (critical: rocksdbBackup.ts, branchDatabase.ts +2; decisions: retain-transactiondetails-fallback, setter-default-restores) @ 5bf547d}

[gemini-code-assist]

Code Review

This pull request updates references from registryStatus to getRegistryStatus across the codebase, including documentation, database resources, transaction managers, and unit tests, to align with the updated @harperfast/rocksdb-js library. Feedback suggests simplifying the initialization of registryStatusFn in resources/longLivedTransactions.ts by removing a redundant runtime check, as getRegistryStatus is unconditionally imported.

[kriszyp]

_{🤖 Reviewed with Codex}