[N0zoM1z0]

Existing maintenance regressions fail on current behavior because they require optional CDP shutdown commands, deletion of a skill directory containing user files, flattened EVM diagnostics after an additional cleanup failure, and another firmware launch while prior process ownership remains unverified. Verify actual socket release while the external browser remains available, removal of the managed skill manifest while preserving caller files, the public projected diagnostic chain with retained-root cleanup after recovery, and blocked firmware launches while the uncertain workspace remains retained.

The EVM fixture now forwards actual cleanup confirmation from the owned launcher or the shared token-aware process-group helper. Its reporting-failure injection no longer discards that confirmation when recovery is enabled. Product code, public schemas and tool inventory are unchanged. The unverified-worker assertion correction is also present in #1616; that PR's actual retained-workspace recovery and provider changes remain separate.

Final integration includes main dce81ca4; at c2e2ed9aa2469b7ff8225018fddc8c6e071ebe5a, npm run test:focused -- tests/boundary/browser/cdpCleanupCancellation.test.ts tests/boundary/browser/cdpConnection.test.ts tests/boundary/browser/cdpBrowserProvider.lifecycle_failures.test.ts tests/boundary/filesystem/qwenClientConfiguration.test.ts tests/boundary/evm/evmoleInterfaceProvider.test.ts tests/boundary/firmware/firmwareIntegration.test.ts passes all 84 runnable cases, with 2 host-only skips across six files. The focused preparation rebuilt the current runtime; the lane exited normally without an external guard stop. npm run check:fast, explicit-file formatting and git diff --check pass.

Linux x64, Node 24.18.0/npm 11.16.0; one lane, CPUs 13/15, nice 10, 2304 MiB family-RSS guard and 2 GiB host reserve. The workflows use real local HTTP/WebSocket, filesystem and subprocess boundaries with controlled producers. No real Chrome, Qwen agent, EVM Python engine or Hopper/Ghidra execution is claimed; Windows/macOS were not verified.

Refs #1065. Helps unblock #1601 without mixing its allocation implementation into this test-only change.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.

[N0zoM1z0]

Superseded by merged #1618 (16d9c5ad). I compared the actual main diff: it covers the EVM retained diagnostic/recoverable cleanup contract, Qwen managed-file removal with user-file preservation, and blocking a subsequent unverified firmware worker. It additionally repairs production CDP late-cancellation behavior so optional shutdown stops promptly and the independent detach attempt retains its budget, which is stronger than my transport-close-only assertion correction.

The extra socket counter/peer-availability checks in this branch overlap existing transport/wire coverage and do not justify a second maintenance merge. I am closing #1617 without deleting its branch/history. Its exact c2e2ed9a verification remains recorded: 84 pass, 2 platform skips, static checks pass; those results are not relabelled as tests of the newer main.

The original #1601 CI blocker can now consume main's existing repair. #1616's actual verified-worker workspace recovery remains separate. Thanks for preserving the production cleanup contract as well as the failure observations.

[morluto]

Thanks @N0zoM1z0 for tightening maintenance coverage around CDP shutdown, retained EVM failures, user-owned skill files, and unresolved firmware-worker ownership. Those boundaries are carried into main in 16d9c5ade across the CDP, EVM, filesystem, and firmware tests (main).