[Haoming9527]

Refs #872 and #932. Stacked on #933; this is a measured performance slice, not full acceptance.

What was wrong

Each Windows session command sealed three private channels, but pending connects woke on a timer and drain polled reader completion. On this host the unchanged driver measured 46.33 ms p50 / 51.10 ms p95 for 300 empty commands.

The change

Signal pending connect operations with a manual-reset event and wait for physical reader-thread termination using owned thread handles. Active reads still drain to authenticated EOF, spill settlement stays checked, and cancellation joins overlapped I/O before releasing buffers.

Add normal-gate real resident-worker assertions for current-generation forged stdout receipts, surviving background jobs (including disowned redirected jobs), sibling preservation, and explicit rejection of unregistered stock asynchronous output. The Windows-only suite is an explicit platform skip on POSIX; missing Windows runtime remains a failure.

Include bilingual performance notes, the supplied harness with executable-path parameters, and both paired raw reports. No dependencies, runtime pins, public contracts, output caps, or model schemas changed.

Validation

• cargo test --locked --manifest-path native/Cargo.toml -p ops-process: 43 passed, 0 failed, 5 ignored child fixtures.
• cargo clippy --locked --manifest-path native/Cargo.toml -p ops-process --all-targets -- -D warnings: passed.
• npm run build: passed.
• DSH_OPS_REQUIRE_RUNTIME=1 node test/run.mjs resident-shell-integrity: 3 checks passed, 1 suite, no skips.
• Named acceptance regression filters (verify-gate, authority, persistent-admission, persistent-mode, resident-shell-channel, resident-shell-output, resident-shell-admission, session-reset-notice, native-ranges, shell-slot, shells-native-spill): 12 suites passed, 0 failed/skipped.
• node scripts/check-file-size.mjs: 950 code files passed; longest 948 lines. git diff --check: passed.
• Full local node scripts/verify-gate.mjs --skip-install: 741 Rust tests passed (8 ignored fixtures/platform cases), 264 JavaScript suites passed, 4 allowed skips, and 92 manifest checks passed. The command failed at package extraction because Windows bsdtar rejected non-ASCII filenames. Repeating that exact package extraction with installed GNU tar 1.35 succeeded; the subsequent packed-runtime closure check passed (1 runtime, SHA-256 1bb7a20ba597c45b6d761df8bc7dbc12fc255dc866ff2f84556d012a65f0d15c). The original gate exit remains reported as failed; PR CI is the cross-platform reference.
• PR-triggered CI required: run 43608861159, exact head 24ac360551c1336453b78268090574f97a1de50a, passed. All 10 jobs succeeded, including Windows/Linux/macOS functional gates and the aggregate CI required check.

Paired empty-command results (p50 / p95): connect wake alone 46.48 / 48.77 → 16.05 / 17.22 ms; connect wake plus thread-handle wait 46.33 / 51.10 → 9.04 / 10.30 ms. Each run uses 300 samples, 30 warmups and the same stock executable. Full raw results include printf tails without filtering.

The ≤1 ms p50 / ≤3 ms p95 target is not met. The 106-operation scenario, cold startup, host orchestration, Linux comparison, channel pooling and redundant prompt/file work remain outstanding. The attached scenario summary disagrees with its raw report and needs reconciliation. These driver-only results do not establish full feature or end-to-end performance acceptance.

Environment

Windows x64, Ryzen 7 6800HS, Node 26.12.1, Rust 1.97.1 release binaries built in isolated target directories. Parent is #933 at 06c22a65, whose PR-triggered CI passed: https://github.com/T-Auto/dsh-ops/actions/runs/43554104077. This host differs from the maintainer's i9-13900H.

See performance evidence for exact hashes and measurement limits.