[loverustfs]

Change

Add a fake whole-block engine ABI harness, typed host DPU simulation messages, and a checked CPU/TCP-versus-host-simulation functional pair. Real engine and hardware execution remain explicitly unavailable.

Add an executable 1 → 3 → 10 shard scaling example with independent localhost TCP services and directories. It drains Get pins, stops/joins old owners, prepares epochs 1 → 2 → 3, publishes complete routes, rejects old local sessions and remote epochs, fences pending writes and demonstrates remapped misses followed by explicit recomputation. Fourteen per-shard roundtrips verify payload bytes and release. It is SINGLE_PROCESS_LOCALHOST, not ten physical hosts or performance evidence.

Document bootstrap, immutable limits, full route publication, remote-fence responsibility, restart/rollback and absent automatic migration. Add scaling-smoke to the CPU regression runner and preserve all hardware/engine deferrals. This branch includes #10–#12 and the reply-budget regression fix.

Validation

Run the retained CPU gate with a fresh evidence directory:

python3 scripts/ci/vm_regression.py --profile cpu --offline --output artifacts/merge-review-cpu

Fetch locked dependencies once before the offline run. The runner checks formatting, default/all-feature Clippy, workspace/doc tests, optional-feature refusal contracts and CPU demos; it writes results.json, environment.json, summary.md, JUnit and raw logs. GitHub CI runs Linux x86_64, native aarch64 and aarch64 qemu-user checks and uploads x86_64 regression evidence.

The final CPU profile contains 21 gates. The scaling example also runs under workspace all-target tests:

cargo fmt --all -- --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace --all-targets
cargo test -p gpucache-transport --all-targets --features rdma-nixl
cargo run -p gpucache-transport --example scaling_smoke

Boundaries

GCK1/GCKR and CacheIdentity remain unchanged. Missing, failed or unverified blocks require recomputation. No hot erasure coding, durability replicas, linked accelerator libraries or product performance samples are added. TCP remains a trusted-fabric plaintext transport; the bearer fingerprint does not provide mTLS.

Real RDMA/NIXL, GDS, BF3/DOCA/SPDK, vLLM export/import and live RustFS remain NOT_RUN_HARDWARE. GPU/model execution remains NOT_RUN_BUDGET; an approved model/license campaign remains NOT_RUN_LICENSE. TCO_UNKNOWN, go_gate=NOT_EVALUATED, Mooncake NOT_LINKED and NOT_A_PERFORMANCE_RESULT remain.

Exact reviewed head af5162912b87b24a722004aa61fece93b72fb7cf: all five required commands PASS, with 277 workspace tests, 93 rdma-nixl transport tests and 21/21 CPU gates. The checkout was clean before and after. Local evidence: artifacts/merge-final-required-commands/ and artifacts/merge-final-cpu/. GitHub exact-head CI passes x86_64, native aarch64 and qemu-user; PR CI also passes all three.