[mengmengmengqiang]

Summary

Slices 5–6 of #547, stacked on fix/547-slice-34. An admitted slot's request now has a deadline, the operator gets a one-shot backend.daemon.reclaim verb, and a refused open is typed from the channel error instead of from the refusal message text.

19 files, +731 −175, two commits on top of the slice 3–4 branch.

What this does

Slice 5 — in-flight expiry and reclaim — d1fc7e228

• A decoded request carries an explicit deadline (Phase::InFlight { started, deadline }, driven by the new request_deadline_ms limit). Past it the daemon frees the slot, closes the worker's socket and records RetainedFailure { cause: Expired, reply_sent: false }, so the caller observes an unknown outcome instead of a silent hang.
• backend.daemon.reclaim { slot, generation } is the authenticated, generation-checked, one-shot operator verb: it closes the named worker's socket, records Reclaimed, answers Invalid on a generation mismatch, never restarts the daemon and never touches Workspaces. It joins the recovery allow-list.
• Reclaimed and expired workers are detached (never joined), and slot fences are keyed by generation, so a stale worker's late clear_fence cannot clobber a newer worker's fence.

Slice 6 — typed refusal outcome — 7a24b34c2

• A failed open is no longer classified by matching the refusal message text. It is typed from the channel error: a notice is Capacity/Unavailable with resource, pool_vacant, reserved_vacant, retryable and a retry_after hint; a legacy EOF answers Unknown { observed } and never guesses "no vacant slot"; a reply that never arrives (reclaimed or expired worker) answers Unavailable { outcome: unknown }.
• AdmissionReason is #[non_exhaustive] with constructor functions; the capacity model's parse/to_sdk/exhausted are #[must_use].
• The doc comment records the version-skew matrix: new daemon × old binding, old daemon × new binding, new control × old worker bundle.

Test evidence

Dell aluminium (x86_64, 88 cores), rustc 1.85.1:

cargo +1.85.1 fmt --all -- --check                                              # rc=0
cargo +1.85.1 clippy --workspace --exclude layerfs-node --all-targets -- -D warnings   # rc=0
cargo +1.85.1 test -p layerfs-daemon --lib                                      # ok, 0 tests
cargo +1.85.1 test -p layerfs-daemon --test daemon_slots --test slot_state      # ok
    daemon_slots.rs: 7 passed, 0 failed
    slot_state.rs:  20 passed, 0 failed
cargo +1.85.1 test -p layerfs-api -p layerfs-control -p layerfs-sandbox --no-fail-fast
    # 11 failures — all pre-existing, see below

The 11 failures are not from this branch. I built and ran the same -p layerfs-api -p layerfs-control -p layerfs-sandbox suite on unmodified e9b903eac: base fails 12 of those tests, and this branch's failing set is a strict subset of base's — zero new failures.

They are a host limitation of the test machine, not of the code: the kernel-mount suites need /dev/fuse owned by the test user and fail as a normal user with kernel attach Mount: Operation not permitted (os error 1) / FUSE device ownership/type. Run as root they pass, but the daemon refuses to start when the command uid equals the owner uid, so daemon_slots/slot_state need a non-root user. The suite therefore cannot be green in one configuration on that host.

Refs #547, #540 — not closing either.