test: fault-fixture coverage gaps for defensive mechanisms

[CMGS]

A whole-repo mechanism audit (every recovery/rollback/idempotency guard traced to its origin commit and test) confirmed all defensive mechanisms are justified, but eight lack a test that exercises their exact failure branch. Each needs a fault fixture; none is urgent — adjacent tests cover the surrounding flow.

sandboxd

• pool/claim.go — Release in-memory rollback on failed persist: no direct test for the persist-fail arm (hibernate/archive equivalents have one via breakStore).
• pool/egress.go — disarmIfReleased: the wake-races-Release window has no test naming this path.
• pool/egress.go — armEgressProxy stale-socket removal before bind: no test with a leftover UDS from a prior life.
• egress/intercept.go — leafFor near-expiry renewal and the 1024-entry eviction: neither is directly asserted.

silkd

• session.rs — Table::create duplicate-id rejection under real concurrency (two racing creates, loser's shell must not leak).
• exec.rs — Started-frame write-failure rollback (needs a broken-pipe writer fixture; the ghost-entry outcome is otherwise untestable).
• proc.rs — Table::remove_if pointer-identity guard: pid-reuse eviction has no isolated test (requires a forged pid collision).
• vsock.rs / net_egress.rs — accept-error retry loop: no EMFILE-style transient-error fixture; a failure here would take down the whole daemon.

Suggested fixtures: breakStore-style fault injection for the pool items, a half-closed pipe writer for silkd exec, fd-budget exhaustion via setrlimit in a child for the accept loops.