fix: stabilize hosted destroyer scenarios

[smiggleworth]

Summary

• restore the five family-actor scenarios omitted from the hosted matrix and forward their required CPU/failpoint controls
• start storage-fault writers before cutting their storage path and keep tmpfs-backed Sqrzl mounted during Fitz restart
• keep same-type recovery requests sequential on one client connection, as required by the wire contract
• reconcile ambiguous Stream pressure outcomes on an isolated replaceable connection and retain strict failure handling when reconciliation fails
• pace sustained pressure according to domain semantics, bound broker sampling overhead, and prevent wall-clock sampling overruns from inventing progress windows
• retry only typed Queue 4005 not-accepted responses with bounded exponential backoff while retaining exact reconciliation for unknown durable outcomes
• allow active clients a bounded three-request observation window after correlated actor failure
• update the harness to @cntryl/fitz 0.0.22

Why

The original full run at https://github.com/cntryl/fitz-destroyer/actions/runs/33666942180 exposed seven scenario failures and five matrix omissions. The failures came from harness ordering, invalid same-type request concurrency, tmpfs loss during storage restart, Stream sessions stranded after ambiguous BEGIN responses, unbounded diagnostic work during sustained pressure, a hosted active-fault observation window that was shorter than the in-flight request lifecycle, and broker liveness gaps corrected by cntryl/fitz#231, cntryl/fitz#232, and cntryl/fitz#233.

Validation

• npm run check: TypeScript compilation and 183/183 tests passed
• branch validation: https://github.com/cntryl/fitz-destroyer/actions/runs/33775542364 passed 60/60 jobs, including all 58 scenarios and aggregate analysis
• post-merge main validation: https://github.com/cntryl/fitz-destroyer/actions/runs/33777076674 passed 60/60 jobs at b8f6cc4172f5ffb6cd412bf8456fafc5dc816885
• actor-supervision evidence: all 7 isolated and 7 correlated domain faults, all 4 active clients, 8 readiness withdrawals, 8 recoveries, and every recovery canary
• Queue overload evidence: all 64 operation identities reconciled for each of 4 producers; the exact local scenario also passed 3 consecutive runs against the corrected broker
• post-merge harness image: ghcr.io/cntryl/fitz-destroyer:b8f6cc4172f5ffb6cd412bf8456fafc5dc816885, digest sha256:51e3b32c2ddf0454f489c0f5db13f8c9d0046157cc49cce47ed34b70b70799f5
• broker image used by the final run: ghcr.io/cntryl/fitz:latest and :main, digest sha256:1a864bf434378f7de36c44e0d9ceca60e091bb2c58a6b3a924417a9a42337bb7

AI assistance

Codex assisted with hosted-artifact analysis, implementation, regression tests, local Docker validation, and end-to-end CI verification. I reviewed the complete diff and retained strict correctness assertions rather than suppressing scenario failures.