fix(hicache): handle finite prefetch I/O failures safely

[mks-hash]

Motivation

Current main (7a719e9a65e7f52b012ab37211b5f174e5d3d38d) now catches storage query/read batch exceptions in unified-memory mode. Two lifecycle gaps remain: a duplicate terminal ACK can reclaim the same tail twice, and an exception after acknowledged batch progress can be published as a partial successful restore instead of an explicit terminal failure.

With the unified-memory flag enabled, the ten regression cases reproduce 2 failures / 8 passes on this base: allocator Double-free detected on duplicate completion, and partial publication after the second batch raises. The latter is a terminal-outcome/cleanup contract difference, not a claim that all shared published KV is leaked. Earlier worker-death reproducers on the old base are retained as regression coverage, not presented as current unified-mode behavior.

Modifications

• Preserve upstream's terminal-ACK finally and query/read fallback for unsupported consumers.
• For supported single-worker resident KV-only operations, route finite exceptions to explicit terminal failure and keep the worker alive.
• Consume terminal ACKs once per operation, preserving cancellation and replacement identity.
• Use upstream's new _retire_ongoing_prefetch cleanup helper; do not reintroduce the old inline reclamation.
• Keep ten real-cache/file-backend regressions. Their config now explicitly enables unified memory so upstream's new exception paths are exercised.

Scope remains resident FULL, synchronous finite I/O and no rank collectives/PP tickets. No distributed exception recovery, permanently blocked-I/O reclamation, scheduler redesign, retries or proactive-prefetch feature is added. Production diff is 74 additions / 3 deletions in three files; the fourth file contains regression tests.

Accuracy Tests

Refreshed against the base above:

• Current-main baseline with unified memory enabled: 2 failed / 8 passed; patched lifecycle suite: 10 passed, zero skips.
• Relevant existing CPU selection: 128 passed / 39 CUDA-fixture skips, 43 subtests. Skips are not passes.
• Additional physical-transfer/dispatch/host-assembler selection: 64 passed / 2 CUDA-only skips, 20 subtests.
• Separate existing two-rank Gloo lifecycle: 1 passed. This checks the existing lifecycle for regressions; distributed exception recovery remains out of scope.
• Applicable changed-file pre-commit hooks and git diff --check passed. Rust hooks had no matching files; no Rust validation is claimed.

CPU fixtures use real pools/controller/file I/O, unpinned host tensors and fixture-only host admission-budget accommodations on this constrained machine. The page-envelope accommodation was external to this branch, applies only to tiny allocations, and does not alter production runtime.

No new GPU H2D/model-inference run was performed against this refreshed SHA. Previous GPU evidence remains tied to its recorded source versions.

Speed Tests and Profiling

N/A: lifecycle correctness contribution; no performance claim.

Checklist

• Applicable changed-file formatting/lint/registry hooks passed.
• Real-cache regression coverage and relevant existing tests run.
• Scope and CPU/GPU validation limits stated explicitly.

---

CI States

Latest PR Test (Base): ❌ Run #37224418283
Latest PR Test (Extra): ❌ Run #37224417998
Latest PR Test (AMD ROCm 10): ❌ Run #37224418252