test(storage): exclude TempDir teardown from durable walltime samples

[DecisionNerd]

Objective

Remove temporary-directory teardown from the measured region of the M6 durable walltime benchmarks so CodSpeed Macro comparisons measure the storage operations rather than recursive fixture cleanup.

Evidence

• Exact main 8156b746: durable_commit fastest/median/mean = 42.01/96.36/243.4 ms.
• PR feat(ontology): add provenance-bearing bridge sets #855 head eb2660c3: durable_commit fastest/median/mean = 94.4/96.73/237.7 ms.
• feat(ontology): add provenance-bearing bridge sets #855 has no graphforge-storage diff. Median and mean are unchanged while the fastest tail moved.
• durable_commit passes two TempDir values through bench_local_values; spill_compaction passes one.
• In codspeed-divan-compat-walltime 5.0.1, by-value input destruction happens inside the benchmark closure, while bench_local_refs drops input after the timed sample boundary.

Acceptance criteria

• TempDir-backed per-iteration fixtures use reference input semantics, keeping recursive teardown outside the measured closure.
• The measured operation and fresh-per-iteration fixture behavior remain unchanged.
• A deterministic repository test prevents TempDir-backed durable walltime fixtures from regressing to by-value input ownership.
• The focused benchmark target compiles and runs for the affected fixtures.
• Repository policy and the CodSpeed workflow remain fail closed; no threshold, retry, skip, or acknowledgement changes.