[skygunner]

A driver receipt can stay below the 64 KiB size cap while containing deeply nested extra values. _load_driver_receipt currently relies on decoder exceptions, so a valid receipt at depth 129 is accepted rather than rejected as bounded JSON.

Enforce an inclusive depth limit of 128 with an iterative walk after decoding. Preserve the existing file identity, regular-file, UTF-8, size and status checks. Add array/object boundary regressions. Keep excessive integers rejected while allowing the existing test to recognize either decoder error form used across Python versions; deep-input failures still require the bounded-JSON error.

Validation on Python 3.14.8, based on upstream 7cf6442f:

• The new depth-129 subtests fail on unchanged upstream and pass with the fix.
• python3 -m unittest scripts.tests.test_agent_loop_benchmark: 51 passed on exact head 37abdd51.
• git diff --check: passed.

This PR is independent of the host-timing parser change and contains only the benchmark parser and its tests.

[yyjeqhc]

Reviewed the scoped implementation and regression coverage against its source baseline; CI and diff checks pass. No blocking issue identified.