[ex1010]

The loader host fixture carries a synthetic stock PC from the stock-switch scenario into unrelated allocator cases: it stores STOCK + 30 at paused + 20, so the real halfword scanner can also see 0x101e0000 at offset 22 and retain an unrelated extent when ASLR puts the truncated host pool there. This deterministically reproduces the first free-space failures and cascade seen in FRACT’s SDK run during test_volatile_service; the CI pool address itself was not logged.

Clear the paused/register snapshots after the stock-switch scenario, before the allocator scenarios begin. Every existing assertion remains, including explicit module-pointer retention, unknown-task refusal and release after clearing the real test pointer; production loader code and native packages are unchanged.

Validation: both real host suites pass (test_volatile_service.py and test_runtime_loader.py), with ASan/UBSan where their existing harness enables them. A controlled synthetic uncached alias to a 768 KiB host pool with low32 base 0x101d0000 reproduces the same first failures on the original fixture; the same mapping passes after only the snapshot reset, with the production loader byte-identical. Local full pr:ready says Ready (163.12 s); pinned isolated Octatrack comparisons and all 64 current Elemod release exports also pass. Jev reviewed the changed fragment: 0 flags, 1 missing-context uncertainty investigated against callers/scenario order and the controlled reproduction, 0 deferred; the changelog is static-only, and this is not a complete file review.

This is a separate host-test correction under the repository’s rule for test fixes, not a change to FRACT’s implementation or hardware evidence. Related to #1380.

Guides: nothing missing; existing guidance already warns about truncated host addresses and ASLR in loader fixtures.