Reproducing and reviewing evidence

bootstrap/frontend-v81-build.json fixes reconstruction sources. Builders verify SHA-256 during copying and reject destinations containing .ncc. This distinguishes a clean generation from reuse of previous objects.

The sequence is to build B with the seed, build its runtime, build C with B and its providers, then repeat the checks on C. docs/NCC_MODULES_V81_REPRO.md records complete builder commands and arguments.

seed -> compiler B -> runtime B
compiler B + runtime B -> compiler C -> runtime C
artifact comparison and contract execution

Reports retain expected output, exit codes, objects used and exact differences. Failed attempts remain separate from final runs. A successful retry should not erase the cause of the previous failure.

Adversarial tests establish rejection of altered inputs. Recorded cases cover invalid alignments, inconsistent offsets and modified images. An image rejected by validation must not be executed as a substitute for resolving the rejection.

The documented unwind tests apply real Windows APIs to metadata and a test context. They do not inject a real exception at every prologue instruction. That limitation belongs to the technical claim.

Reproduction can require system tools, stack settings and external runners. Record these conditions before comparing a local result with the dossier.