[gspark-etri]

Native bindings cannot load the synthetic model.bin fixtures. Add an external fixture provider and traceable JSON run reports while retaining all default golden cases and verification assertions. Reused same/elsewhere ranges are checked against actual bytes with bounded reads; B2 requires an actual changed units case.

Stacked on #80 (ADR-0010 PR #76). No vendor-format logic is added.

Validation: 26 Python regressions, default four-case B1 and B2 runs, Python 3.6 syntax, and diff checks pass. Independent review findings were reproduced and corrected. Native RNGD/ATOM runs are pending and will be attached by the runtime PRs. Refs #79.

[gspark-etri]

Validation evidence read: CI 41062510142 is green for Go/race/vet/build and both S3 backends; default four-case B1/B2 remains 0 failed. Local 26 regression tests pass, with independent review confirming the byte-equality and actual-change gates. Runtime RNGD native attempt02 has now passed all three declared cases (identical/interior projection/LM) and nine fresh-token generation gates; final fault gate/export are in progress. The original coarse E7 projection fixture correctly requested reload when its changed chunk also covered equal host embedding/scaled norm bytes; that failed expectation is preserved in runtime PR40 rather than weakening coverage. Stacked on #80; author PR76 branch is unchanged.

[gspark-etri]

Both runtime outcomes are now published and merged into rngd-pd's integration branch: RNGD #40 passes bounded-subset B2 with actual serving D2H/journal/undo; ATOM #41 passes public B1 plus n=5 reload timing. Frozen raw reports: RNGD, ATOM. Both use this neutral kit at 71c37d2 and unchanged PR76 schemas/assertions. The provider inputs declare supported native subsets, with actual same/elsewhere byte checks and an actual changed-case gate. No vendor parser was added to onggi. PR80/81 remain reviewable as an isolated stack; the ADR PR76 author branch has not been modified. NUMBERS.md is untouched.