ci-data-v8 — post-P26 subset re-cut (ADR169 follow-up)

Re-cut of the CI reference-data estate, executing the owner-gated ADR169 follow-up.

Why: ci-data-v7 was cut 2026-07-20. P26 (International Trade, merged 2026-07-28, commit 84e29ed) added fact_ricci_unequal_exchange_gvc, which _bootstrap_external_nodes reads unconditionally at session init. Every CI job that hydrates a session therefore crashed at tick 0 with no such table — including qa:michigan-rollover-smoke, the tick-52 year-rollover gate, which had consequently never once executed on its merits.

Provenance. The source reference DB was rebuilt from parquet sources + schema.sql inside the vendored flake's pinned dataBuild devshell (python 3.12.13, sqlite 3.53.1) and reproduced the manifest's registry product sha 5e1e60fc809708a0801f91aa91c90074b1a75c52ef27d9a3ee32a00ab02a808e exactly — 76 tables, 59,828,266 rows. The ADR098 byte-identity contract holds. (Both local checkout copies of the DB were off-pin; the registry sha is the only trustworthy source, which is precisely the discipline ADR098 exists to enforce.)

All 85 manifest-pinned dist artifacts (84 parquet + schema.sql) were verified against data-artifacts.yaml with sha256sum -c before upload — zero mismatches. The TIGER tarball is byte-identical to v3–v7 (516fe00e…, re-verified on download).

Subset: reference-subset-20260730-v8.sqlite, 14,192,298 rows, sha c139c39859dd0975910a187737c4ef3cb520a1789689a958a2aeb1343a2e57e2, cut by tools/make_reference_subset.py from the byte-verified product. fact_ricci_unequal_exchange_gvc present with its full 51 rows (TablePolicy("full"), no generator change needed). The subset tool carries no sqlite pin — only the source product is under the byte-identity contract; the asset itself is verified by the sha256 above at fetch time.