Discovery implementation verification
This report covers the measurement branch above foundation commit 49eec145a. It records local implementation evidence, not production readiness or measured recommendation uplift. No production migration, flag, purchase or deployment ran.
Passing checks
| Check | Result |
|---|---|
Hosted pnpm typecheck | Passed, eight tasks |
Hosted pnpm build | Passed, five tasks; existing bundle-size warnings |
Recommendation and affected server selection in discovery-check.yml | 355 passed, zero failed, eight service-dependent skips |
| Recommendation and affected client selection in that workflow | 55 passed, zero failed |
python -B -m unittest trainer.test_discovery_data -q | 86 passed |
| Client edition boundary | Passed |
| Generated OSS discovery boundary | Passed |
| Real testing Redis adapter, complete snapshot fixture | Eight checks passed; 41 exact temporary keys removed |
| Deterministic session-revert fixture correction | Nine passed |
The new credential-free GitHub workflow covers the recommendation selections, hosted typecheck/build, temporal dataset tests and exported discovery boundary. It installs a disposable local Redis runtime for Lua tests. The existing main-PR OSS workflow remains responsible for full exported-tree verification.
Full hosted suites are not entirely green
The full server run reports 2,306 passed, two failed, nine skipped. The full app run reports 1,180 passed, eight failed. Independent targeted baseline probes reproduced all ten named failures at foundation 49eec145a. Thirty relevant files have identical Git blobs before the foundation (edcbd8f39), at the foundation, and in the measured working tree. The navigation audits also have identical predicate deficits across all three states. These ten failures predate this recommendation work; they have not been suppressed or represented as passing.
| Existing failure | Observed baseline cause |
|---|---|
billing-coverage.test.ts: endpoint registry | generation-enhance has no declared billing policy. |
usage-log.test.ts: direct inserts | Static scanner flags the model-popularity.test.ts fixture insertion. |
top-bar.test.ts: reading navigation | Expected hidden-header state is absent on its hub scroll replay. |
back-navigation-audit.test.ts: stateful controls | Test expects admin-users:search; source uses admin-users:search-v2. |
| Same audit: nested scrollers | Admin layout lacks admin-main restoration identifier. |
mobile-nav-drawer-layout.test.ts: rows | Source and asserted minimum heights differ. |
| Same test: chrome | Source and asserted icon heights differ. |
| Same test: geometry | Source and asserted item gaps differ. |
plan-theme.test.ts: credit pill | Expected literal formatting expression is absent. |
reading-page-canvas.test.ts: gradient | Single-line source regex does not accept the existing multiline gradient. |
Only the failing predicates were rerun against isolated baseline source; there was no duplicate full-suite baseline run. File identities and original tests establish the baseline comparison. They do not establish whether every static UI assertion corresponds to an actual visual defect. Detailed probe commands, identities and output are retained locally under .local-artifacts/discovery-foundation-49eec145a/.
Exported-tree verification
The final export installed, typechecked and built successfully. Its first test run exposed an existing nondeterministic session-revert fixture: sequential inserts can share a millisecond, so UUID tie-breaking can prune a different snapshot than the test expects. The fixture now supplies explicit turn times; production snapshot logic is unchanged and all nine affected tests pass.
A subsequent test launch hit a Windows EPERM race because shared-package build and test tasks both clean/rebuild the same output directory. The exported test phase is therefore run serially with pnpm exec turbo test --force --concurrency=1. This is recorded separately from a successful one-command oss:verify; no failed invocation is presented as successful.
The final serial run passed 3,078 tests, zero failures, two service-dependent skips: 47 shared, 985 engine, 1,296 server and 750 app tests. All seven Turbo tasks succeeded with forced execution. The generated discovery boundary test also passed. These results validate the final exported source including the guest handoff and the deterministic revert fixture. The exported engine was also run directly (pnpm --filter @yumina/engine test) to retain its complete standalone output: all 985 tests passed.
Independent reviews and live gates
Independent reviews cover server receipts/outcomes/history/erasure, warehouse contracts, client flows, Redis admission/recovery, immutable ranking/delivery capture and the Python dataset/evaluator. Reproduced implementation findings were fixed and rechecked. See the client, measurement re-review and offline review reports in this directory.
The rollout plan retains actual ClickHouse DDL/FINAL/TTL/query-plan validation, source reconciliation and completeness, retained-volume ingestion freshness, isolated application HTTP/SQL load, controlled activation and matured online outcomes as live gates. Local mocks, PGlite, adapter checks and offline diagnostics cannot replace those measurements.