nordisk/nornir

The multi-dimensional viz matrix test (tab × mode × data)

Rickard's explicit ask: "test all variants — tab × mode × data." This doc is the contract for tests/viz_md_matrix.rs — the axes it crosses, every cell, and which cells are asserted vs. logged known-gaps.

Why it exists — the gap it closes

tests/viz_live_data.rs proves each data-bearing tab renders NON-EMPTY on the embedded (fat) path — a SINGLE axis. The live audit that motivated this file found the hole: the thin / remote path is what a real user runs against the server, and several tabs rendered EMPTY there even though the embedded path was green. A single-axis test can never catch that.

This file crosses three axes so each (tab, mode, data) crossing is its own assertion. The whole point: it FAILS if a tab is empty-with-data in EITHER mode — catching the remote-path emptiness the single-axis test missed.

The axes

10 × 2 × 2 = 40 cells.

How a cell is read — the CORRECT state_json keys

Each cell reads the load-bearing rendered-row count for its tab, with the right key:

For the still-un-wired thin Test stub, "clean" additionally requires the Test.Results RPC TODO note in state_json.test.source. The now-wired thin Release / Leaderboard cells use the same clean signal as every other wired tab — a null state_json[tab].error.

Seeding (real data in — LAW 1)

• fat: a temp Iceberg warehouse seeded with release_lineage + dep_graph (the tables Viz.Timeline is built from) + release_events + test_results + agent_model_runs, driven through the embedded inject_*_for_test hooks and the warehouse read path. The Knowledge cell lays a tiny real crate per repo on disk and runs the scan.
• thin: an ephemeral nornir-server serving a monitored workspace matrixws whose <home>/.nornir/workspaces/matrixws/builds/ warehouse is pre-seeded the same way before the server opens (locks) it, so the remote Viz.* / Warehouse.* RPCs return real data — the exact path the GUI drives. A second, un-seeded server provides the data-absent column. The thin Knowledge cell lays a local checkout on the client (Knowledge is a client-side local-disk scan even in remote mode).

The seeded warehouse is always a temp dir — never committed (warehouse-never- committed LAW).

The cell table — asserted vs known-gap

37 cells asserted (20 fat + 17 thin), 3 cells known-gap (logged #[ignore]-with-reason, all thin × present).

Fat column (20 cells — all asserted)

Every tab × {present, absent}. Present asserts NON-EMPTY rendered rows; absent asserts the clean empty placeholder (0 rows, no crash). The driver sets NORNIR_VIZ_NO_DURABLE_ACTIONLOG=1 so N5's durable action-log does not hold the catalog's exclusive redb lock — otherwise the Funnel cell dead-locks opening the same warehouse via the funnel Store.

Thin column (17 asserted + 3 known-gap)

Known-gap cells (why, and what flips them to required)

Flipped to asserted (2026-06-13): Release (Viz.ReleaseEvents) and Leaderboard (Viz.BakeoffResults) remote reads are now wired — their thin × present cells assert real rows over the wire and their *_known_gap tests are un-ignored (now plain regressions on the wired RPC).

Each remaining known-gap cell has a dedicated #[ignore]-with-reason #[test] that asserts the INTENDED contract (real rows over the wire) — so flipping it on is a one-line change, and it never asserts false-green in the meantime.

The tests

• md_matrix_fat_all_tabs_present_and_absent — the 20-cell fat column.
• md_matrix_thin_wired_tabs — the 17-cell thin column (two ephemeral servers: one seeded, one empty). This is the test that catches the live remote-path emptiness.
• md_matrix_shape_and_known_gaps — guards the cell table itself (40 cells; exactly 3 known-gaps — Test/Funnel/LiveRun — all thin × present, each carrying a reason; Release + Leaderboard are NOT gaps).
• md_matrix_thin_{release,leaderboard}_present_known_gap — now-WIRED, un-ignored, assert real rows over the wire.
• md_matrix_thin_{test,liverun}_present_known_gap — the remaining #[ignore]-with-reason known-gap cells (flip on when the read lands).

Run

cargo build --features viz --bin nornir-ui
cargo test --features viz,server --test viz_md_matrix -- --nocapture

Runs under nornir test as plain #[test]s (LAW). The server feature is required for the thin column's ephemeral nornir-server.

The workspace axis — tests/viz_workspace_matrix.rs (2026-06-13)

viz_md_matrix.rs crosses tab × mode × data for ONE workspace. The sweep adds the missing workspace axis (Rickard's law: every workspace × every function must retrieve its OWN functional data, and a workspace switch must re-scope every function with no stale leak — the Security re-scope bug that shipped in 0.4.15).

One ephemeral nornir-server monitors THREE workspaces (alpha/beta/gamma), each builds/ warehouse pre-seeded with a DISTINCT count (2/3/4 release_events rows) and a DISTINCT repo set. Two tests:

• every_workspace_every_function_retrieves_its_own_data — a thin app scoped to each workspace reads its OWN counts over Viz.ReleaseEvents / Viz.BakeoffResults / Viz.Timeline / Warehouse.Scan, and the Security tab lists that workspace's repos. Asserts the EXACT distinct count, so a leak fails.
• switching_workspace_rescopes_every_function_no_leak — one app drives the picker alpha→beta→gamma→alpha→gamma→beta via switch_workspace_for_test and asserts every function follows the current workspace, never the previous.

Heavy is default-ON in the release gate

run_cargo_test now runs --include-ignored (heavy: container/corpus/matrix tests) by default. Opt out with nornir --skip-heavy … (or NORNIR_SKIP_HEAVY=1). nornir's own release runs the feature-gated viz matrix via NORNIR_TEST_FEATURES=viz,server (appended as --features before --).