[frankmcsherry]

Now directly on master-next. One commit; it no longer waits on #981 or #982. It adds interactive/tests/difftest.rs and a pointer in interactive/README.md, and nothing outside tests changes. #978 and #986 fixed bugs this harness found, and #981/#982 fix the one it found that is still open (below).

What it is. An #[ignore]d test that generates random, well-typed lambda programs over random update streams. Each case is checked against oracles that need no reference implementation.

• Programs: declared inputs of several shapes; map, value, filter, flatmap, join, negate, concat and difference, arrange, the reductions, delay, join_function; small and nested iterative scopes.
• Feeds: 3–30 epochs of inserts and retractions over small values, so keys collide. Some cases use ints of every storage width; some have one big epoch.

The oracles, each comparing every export after every epoch with another way of computing it:

• workers: 2–5 workers against 1.
• scratch: the accumulated input evaluated from scratch.
• optimize: the program unoptimized.
• delta: every join as a delta join.
• ahead: every row fed up front at its future time.
• split and bind: the program cut into a producer and a consumer. The consumer imports what the producer publishes, as containers or arrangements (split), or has its inputs bound to the producer's traces (bind).
• reinstall: the split consumer dropped, and installed again later. Off for now, see below.
• explain: every row regenerated by its explanation alone.
• On one worker, every snapshot also round-trips through a binary frame and feed text, taps sum to the contents, and a keyed peek matches the contents at its key.

Off until #982: late installs. One switch, LATE_INSTALLS, is false.

• With it off, the split and bind consumers are installed with their producer, and reinstall doesn't run.
• Why: a consumer installed after its producer has run imports traces whose history has been compacted. On master-next such a consumer does not converge, and its tick hangs. DDIR server: every import reads a trace as of the time its reader starts #982 (on Proxy backends that read entered and frontiered batches #981) fixes it: each import reads its trace as of the time its reader starts.
• Turning it back on is a one-line change once DDIR server: every import reads a trace as of the time its reader starts #982 lands. The late epoch is drawn either way, so every seed generates the same case with the switch on or off.
• Checked on master-next: with the switch on, 360 seeds of split, bind and reinstall hung twice. Seed 40000075 hung in reinstall, and 40000160 in split with a late consumer. Both pass with the switch off.

Running it. A seed always generates the same case. A failure is minimized and printed as a case file, with the rows each side has that the other lacks and the commands that replay it, by seed or from the file. The module docs cover running, picking oracles and replaying.

DIFFTEST_SEED=0 DIFFTEST_CASES=1000 cargo test --release -p interactive --test difftest -- --ignored --nocapture

• DIFFTEST_ORACLES=workers,delta runs only those oracles.
• DIFFTEST_SEED=<seed> DIFFTEST_CASES=1 replays one case.
• DIFFTEST_CASE_FILE=<file> checks a printed case, edited by hand if you like.
• With every oracle on, it runs about 25 cases a second.

What it skips, deliberately

• The explain oracle skips programs that read or set times, whose answers depend on arrival.
• A min loop's step filter reads the node or the label, never both. A filter that weighs a label against the node it flows to isn't label propagation. A node can hold a label it can't pass on, in place of one it could, and the loop need not have a fixed point. Sweep seed 5021723 cycled with period 3; a plain simulation of the program confirms it, and DDIR computed it correctly.

Checks:

• On master-next bdc950a, seeds 40,000,000–40,024,999 with every oracle: 25,000 cases, 0 failed, 0 refused, 0 hangs.
• Rebased onto master-next 53f3f63 (with LDBC: reduce query intermediates and profile additional memory #996 and DDIR explain: explaining a flatmap of a collected list is not quadratic #998), and the explain skip for big exploded lists removed, since DDIR explain: explaining a flatmap of a collected list is not quadratic #998 fixed the quadratic. Seeds 41,000,000–41,009,999 with every oracle: 10,000 cases, 0 failed, 0 refused.
• Clippy shows nothing in the new file.

🤖 Generated with Claude Code

https://claude.ai/code/session_019XsakwFGwtESwZ1gSYwG4o