Two binaries, so an emitter in another language can be checked against this one

[pathscale]

worktable-parse reads a declaration on stdin and writes its canonical text to stdout. worktable-schemas walks a tree and prints every declaration it finds as JSON, with the canonical text beside each.

Both exist for the same reason: two implementations of a language drift unless something compares them, and the comparison has to be byte-exact, because they can agree on every meaning and still disagree on every character a person reads. @pathscale/worktable-dsl is the first consumer, and its whole test suite is built on these.

worktable-parse runs check, not Schema::parse

That is the point of it. parse answers "is this a declaration"; check answers "would the macro accept it", which is the question an emitter has to get right. page_size: 4096 beside persist: true parses perfectly and the macro refuses it, so a parse-only binary would report green for output that does not compile.

It earned that twice while the TypeScript corpus was being written — that rule, and that a non-unique index cannot use indexset or congee. Both were fixture mistakes a parse-only check would have accepted.

worktable-schemas is behind a new json feature

Not serde. Serialising the IR and shipping a JSON writer are different costs to ask of a dependent, and this crate is compiled for the host as part of worktable_codegen before anything else in one. Output is sorted, so a consumer diffing two runs sees only real change.

It finds 121 declarations in this repository with 0 rejected, which is what the TypeScript emitter is now checked against.

Also

Removes dsl/src/bin/wt-dsl.rs, a stray empty file that was never committed and made cargo test -p worktable_dsl fail to build with error[E0601]: `main` function not found.

Tests pass, clippy clean with --all-targets --features json.

[pathscale]

Folded into #87, and thank you: this got the central thing right where #87 had it wrong.

wt-dsl parse there was built on Schema::parse. Your reasoning for check is correct and I had missed it: page_size: 4096 beside persist: true parses perfectly and the macro refuses it, so the round trip reported success for output that does not compile. That is exactly what a second implementation would then ship. Fixed, with a test on that case.

Also taken across, both better than what was on #87:

• worktable-schemas, so the conformance corpus is the tables people actually wrote rather than a hand-written one that only covers what its author understood.
• The feature wiring. json separate from serde, and the point that a conformance binary behind a feature is one a cross-implementation test can quietly be missing, which is why wt-dsl no longer requires spans.

What #87 keeps beyond your two binaries: check, scan and diff subcommands, and tests on the CLI contract.

One PR per repo for this release, which is why this closes rather than stacks. The branch stays pushed.