test(corpus): gate on node kinds, because a grammar is not a bag of words

[rmichaelthomas]

The hole

require each (v8a §49) had no corpus case and is not ported to TypeScript — and the reserved-word gate was green the whole time, because require and each both appear in other programs. The port's divergence list was empty and read as complete coverage. It was measuring a different thing.

This is the Calendar Era lesson one level up. v29 changed behaviour while adding no word, so a word-list gate couldn't see it, and the answer was a behavioural corpus. But a corpus indexed by words has the identical hole: it cannot see a construct whose words are already spoken elsewhere.

The gate

A second gate, over the AST node kinds the parser can produce. Running it found require each and four more the word gate had been silently satisfied about:

Node kind	The construct	Why the word gate missed it
RequireEachNode	require each n in nums is above 0	both words used elsewhere
EachNode	each the numbers show	each appeared only as the pronoun in a where clause — a different thing
FieldAccessNode	require total of o1 is above 50	a different operand path from show total of o1, which produces no such node
CompositionCallNode	greet, and greet-name from "bob"	the call form was never exercised, only the definition

Eight cases added. Node coverage 39 of 43.

The four that remain

Excused with a stated reason, on the same contract the word list already uses — no entry without a reason, and a test that fails if a reason stops being true:

• AboutNode, WhenNode — validate()-level, never seen by the per-line pipeline
• FinishNode — finish outside an event handler is refused by the analyzer, and the corpus has no handler
• PackVerbNode — needs a pack the generator does not load

Three tests hold that contract: every node kind covered or excused; no excuse that has stopped being true; no excuse for a node the parser no longer defines.

Verified by deletion

Removing the require each cases fails the gate with assert not ['RequireEachNode']. A gate that has never been seen to fail isn't known to be a gate.

Consequence for the port

@liminate/ts-validator now diverges on require each, which is correct — it doesn't implement it. That is the gate reporting a real gap instead of an empty list. The port PR follows.

Verification

1902 passed, 0 failed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01E6qcDj1e1dEYvc8jLnpGZy