grow the fixture corpus to cover every rule, negatives included

[lex00]

Part of #4.

What

Grow the fixture corpus from the single rule in #10 to full coverage of the specification, one rule at a time.

Ordering

Follow the spec issues rather than the implementation: a rule gets fixtures when its normative text lands, so the gate in #8 stays green continuously rather than going red for a stretch and being silenced.

Include negative fixtures deliberately

A fixture asserting that something is rejected, and that the rejection names the right node, is worth more than one asserting acceptance. Acceptance can be right by accident. Rejection with a located, correctly-attributed message usually cannot.

Done when

Every identifier in spec/ has at least one accepting and one rejecting fixture where both are meaningful, and #8 passes without exemptions.

[lex00]

Progress: 25 fixtures, 32 of 122 rules covered, UNCOVERED.md shrunk accordingly. chant (0.63.0, public API) and the reference agree on all 25.

Worth noting from this batch:

• The two F-Exc-Lazy fixtures (false && getId(), true ? "yes" : getId()) expect shape reject, fold fold — the one permitted inversion of F-Direction — and both implementations do exactly that.
• S-Member/nullish-plain-refuses and S-Member/optional-chain pin chant-v0.63.0's #2328 behaviour from the outside.
• The fixture format gained a "$undefined" sentinel because JSON cannot write the folded value undefined. That was a real gap in conformance fixture format and runner, with an adapter interface a foreign implementation can satisfy #7's format that the first ?. fixture exposed.

Negative fixtures with rejectAt are still thin (one). Positions are worth adding only when computed, not hand-counted — see #10.

[lex00]

Correction: 31 of 122 rules covered, not 32 — the generated UNCOVERED.md is authoritative and says 31. Same slip as before: wrote the number before the script printed it.