feat(M5d): ~> function-composition ($f ~> $g) + partial composition + T2006

[flearc]

Summary

M5d adds the compose form of the ~> operator — $f ~> $g (when both sides are functions) → a new function λ($x){ $g($f($x)) } — completing the ~> operator (the apply form value ~> $func already shipped). This is the last M5 sub-milestone, closing out the function-signature / $eval / composition arc.

• Compose branch — in the ~> else path: evaluate the RHS; if it's not a function → T2006; if the LHS is also a function → compose by evaluating a memoized chainAST meta-lambda (function($f,$g){function($x){$g($f($x))}}) and applying it to {lhs, rhs}. The composed result is a real lambda, so it re-composes / curries under further ~> for free.
• Partial composition — a partial-application RHS ($g(?, x)) is detected by scanning for a ? placeholder and routed through the existing partial machinery (it evaluates to a function value, then composes or single-applies). Delivers $substringAfter(?,"@") ~> $substringBefore(?,".").
• T2006 — 42 ~> "hello" now raises the jsonata error code (was T1006), via a shared M.is_function helper that also backs the partial-application check.

Results

• Official jsonata suite: 1147 → 1153 / 1682 (+6, 68.6% → 69.0%), zero regressions. The function-applications group goes to 20/22 — the 6 compose cases (005, 009, 012, 015, 016, 017) flip green.
• 428/428 unit tests green.

Verification

An adversarial fidelity review against a jsonata-js v2.2.1 oracle confirmed compose order (g∘f), re-composition/currying, the T2006 boundary, partial-composition, and input-threading all match. The remaining function-applications failures (case018 parser apply-chain form, case021 regex literals) are out of scope (M7).

Test Plan

• busted full unit suite — 428/428
• scripts/run-suite.sh — 1153/1682, zero regressions
• Adversarial fidelity review vs jsonata-js v2.2.1 oracle

🤖 Generated with Claude Code