Push idiomatics into an algebraic higher-order core

[godofecht]

Adds a maximal program that deliberately pushes the Haskell-esque side of current Flow without inventing syntax:

• first-class function values and higher-order application;
• escaping affine closures;
• compose returning a closure that captures two other closures;
• fold-map over a borrowed span;
• pure transformation/classification core;
• algebraic match routing instead of a flag/branch ladder;
• scoped effects only at the observation shell.

It replaces duplicated CI program arrays with portable-programs.txt, making the portable corpus a single source of truth. maximal/algebraic_core.flow is entry #8, so the exact same source must execute through C/native and independently lower to MLIR accepted by LLVM 18 mlir-opt.

The hostile corpus exposed two general MLIR parity layers rather than being simplified: Flow #730 added spans/matches/first-class and capturing closures, and Flow #756 fixed nested exhaustive-control-flow terminator placement. Both are merged. The temporary compiler pin is removed; this PR now validates directly against current Flow main before merge.