Add @compact primitive and free-list memory recycling

[gavlooth]

Summary

• Adds @compact(term) primitive: normalizes term to SNF, then deep-copies the result tree to fresh heap positions. Safe to call from any evaluation context because it never modifies existing heap data. Enables iterative algorithms (e.g. Bellman-Ford with early termination) to shed accumulated intermediates between rounds.
• Adds per-thread, size-segregated free lists (sizes 1-16) with heap_free calls in all interaction handlers (APP-LAM, DUP-NOD, DUP-SUP, MAT-CTR, MAT-NUM, OP2, EQL, AND, OR, etc.). Disabled in multi-threaded mode to avoid cross-thread races.
• Adds OP2-NUM fast path: skips frame push/pop when both operands are already NUM.
• Splits dup_nod allocation into 3 separate allocs for better free-list reuse.

Motivation

HVM4 programs with iterative loops (e.g. @repeat_until calling Bellman-Ford rounds) accumulate heap garbage from evaluation intermediates. Without reclamation, heap exhausts after a few rounds. The free list recycles interaction-freed blocks immediately, and @compact provides a periodic deep-copy escape hatch for long-running computations.

Usage

// In a loop body — compact state between rounds to shed dead heap
@check_go = λ{
  0: λn. λf. λdist. #S{dist, 0};
  λm. λn. λf. λdist. @repeat_until_go(n - 1, f, %compact(#S{dist, 1}))
}

Test plan

• All 196 existing HVM4 tests pass (8 FFI tests fail pre-existing on Linux due to -dynamiclib)
• Compact works with linear use, DUP of compacted terms, sibling pairs, nested constructors
• Bellman-Ford with compact + early termination produces correct shortest paths ([0,3,2,6,7])
• Repeat loops with compact (5+ rounds) produce correct results
• Free list recycling reduces heap usage across all interaction types

🤖 Generated with Claude Code