Add DAG-based constraint propagation (HC4Revise)

[dpsanders]

Summary

• Adds an explicit DAG (directed acyclic graph) engine for forward-backward interval constraint propagation, as an alternative to the existing code-generation approach via ReversePropagation.jl
• New types DAGContractor, DAGSeparator, and ConstraintDAG in src/dag/ that are drop-in replacements for Contractor and Separator
• Implements iterative HC4Revise: forward evaluation (leaves→root) + backward contraction (root→leaves using IntervalContractors reverse functions), iterated to a fixed point

Motivation

The current approach compiles a symbolic expression into a Julia closure via eval(). This works but produces an opaque closure that can't be inspected, iterated, or shared across constraints. An explicit DAG enables:

• Inspection/debugging — print or visualize the computation graph
• Iterative narrowing — repeat forward-backward passes until convergence (important when variables appear multiple times)
• Multi-constraint propagation (future) — share variable nodes across constraints so narrowing from one constraint benefits others
• No eval() — avoids world-age issues

Current status

Both approaches produce identical paving results on all test cases. The DAG is ~10x slower currently due to DAG reconstruction per call and dynamic dispatch. Optimization paths are documented in CLAUDE.md.

Problem	ϵ	Codegen	DAG	Ratio
Unit disk 2D	0.1	287 μs	3.2 ms	11x
Unit disk 2D	0.01	2.3 ms	27 ms	12x
Annulus 2D	0.1	1.6 ms	14 ms	9x
3D torus	1.0	12 ms	120 ms	10x

Usage

# Existing (unchanged):
S = Separator(x^2 + y^2 <= 1, [x, y])

# New DAG-based:
S = DAGSeparator(x^2 + y^2 <= 1, [x, y])

# Both work with pave:
inner, boundary = pave(X, S, 0.1)

Test plan

• 27 new tests in test/test_dag.jl — DAG construction, forward/backward propagation, contractor, separator, paving, comparison with original, combined separators, 3D
• All existing tests in test/runtests.jl still pass
• Benchmarks in benchmark/bench_dag_vs_codegen.jl confirm identical box counts

🤖 Generated with Claude Code