pred path with an instance: one failing path aborts the whole batch

[isPANN]

Summary

When pred path runs on a source instance, one failing path aborts the entire batch and shows no results for the paths that succeeded.

pred create MIS --graph 0-1,1-2,2-3,3-0,0-2 -o mis.json
pred path MIS QUBO mis.json
# Error: concrete path 4 failed during reduction: ILP -> ILP: binary encoding requires a finite upper bound ...
pred path MIS QUBO mis.json --limit 4   # works: paths 1–4 succeed

Root cause

ReductionGraph::execute_paths (src/rules/graph.rs) returns Result<Vec<ExecutedPath>, ExecutePathsError> and uses ? on each reduction step. A reduction that rejects one particular instance is a valid outcome for that path, not a fatal error for the whole request.

Paths legitimately fail for particular instances:

• ILP<i64> → ILP<bool> rejects unbounded variables (ILP<i64> rules declare bounds as constraints, so ILP<i64> → ILP<bool> always fails #1176).
• MaxCut → MinimumMatrixCover rejects negative edge weights on the i64 variant: edge (u, v) has negative weight -w.

Proposal

• Return one outcome per path: executed (with measured parameters) or failed (with failing step index and ReductionError). Keep batch-level errors (empty path, different sources, unknown node) as errors.
• Keep prefix sharing. A failed prefix should mark every path that shares it as failed without running it again.
• Update the text and JSON output of pred path and the MCP tool so failed paths show up with their reason instead of aborting.
• Tests: a batch with one failing and one succeeding path, and a shared failing prefix.

Related: #1152, #1178 (size budget), #1179 (path ranking).