feat(mir): Phase 69-3 Fix MIR non-determinism with BTreeSet

Replace HashSet with BTreeSet for CFG predecessors/successors:
- BasicBlock.predecessors: HashSet → BTreeSet
- BasicBlock.successors: HashSet → BTreeSet
- LoopFormOps.get_block_predecessors(): returns BTreeSet
- BasicBlock.dominates(): takes &[BTreeSet<BasicBlockId>]

This ensures deterministic PHI generation and test stability.

Test results:
- loop_with_continue_and_break tests: now deterministic (3/3 same output)
- loopform tests: 14/14 PASS (no regressions)
- merge_exit_with_classification tests: 3/3 PASS

Technical changes (6 files):
- basic_block.rs: BTreeSet types + new() initialization
- loopform_builder.rs: trait signature + 2 mock implementations
- phi_ops.rs: return type
- json_v0_bridge/loop_.rs: return type

Same pattern as Phase 25.1 (MirFunction.blocks HashMap → BTreeMap).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

src/mir/loop_builder/loop_form.rs

@@ -395,7 +395,7 @@ impl<'a> LoopBuilder<'a> {
                             merge_block.add_instruction(MirInstruction::Phi {
                                 dst: phi_id,
                                 inputs: final_inputs,
-                type_hint: None, // Phase 63-6
+                                type_hint: None, // Phase 63-6
                             });
                         }
                     }