Fan-in / frontier-expansion ordering — port high-fan-in scaffolding first

[Tokarzewski]

Part of #64.

Problem

Pure leaf-first ordering is locally safe but globally slow: each step frees only the next thin layer. The foundational tier (OutputProcessor, DataLoopNode, ScheduleManager, the Array1D/constants types) gates most mid-tier modules — porting these expands the leaf frontier the most, even though they are not leaves.

Approach

• Rank dependency-graph nodes by fan-in (how many downstream fns/modules each unblocks); optionally PageRank for transitive unlock.
• Migration loop: port the max-unlock scaffolding → recompute the frontier → sweep newly-freed leaves with the model + verify gate.
• Feed the now-ported deps' signatures into scripts/sft/dep_context.py / the turbovec RAG (scripts/rag/) so the 1.5B stops emitting unknown symbols.

Done when

• migration_plan.py can emit a fan-in-ranked order (not just topological leaves).
• Empirically, frontier size grows faster than under pure leaf-first on the EnergyPlus graph.

🤖 Generated with Claude Code

[Tokarzewski]

Done — landed on main in merge 90a2971 (feature commit 2baa310).

scripts/sft/migration_plan.py now computes fan_in/fanin_ranked and reports top_fanin + top_fanin_scaffolding (max-unlock, non-leaf functions to port first). The cbm-backed graph (cbm_graph.py) annotates callables with fan-in over the live EnergyPlus call graph too. Covered by unit tests in tests/test_migration_plan.py. Part of #64.