Epic: Tackle dependency closure for C++→Mojo migration

[Tokarzewski]

Parent epic for the dependency-closure strategy. On real EnergyPlus code the limiter is dependency closure, not translation skill — the 1.5B draft model is strong, but real functions call siblings, touch the EnergyPlusData god-object, and link libs, so they cannot be ported or verified in isolation.

Reframe: closure is two coupled problems

• Verify-closure — to confirm F is correct its callees must produce real values → seems to need them ported first.
• Compile-closure — to make F's Mojo build, every symbol/type it references must exist.

Leaf-first ordering attacks both at once, which is why it stalls. The leverage is to decouple them: make correctness independent of porting order (record/replay), and treat compile-closure as a build-ordering chore (shims + fan-in order). Correctness stays guaranteed at every step by the C++ numeric oracle (/home/bart/Github/EnergyPlus/).

Children

• C++ I/O recorder + Mojo replay shim — decouple verify-closure from compile-closure #65 — C++ I/O recorder + Mojo replay shim — MVP landed (main 787343c): scalar record/replay decouples verify-closure, proven end-to-end; bool/int/Array deps + gate-integration remain
• Dependency-boundary shim library (ObjexxFCL Array, Constant::*, libm intrinsics) #66 — Dependency-boundary shim library — Array1D/Array2D + Constant + libm done (main 61db4bf); psychrometric shims remain
• SCC/cycle detection + trait-seam cycle breaking in migration ordering #67 — SCC/cycle detection + trait-seam cycle breaking in migration ordering ✅ landed in main (90a2971): tarjan_scc/cyclic_components in migration_plan.py, SCC-condensed layering, 18 unit tests
• Fan-in / frontier-expansion ordering — port high-fan-in scaffolding first #68 — Fan-in / frontier-expansion ordering (port high-fan-in scaffolding first) ✅ landed in main (90a2971): fan_in/fanin_ranked + top_fanin_scaffolding in the plan
• God-object vertical slicing — path-specific EnergyPlusData slice + per-module sub-state split #69 — God-object vertical slicing (path-specific EnergyPlusData slice + per-module sub-state split) ⚠️ enabling graph only: cbm_graph.py now emits writes_field/reads_field edges with Field owners (live: 14,704 writes / 3,792 reads over the EnergyPlusData family), so the slice is queryable. The actual path-specific slice + per-module sub-state port is not done — keep open.

Progress note

The directed-graph foundation for #67/#68/#69 is on main:

• A persisted directed graph (node-link JSON) replaces the flat layer summary.
• cbm_graph.py queries the codebase-memory-mcp knowledge graph (via its CLI bridge) to build a multi-relational directed graph — calls + writes_field + reads_field — over the indexed EnergyPlus C++ oracle (18,996 nodes / ~37k edges).

Shape of the approach

Top-down to define the seams (runtime path, god-object slice, trait boundaries); bottom-up to fill and verify leaves along that path — with record/replay so verification never waits for the closure. Compile-closure becomes a build-ordering chore; correctness is guaranteed at every step by the oracle.

Relates to north-star #56.

🤖 Generated with Claude Code

[Tokarzewski]

Status — dependency-closure epic, this session's increments.

Child issues moved (all draft PRs against main, pytest-green, no Mojo/g++ run
since that toolchain is unavailable in the current env):

• Dependency-boundary shim library (ObjexxFCL Array, Constant::*, libm intrinsics) #66 (shim library) — first real psychrometric shim landed (PR feat(#66): real PsyPsatFnTemp psychrometric shim (replaces pvstar stub) #75):
  PsyPsatFnTemp ported faithfully from the oracle PsyPsatFnTemp_raw, Python
  reference + Mojo/C++ shim codegen, 12 pytest cases vs textbook saturation
  pressures. Replaces the numerically-wrong pvstar stub. Mojo↔C++ compile-match
  gate deferred (plan on Dependency-boundary shim library (ObjexxFCL Array, Constant::*, libm intrinsics) #66).
• God-object vertical slicing — path-specific EnergyPlusData slice + per-module sub-state split #69 (god-object slicing) — codegen half landed (PR feat(#69): emit per-module sub-state Mojo structs from a god-object slice #76):
  emit_substate_structs turns a slice manifest into per-module data<Module>
  Mojo structs + a StateSlice container, so a fn pulls in its sub-state instead
  of the full EnergyPlusData. 6 pure-graph pytest cases. Path-specific
  end-to-end port for one IDF deferred (plan on God-object vertical slicing — path-specific EnergyPlusData slice + per-module sub-state split #69).

The directed-graph foundation (#67/#68 done, #69 enabling edges) plus the
record/replay decoupling (#65) remain the backbone; the two increments above chip
at the two concrete remainders (psychro shims, sub-state split). No change to the
#65 scalar record/replay MVP this session (owned by another agent).

Closure-coverage instrumentation (the north-star metric, #56) is the suggested
next lever: stand up the % number on Minimal.idf's pruned cone so the slice work
above reports a coverage delta rather than just compiling.