God-object vertical slicing — path-specific EnergyPlusData slice + per-module sub-state split

[Tokarzewski]

Part of #64.

Problem

EnergyPlusData is in everyone's closure — it gates ~96% of compile failures, is too big to port wholesale, and impossible to avoid (every fn takes state). Porting it fully blocks everything; not porting it blocks everything.

Approach

Do not port the object — port a path-specific slice, the Route-B prune that already worked (Minimal.idf: 46→22 fields, cone 152→23):

• Pick one runtime scenario (e.g. Minimal.idf), trace which state.dataXxx->... fields are actually read/written on that path, port only those.
• Split the god-object into per-module sub-states (dataCoolTower, dataInput, ...) — already namespaced in C++ — so each fn pulls in only the sub-struct it uses.
• Keep /home/bart/Github/EnergyPlus/ READ-ONLY as the oracle for the slice.

Done when

• A minimal viable EnergyPlusData slice compiles and supports one IDF end-to-end.
• Functions depend on per-module sub-states, not the full god-object.

Note: end-to-end-runnable for one IDF beats fully-ported-for-nothing.

🤖 Generated with Claude Code

[Tokarzewski]

Slicer landed (main 30308a7) + finding on the data source.

scripts/sft/god_object_slice.py now computes the slice two ways:

• reachability (slice_manifest): transitive calls closure from runtime entries → touched fields, grouped per sub-state.
• field-centric (field_centric_slice): per-sub-state field inventory + writer/reader functions, independent of call-graph completeness.

Finding — cbm's C++ edge resolution is too sparse for reachability slicing (for now):

• Only 516 / 9048 callables have any outgoing CALLS edge; heavy functions like CalcHeatBalanceInsideSurf have 0 CALLS even unfiltered. C++ call resolution (overloads, virtual dispatch, macros) is hard and this index kept only high-confidence edges.
• So transitive-closure slicing returns near-empty cones today.

What works now (field edges are dense — 14,704 writes / 3,792 reads):

• Field-centric slice over the EnergyPlusData family: 566 sub-states; top by fields touched: OutputReportPredefinedData (1217), EnergyPlusData core (252), DXCoilData (161), ZoneSizingData (103). This is the per-module sub-state split this issue calls for.
• Caveat: writer/reader attribution under-counts — state.dataXxx->field = … member-access writes aren't resolved to the Field (EnergyPlusData shows 2 writers for 252 fields). The field inventory is reliable; the function attribution is a floor.

To unblock the rest of #69: improve the C++ call/write resolution feeding the graph — e.g. re-index with a clang-based resolver, or augment cbm edges with a libclang AST pass (ties into the ground-truth extraction already in the strict engine). Until then, use field-centric inventory to scope sub-states and pair with the regex planner's verifiable leaf frontier. Part of #64.

[Tokarzewski]

Progress — per-module sub-state codegen landed (PR #76, draft).

The slicer already emits the manifest (reachable callables → touched fields →
grouped by owning sub-state). PR #76 adds the codegen half:
emit_substate_structs(manifest) turns a slice into struct data<Module>
scaffolds holding only the sliced fields + a StateSlice container — so a
function depends on state.dataHeatBalance.<field> instead of the full
EnergyPlusData. --emit-mojo writes it; 6 pure-graph pytest cases verify it
(unreached fields excluded, read+write dedup, Data prefix dropped).

Toolchain-gated remainder (deferred — no Mojo here), to hit "done when":

1. Real types, not name heuristics. The graph carries field names, not C++
   types, so emitted fields default to Float64/Int/Bool by name. Join the slice
   against the oracle headers (Xxx.hh struct XxxData) — the same source the
   keystone port reads — to stamp exact types before compiling.
2. Pick the runtime scenario + entries. Use Minimal.idf and the verified
   Route-B prune (46→22 fields, cone 152→23) as the first slice: run
   god_object_slice.py --entry <runtime entry fns> --emit-mojo slice.mojo.
3. Compile the slice. Feed the emitted structs through the Mojo build; fix the
   handful of type/field gaps (this is where Dependency-boundary shim library (ObjexxFCL Array, Constant::*, libm intrinsics) #66 shims + the field-default-fixer
   come in). The slice should compile standalone before wiring.
4. End-to-end one IDF. Rewire the runtime-path functions to take the
   per-module sub-state instead of EnergyPlusData, and run Minimal.idf through
   to a matching .eso vs the C++ oracle. That closes the issue.

Items 1–4 are build/verify chores on a Mojo box; the slice→struct codegen they
build on is the pytest-verifiable piece, now in place.