# Local CDSL Engine

This package rebuilds `cad.cdsl.llm.v1` models through the CDSL-only path:

`semantic validation -> capability analysis -> sketch resolution -> session runtime -> STEP`

## Package layout

| Module | Responsibility |
|---|---|
| `specs.py` | Vector math, plane/axis helpers, parametric feature specs (no kernel deps) |
| `topology.py` | Diagnostics, planning contracts, `TopologyRegistry`, selector resolution |
| `capabilities.py` | `CapabilityAnalyzer`: preflight blockers before any geometry runs |
| `sketch_solver.py` | Profile expansion (circle / polygon / analytic contours) |
| `session.py` | `ExecutionSession` and the kernel-facing `GeometryAdapter` protocol |
| `runtime_base.py` | Shared error types and the `ExtentVector` value |
| `extents.py` | End-condition planning (blind / through / up-to-surface / ...) |
| `pattern_transform.py` | Translate/mirror/rotate parameter algebra for pattern replay |
| `registry.py` | `EXECUTORS`, the `atomic_executor` decorator, and `execute_node` dispatch |
| `executors/` | One module per executor family; importing the package registers all |
| `build123d_adapter.py` | The only layer that creates or mutates B-rep objects |
| `topology.py` / `runtime_types.py` | Historical re-export shim (`runtime_types`) |
| `translator/` | Frozen SolidWorks-exact code generation (`ir` / `codegen` / `runtime_lib`) |
| `legacy/` | Frozen legacy engine paths (`llm_compiler`, `llm_engine`, exact rebuilds) |
| `rebuild.py` | Legacy three-way facade plus `compare_with_gold` acceptance |

`runtime.py` keeps the `analyze_cdsl` / `rebuild_cdsl` entry points and
re-exports the historical names. `llm_compiler.py` and `llm_engine.py` are
compatibility shims for the frozen `legacy/` implementations and are not used
by `run_cdsl_only`.

## Adding an atomic operation

1. Add the operation contract to `profile_schema.json` (`operation_contracts`),
   including its `runtime_capability` flags — this is the single source of
   truth for preflight classification.
2. Add a decorated executor function in exactly one `executors/<family>.py`
   module (`@atomic_executor("...")`); the shared registry never changes.
3. Add the atomic id to `ALL_ATOMIC_IDS` in `registry.py` (registration fails
   fast on unknown or duplicate ids, and `executors/__init__` fails if any
   declared id has no registered executor).
4. Update `cdsl_schema.json` in the same change.
5. Extend `backend/tests/test_profile_schema.py` fixtures if the contract
   shape changed.

Multiple people can add different operations in parallel without touching a
shared registry file: the only shared edits are the two schema documents.

Supported profiles are defined by `SHAPE_GENERATORS` in `sketch_solver.py`.
Supported feature atomic operations are defined by `EXECUTORS` (populated from
`executors/` at import time). Their human-readable contract is in
`profile_schema.json`; the complete, machine-enforced CDSL object contract is
in `cdsl_schema.json`.
The Studio only accepts self-contained profile data and requires successful
`engine=cdsl_only` output. It never uses the legacy translator fallback or
`compiler_context`.

## Engine schema maintenance

`profile_schema.json` and `cdsl_schema.json` together are the source of truth
for the engine contract exposed to the CAD Agent and the backend validator.
Any addition, removal, rename, or parameter-contract change in
`sketch_solver.py`, `executors/`, or the build adapter must update both files
in the same change.
`backend/tests/test_profile_schema.py` fails when the registered profiles,
supported atomic operations, or `runtime_capability` flags diverge from the
document.

## Batch baseline

Use the resumable batch entry point to produce feature-level eligibility and
rebuild reports. `--build` invokes only the session-based CDSL runtime; it
never falls back to `compiler_context` or the legacy translator.

```bash
PYTHONPATH=backend/engine python -m cdsl_engine.batch_rebuild \
  json_to_cdsl/output --out /tmp/cdsl-batch --build --build-timeout 15
```

The output directory contains `manifest.json`, one report per part under
`parts/`, `summary-by-atomic.json`, and `summary-by-blocker.json`. Re-run the
same command to resume completed work; use `--max-parts` to run a bounded CI
shard.

Use `--part-ids 046112,053393` to run an exact, comma-separated regression
subset. Unknown ids are rejected so a phase baseline cannot silently omit a
requested part.

Use `--phase p3`, `--phase p4`, or `--phase p6` to run a documented
strict-closed static pool. The
selector lives in `phase_pools.py`; its membership is regression-tested
against the committed exports rather than copied into a shell command.

The runtime foundation test suite also analyzes every committed export without
building STEP, so CI verifies that the full corpus always yields one
machine-readable capability result per input. Use the batch command above for
the slower, resumable truth-build layer.