mrg_build: YoWASP WASM toolchain backend (Docker-free local synth/PnR, #8)

[santolucito]

Implements the mrg_build half of #8: a toolchain-abstraction layer that resolves each tool to either a native binary on $PATH (oss-cad-suite / the sandbox image) or a YoWASP WASM entry point (yowasp-yosys, yowasp-nextpnr-ecp5) — so local synth/PnR reports work in Docker-less environments with nothing but pip wheels.

What's in this PR

• toolchain.backend() — picks native (tools on PATH, preferred) or wasm (YoWASP wheels importable) once per run, so a report is never a native/wasm hybrid. Forceable via MRG_TOOLCHAIN_BACKEND=native|wasm for parity testing.
• WASM execution via subprocess shim — YoWASP entry points (yowasp_yosys.run_yosys(...)) are invoked in a sys.executable -c subprocess, so cwd, env, and output capture behave identically to the native tools. All call sites (synth, pnr, pnr_soc — including the LiteX build-script lines) route through one _cmd() resolver.
• BuildReport.backend — every report now records which backend produced it (native / wasm; the SDK's Docker path can set docker), alongside the existing toolchain version string (Proposal: [local] extra — Docker-free local synth/PnR via YoWASP #8 acceptance criterion).
• Error messages — a missing toolchain now suggests pip install 'manhattan-reasoning-gym[local]' before oss-cad-suite (Proposal: [local] extra — Docker-free local synth/PnR via YoWASP #8 acceptance criterion).
• Tests — test_toolchain.py is parametrized over both backends, each skipping independently when its backend isn't installed, plus unit tests for backend selection and the error message. Also fixed a vacuous pass in the determinism test (None == None when both runs fail).
• README — documents the two backends under sandbox/.

Verified

On this host (macOS, arm64), with yowasp-yosys==0.66.0.0.post1165 and yowasp-nextpnr-ecp5==0.10.0.0.post753:

• wasm synth on tests/fixtures/mac.v: ok, exactly 1 MULT18X18D, backend/toolchain recorded (Yosys 0.66 / nextpnr-0.10, backend: "wasm").
• wasm PnR: fits, Fmax 339 MHz, timing_met at 65 MHz, DSP 1/156, deterministic across runs with the same seed.
• Absolute host paths resolve fine through YoWASP's wasmtime runtime (no cwd-only restriction).
• pytest tests/ — all wasm-backend and backend-selection tests pass; ruff check clean.

Note: the three [native] test failures on my host are pre-existing (also fail on main): the local ~/oss-cad-suite bundle is a stale yosys 0.47 that misparses the quoted tee -o filename, vs the repo's 0.62 pin. The image's native path is unaffected — argv construction for the native backend is byte-identical to before.

Out of scope (tracked in #8)

• The SDK-side fallback order in _local_build.py (native → wasm → Docker → error) and the [local] extra in pyproject.toml with pinned wheel versions — those live in the SDK repo.
• Shipping the pre-elaborated base SoC for full-SoC pnr without LiteX installed.
• Version pinning of the YoWASP wheels to the sandbox image's yosys/nextpnr builds (the wheels used here are current releases, newer than the image pin).

Closes nothing on its own; part of #8.

🤖 Generated with Claude Code