test: move slow emit-zig behavioral smoke out of default zig build test

[jeffersonmourak]

Summary

The dev-loop zig build test had grown to ~221 s (warm), and almost all of it was a single serial tail. Two emit-zig behavioral suites (tests/emit/behavior_test.zig, tests/emit/project_behavior_test.zig) each spawn a nested zig build wasm per fixture — ~33 cache-missing WebAssembly compiles per run, one of which alone accounted for ~180 s.

Those suites compile a circuit through the experimental --emit-zig backend (standalone Zig source → wasm) and assert its runtime behavior. But that same behavior, over the same 32-fixture corpus and the same expected vectors, is already validated far more cheaply by tests/e2e/serializer_fixtures_test.zig, which drives the production path: splice topology sections into the prebuilt embedded runtime and run via Node, with no per-fixture compile. The two emit suites were therefore ~100% redundant for behavioral coverage; their only unique value is smoke-testing the emit-zig backend itself.

Change

• Trim the two emit-zig suites to a small backend smoke (2 fixtures each, covering the single-file and multi-file project emit paths).
• Split the build into three steps:
  • zig build test — fast dev-loop suite (default); full circuit-behavior coverage retained via serializer_fixtures_test.
  • zig build test-emit — opt-in emit-zig backend smoke (the nested-build suites).
  • zig build test-all — test + test-emit; the full gate.
• Point pr-tests.yml at test-all so emit-backend regressions still block merges (the job name is unchanged to preserve the required-check identity).
• Add a build-once guard to tests/cli/integration_test.zig, which was invoking zig build circ-compile 16 times per run.

Result

	before	after
zig build test (dev loop)	~221 s	~31 s (7×)
zig build test-all (CI)	—	~27 s

Coverage was verified intact: deliberately breaking the AND-gate evaluator turns the fast test red via serializer_fixtures_test (plus the truth-table suites).

A prior round of test-speed work (96 s → 25 s, via a persistent wasm workspace) is recorded in DOCS/archive/plan-test-speed.md. A natural follow-up — grouping the ~57 per-module test binaries and reusing the installed CLI binary in the integration test instead of rebuilding it — would push the dev loop below ~31 s.

[github-actions]

Engine bench: deltas vs main

(bench produced no JSON output; see workflow logs for the actual error.)

See the workflow run logs for the actual error.