ci: add a release-mode lane so C-LEARN-scale tests can be un-ignored

[bpowers]

Problem

All C-LEARN-scale integration tests are marked #[ignore] so they don't blow the 3-minute cargo test --workspace debug-mode wall-clock cap (see docs/dev/rust.md test-time budgets and the CLAUDE.md "Hard Rules" section). Affected tests include (in src/simlin-engine/tests/):

• simulates_clearn
• clearn_with_ltm_simulates_model_vars_identically
• clearn_ltm_discovery_compiles
• clearn_pinned_climate_loop_is_scored
• simulates_clearn_wasm

Because they're #[ignore]d, they only run when someone remembers to invoke them manually (cargo test -- --ignored). Regressions in C-LEARN-scale behavior are therefore not caught automatically by CI.

Why it's now worth fixing

Measured 2026-06-01: a single C-LEARN test takes 39-60s in debug mode (parse + LTM compile alone is ~39s), i.e. 22-33% of the entire workspace budget -- which is exactly why they're ignored. But in release mode the same test is only 6-15s, thanks to the recent LTM performance work (the LTM compile dropped from 47.6s to 4.0s in release; see commits 477af2a / 94f9c07 and the recent perf session).

The perf wins make these tests cheap enough in release mode that a dedicated CI lane running them automatically is now practical, where it would not have been before.

Why it matters

• Correctness/regression coverage: C-LEARN is the largest real model in the corpus and exercises code paths (large-model LTM compile, fragment assembly, pinned-loop scoring, WASM sim) that smaller fixtures don't. Right now nothing guards those paths in CI.
• Developer experience: relying on humans to remember --ignored is fragile; the recent perf regressions/fixes in this area show the paths are actively churning.

Proposed approach

Add a separate CI lane in .github/workflows/ci.yaml (or a scheduled/nightly job) that runs the ignored C-LEARN-scale tests in release mode, e.g.:

cargo test --release -p simlin-engine -- --ignored <clearn test filters>

This stays off the 3-minute debug cargo test --workspace path (which remains the canonical PR gate) while still exercising the C-LEARN tests automatically. Options to weigh:

• A dedicated release-mode job in the existing PR CI matrix (runs on every PR but only the C-LEARN-scale ignored tests, in release).
• A schedule:-triggered nightly job (lower PR cost, slightly delayed signal).

Either way, this is a CI-config change; it does not require un-marking the #[ignore] attributes for the default debug flow.

Components affected

• CI configuration: .github/workflows/ci.yaml
• The #[ignore]d tests in src/simlin-engine/tests/ (no source change needed beyond possibly a shared filter/feature to select them)

Related

• Closed engine: gate wrld3_element_level_enumeration_is_uncapped behind cfg(not(debug_assertions)) #465 gated wrld3_element_level_enumeration_is_uncapped behind release-only for the same debug-vs-release cost asymmetry -- precedent for release-gated heavy tests.
• tech-debt.md item 13 (Ignored Rust Tests) catalogs the ignored-test inventory; item 60 (No CI feature matrix) is an adjacent CI-gap item but distinct.

Discovery context

Identified during the GH #653 pinned-loop work (design plan docs/design-plans/2026-06-01-ltm-653-pinned-loop-dimensions.md, "Additional Considerations" section records the timing measurements). The question was whether the new C-LEARN pin test (clearn_pinned_climate_loop_is_scored) could be un-ignored given the recent perf improvements; the answer was no for debug mode, yes if a release-mode lane existed.