Fast PR CI: small hostile-test caps, heavy tests in a separate job (#43)

* test: hostile-params test runs under small session work caps

Session gains Limits, the defaults of mesh.tet.generate's max_steiner and
max_tets (production values unchanged; a parameter still overrides them;
file.open keeps them). The registry-wide hostile-params test runs on
sessions with small caps, so the needle-triangle tessellation it meshes is
refused at once instead of meshing to the 50k Steiner cap: the engine test
binary drops from ~350 s to ~20 s locally.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NmAttHhUZTokgPeEJia8sm

* ci: move heavy tests to a heavy job on main, nightly and on demand

Four slow tests are #[ignore = "heavy: ..."] (every corpus recipe, the
cantilever convergence study, the explicit-vs-modal cantilever period, the
generic-predicate rational cross-check); fast representative tests of each
area stay in the PR gate. cargo xtask test --heavy runs every ignored test
(benchmarks included) and the doc tests; cargo xtask test and ci run the
rest without the doc tests. The new heavy CI job runs it on pushes to main,
nightly and on workflow_dispatch. Branch jobs restore main's rust cache and
no longer save their own. AGENTS.md documents the rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NmAttHhUZTokgPeEJia8sm

* test: the three slow LBM verification cases run in the heavy job

The Stokes-drag sphere array, the staircase-pipe convergence study and the
developing-channel case are heavy (21 s, 11 s, 6 s alone); plane
Poiseuille, Couette, Taylor-Green, reproducibility and the stability
limits stay in the PR gate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NmAttHhUZTokgPeEJia8sm

---------

Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

AGENTS.md

@@ -97,6 +97,14 @@ People trust ModelCraft with their models; a crash loses their work. **This outr
   warnings, tests, wasm) on every PR; don't run the full gate locally. Before pushing: `cargo fmt
   --all`, `cargo check` the touched crates, and run the tests you added or changed. Commit after
   each working step (a layer that compiles with its tests passing), not one big commit per PR.
+- **Keep the PR gate fast** (under ~3 minutes on CI). A test that takes more than a few seconds
+  (convergence studies, the whole corpus, cross-checks between solvers) is
+  `#[ignore = "heavy: cargo xtask test --heavy (CI heavy job)"]`, with a fast representative
+  test left in the PR gate. The `heavy` CI job runs `cargo xtask test --heavy` (every ignored
+  test, benchmarks included, and the doc tests) on pushes to `main`, nightly and on demand
+  (`gh workflow run CI --ref <branch>`). Hostile-input tests run commands on sessions with small
+  work caps (`Session::limits`) so refusals are instant; never lower the production caps for a
+  test.
 - **Publishing**: `origin` is github.com/jyviko/modelcraft. Changes land through pull requests:
   branch, **rebase onto the latest `origin/main` right before opening the PR and before every push
   that updates it** (never merge `main` into a branch), push, open a PR, let CI run. A PR merges