test: wire suite helpers, topics, and drop cantilever Gmsh mesh

The default harness loads shared helpers and smoke checks before topic
includes, registers poroelastic and thermo_poroelastic, drops the retired IO
topic, and removes the orphaned ASCII fixture.

- Add zero_alloc_helpers.jl, test_api_contract.jl, test_pkg_hygiene.jl, and
  test_inference_smoke.jl included once from runtests.jl
- Extend TOPICS; remove io; clarify test_args accepts topic names only
- Delete test/testdata/cantilever_beam.msh

test/zero_alloc_helpers.jl

@@ -0,0 +1,76 @@
+# SPDX-FileCopyrightText: 2015-2026 Jukka Aho
+# SPDX-License-Identifier: MIT
+
+"""
+Shared helpers for zero-allocation regression tests.
+
+`@allocated` can miss some paths or mask one-shot caches; scanning optimized
+LLVM IR for Julia GC entrypoints catches heap traffic inside compiled code.
+Names differ across Julia versions (`jl_*` vs `ijl_*`, `julia.gc_alloc`, …);
+`count_llvm_gc_alloc_sites` tracks the union used in practice.
+
+Note: `Pkg.test` runs Julia with `--check-bounds=yes` (`JLOptions().check_bounds == 1`).
+Some small indexed kernels then show GC-related `call` lines in optimized IR even when
+`@allocated` is zero. Prefer `@allocated` plus LLVM scans for large stable kernels
+(`assemble!`, `apply_K!`, …); for tight index loops, gate IR scans or accept that CI
+skips the IR assertion.
+"""
+
+using InteractiveUtils: code_llvm
+
+# Every needle must appear only inside a `call` line in IR we care about.
+const _LLVM_GC_ALLOC_MARKERS = (
+    "julia.gc_alloc",
+    "jl_gc_pool_alloc",
+    "jl_gc_big_alloc",
+    "jl_gc_alloc_typed",
+    "jl_gc_small_alloc",
+    "ijl_gc_pool_alloc",
+    "ijl_gc_big_alloc",
+    "ijl_gc_alloc_typed",
+    "ijl_gc_small_alloc",
+)
+
+"""
+    count_llvm_gc_alloc_sites(ir::AbstractString) -> Int
+
+Count LLVM `call` lines that invoke a Julia GC allocation runtime function.
+Used on the output of `llvm_ir` / `InteractiveUtils.code_llvm`.
+"""
+function count_llvm_gc_alloc_sites(ir::AbstractString)
+    n = 0
+    for line in eachsplit(ir, '\n')
+        occursin("call ", line) || continue
+        for needle in _LLVM_GC_ALLOC_MARKERS
+            if occursin(needle, line)
+                n += 1
+                break
+            end
+        end
+    end
+    return n
+end
+
+"""
+    llvm_ir(f, ::Type{<:Tuple}; optimize::Bool=true) -> String
+
+Optimized LLVM IR for `f` at argument tuple type `T`, as a single string.
+"""
+function llvm_ir(@nospecialize(f), @nospecialize(T::Type{<:Tuple}); optimize::Bool = true)
+    io = IOBuffer()
+    code_llvm(io, f, T; optimize = optimize)
+    return String(take!(io))
+end
+
+"""
+    llvm_gc_alloc_site_count(f, ::Type{<:Tuple}; optimize::Bool=true) -> Int
+
+Convenience: `count_llvm_gc_alloc_sites` ∘ `llvm_ir`.
+"""
+function llvm_gc_alloc_site_count(
+    @nospecialize(f),
+    @nospecialize(T::Type{<:Tuple});
+    optimize::Bool = true,
+)
+    return count_llvm_gc_alloc_sites(llvm_ir(f, T; optimize = optimize))
+end