Depends on #2197.

Part of #2152 (stage T).

Problem

Wasm currently selects nogc because BDWGC is unavailable. That leaves J32, J64, and P1 without a runtime-owned allocator, reclamation, memory statistics, or linear-memory growth policy.

Enabling collection by default is not yet correct: LLGo may keep a live Go pointer only in an SSA/wasm local across runtime.GC, and a linear-stack scan cannot observe that value. Suspended-G root publication is also intentionally deferred to stage D. This PR therefore provides the collector behind the temporary internal llgo_wasm_gc build tag while preserving the existing default nogc behavior.

Implementation

• Refactor the existing bare-metal block allocator and conservative mark/sweep core behind narrow memory-layout, growth, root-scan, and stack-stat hooks.
• Reuse the same runtime integration for bare-metal and wasm instead of duplicating allocation, defer-node, and runtime.ReadMemStats shims.
• Add J32/J64/P1 adapters based on __global_base, __data_end, __heap_base, the active linear stack boundary, and WebAssembly page growth.
• Relocate collector metadata when linear memory grows and retain at least one MiB of growth per step.
• Route Emscripten standard and emscripten_builtin_* allocation entry points to the collector. -sMALLOC=none is injected only when llgo_wasm_gc is selected, so default wasm builds retain their current allocator.
• Keep finalizers and cleanup hooks unsupported in this initial collector.
• Reject llgo_wasm_gc with threaded P1 builds; the collector is intentionally single-worker and requires LLGO_WASI_THREADS=0 for P1.

The collector is single-worker. It scans linear-memory globals and the active linear stack, but does not claim that wasm locals or suspended Asyncify contexts are roots. Stage D will add compiler-maintained root records and then remove the opt-in gate.

Validation

The fixture covers:

• runtime.ReadMemStats allocation and reclamation counters;
• global-root retention and recursive retention through a global slice;
• reclamation of 1,024 unreachable objects;
• 64 KiB aligned Emscripten allocation and data integrity;
• bounded linear-memory growth past each target initial heap, followed by collection and live-data verification.

Local bounded checks used GOMAXPROCS=2 and -p=1. The full macOS internal/build and internal/crosscompile tests passed with 73.9% and 80.9% package coverage; both newly added build helpers report 100% coverage. The runtime module passed on macOS and in an Ubuntu 24.04 container limited to 15 GiB and 2 CPUs.

The updated #2197-based stack passes the full GitHub Ubuntu/macOS matrix and Codecov; local macOS target and coverage checks also pass.

Native and embedded cost

Compared with the direct #2197 base using identical inputs:

• native Mach-O section sizes are identical in every section, total 67,945 bytes; only the linker UUID differs;
• cortex-m-qemu embedded ELFs are byte-for-byte identical (same SHA-256), with .text 92 bytes and total sections 4,353 bytes.

Wasm-only sources and the allocator linker option are excluded unless the internal tag is selected, so this PR adds no default native, embedded, or wasm runtime path.

Opt-in wasm size

Using the same minimal wasm-runtime fixture for both modes:

These bytes are paid only by explicitly tagged builds. Runtime throughput and pause measurements remain follow-up data once compiler-maintained roots make collection semantically usable for general programs.

Review scope

Excluding #2197, this PR changes 29 files with 546 additions and 62 deletions. The first commit is a behavior-preserving platform-hook refactor; the second contains the wasm facility; later commits add CI, make the growth test target-relative, reject unsafe threaded-WASI selection, and preserve the official nil-argument behavior.