[Based on #139] debug: validate fault boundaries and final Wasm DWARF

[cpunion]

Continues xgo-dev#2164 on top of #139.

This change closes native fault/boundary acceptance gaps and makes clean final WebAssembly DWARF a hard gate without changing runtime or scheduler behavior:

• compiles C, C++, cgo preambles, and LLGoFiles with DWARF whenever the shared typed debug policy emits Go DWARF;
• verifies exact user source locations for explicit panic, integer division by zero, invalid memory, and a real host trap;
• verifies ordered Go -> C -> Go callback frames;
• builds a separate -O2 artifact, requires nested LLVM inline frames, and steps back to the physical caller;
• temporarily pins the conservative Binaryen DWARF range repair from Repair invalid DWARF scope ranges after transforms WebAssembly/binaryen#8964;
• strictly verifies the final Asyncify-transformed browser embedded Wasm and external DWARF sidecar with llvm-dwarfdump --verify;
• makes final WASI verification mandatory and covers both embedded and external DWARF artifacts;
• runs the native fixture in the existing Darwin/Linux primary LLDB lane, the WASI gates in llgo.yml/wasi-debug.yml, and the browser fixture in Chrome 151.

Validation:

• python3 check.py wasm-opt in the pinned Binaryen tree
• 73-CU LLGo + Emscripten final embedded browser Wasm: llvm-dwarfdump --verify reports No errors.
• browser external deployment module and .debug.wasm sidecar: No errors.
• 90-CU WASI embedded artifact and external .debug.wasm sidecar: No errors.
• go test -timeout 10m ./internal/debugabi ./internal/wasmdebug ./internal/browserdebug ./cmd/internal/browser ./cmd/internal/debug
• full go test ./internal/build: pass in 290.468s after the final-gate audit
• native LLDB acceptance on Darwin/arm64 with Apple LLDB 17: NATIVE_DEBUG_ACCEPTANCE_OK
• native LLDB acceptance on Linux/x86_64 with LLDB 19.1.7: NATIVE_DEBUG_ACCEPTANCE_OK
• Node language-extension tests, shell syntax, workflow YAML, and diff validation

The Binaryen pin can return to the supported Emscripten release after the upstream repair lands and is released.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[github-actions]

LLGo baseline benchmarks

6032fec376e8 | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs main	Build	vs main	Run	vs main
Linux	cprintf	20592 B	+11.0% (worse)	568.455 ms	+37.9% (worse)	1.278 ms	-17.7% (better)
Linux	fmtprintf	5828200 B	+163.3% (worse)	6.229 s	+61.6% (worse)	2.450 ms	-25.5% (better)
Linux	println	691544 B	+851.8% (worse)	558.065 ms	+39.5% (worse)	1.594 ms	-4.7% (better)
macOS	cprintf	50784 B	-40.0% (better)	467.011 ms	+12.7% (worse)	3.518 ms	-1.9% (better)
macOS	fmtprintf	2749760 B	+17.2% (worse)	6.280 s	+59.4% (worse)	19.575 ms	-16.2% (better)
macOS	println	111264 B	-12.4% (better)	449.987 ms	+8.5% (worse)	3.968 ms	-45.3% (better)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs main
Linux	BenchmarkLookupPCRandom	13.410 ns/op	-0.5% (better)
Linux	BenchmarkMergeCompilerFlags	150.500 ns/op	-2.6% (better)
Linux	BenchmarkMergeLinkerFlags	94.980 ns/op	-6.2% (better)
Linux	BenchmarkChannelBuffered	33.400 ns/op	-3.9% (better)
Linux	BenchmarkChannelHandoff	25351 ns/op	-20.9% (better)
Linux	BenchmarkDefer	45.280 ns/op	-19.6% (better)
Linux	BenchmarkDirectCall	1.557 ns/op	-0.4% (better)
Linux	BenchmarkGlobalRead	1.873 ns/op	+20.1% (worse)
Linux	BenchmarkGlobalWrite	2.490 ns/op	-0.2% (better)
Linux	BenchmarkGoroutine	34817 ns/op	-53.9% (better)
Linux	BenchmarkInterfaceCall	9.386 ns/op	+20.4% (worse)
Linux	BenchmarkRuntimeGetG	2.181 ns/op	-58.9% (better)
macOS	BenchmarkLookupPCRandom	12.370 ns/op	-0.6% (better)
macOS	BenchmarkMergeCompilerFlags	130.200 ns/op	+4.6% (worse)
macOS	BenchmarkMergeLinkerFlags	86.940 ns/op	+25.4% (worse)
macOS	BenchmarkChannelBuffered	22.020 ns/op	-20.6% (better)
macOS	BenchmarkChannelHandoff	6915 ns/op	-1.3% (better)
macOS	BenchmarkDefer	26.220 ns/op	-29.9% (better)
macOS	BenchmarkDirectCall	1.017 ns/op	-30.0% (better)
macOS	BenchmarkGlobalRead	1.020 ns/op	-15.8% (better)
macOS	BenchmarkGlobalWrite	1.016 ns/op	-31.7% (better)
macOS	BenchmarkGoroutine	25140 ns/op	-20.6% (better)
macOS	BenchmarkInterfaceCall	5.111 ns/op	-5.3% (better)
macOS	BenchmarkRuntimeGetG	2.028 ns/op	-21.3% (better)

Compared only with the latest matching platform in the main series.

[cpunion]

Benchmark context: the large Linux whole-file deltas are the expected debugger-owned DWARF cost of xgo-dev#2164 Go-compatible defaults (native = embedded), not an equivalent loadable-footprint increase. Against the latest main baseline, Linux text/data are cprintf 410/13,813 vs 410/12,949 bytes; println 21,929/30,789 vs 21,337/29,283 bytes; and fmtprintf 768,473/997,255 vs 759,177/982,167 bytes. The proposal artifact-role gate separately measures debug and deployment outputs, and the Targets lane verifies that host DWARF leaves flashed bytes identical. Production stripping remains explicit through -debug-artifact=none or -w; full-DWARF build-time cost remains tracked.

[cpunion]

Compatibility shard follow-up:

• The first run's only failure was TestRuntimeSetFinalizerTinyObjects observing 1/32 finalizers within its 3-second window.
• The current head does not change that test or the finalizer/GC implementation.
• A Linux/x86_64 Go 1.24.2 reproduction capped at 6 GiB / 4 CPUs passed 10/10 repetitions of the focused test.
• The failed job was rerun without code changes. Attempt 2 passed; github.com/goplus/llgo/test/go and the full 1075-second shard both completed successfully.

This is therefore treated as a timing flake. No debugger acceptance assertion or runtime behavior was weakened.