[Based on #138] debug: integrate DWARF-safe defaults

[cpunion]

Integrates the source-debugging dependency chain required by xgo-dev#2164 on top of #138 without pulling scheduler implementation into the proposal.

• uses the reviewed independent DebugRef return/select repairs instead of the older duplicate commits in debug: keep optimization and SSA rewrites DWARF-safe xgo-dev/llgo#2143
• keeps the normal LLVM optimization pipeline enabled with DWARF and preserves defer locations and C ABI variable homes
• keeps aggregate parameter homes authoritative in debug builds, avoiding invalid LLDB byval dereferences while retaining the optimized rewrite for non-debug builds
• restores Go-compatible linked DWARF defaults while preserving typed embedded, host, external, and none artifact policy
• preserves Darwin PC-line sites while suppressing only the address-site class that disturbs LLDB lexical scopes
• stabilizes Darwin cold/warm debug-map archive paths
• resolves outstanding review notes by guarding a moving anchor, clarifying site policy, using explicit PC-line naming, and updating the design documentation

Upstream review units remain xgo-dev#2183, xgo-dev#2184, xgo-dev#2143, xgo-dev#2148, xgo-dev#2157, xgo-dev#2142, and xgo-dev#2202. This PR is the latest-main combination and acceptance layer.

Validation:

• full internal/build package: pass in 249.939s
• full internal/cabi package: pass, including the aggregate debug-home regression
• focused ssa, cl, and cmd/internal/debug: pass
• native LLDB acceptance: 276/276, including slices and mutated aggregate parameters
• DWARF return-order, Darwin pclntab line-site, and large-array C ABI probes: pass
• real GOOS=js embedded and external browser artifacts build and load
• exact pre-Asyncify objects pass llvm-dwarfdump verification across all 71 compile units

Binaryen's released Asyncify still exhibits WebAssembly/binaryen#6406 on this layer's final browser artifact. The stacked #140 closes that final-artifact gate with the repair proposed upstream as WebAssembly/binaryen#8964 and keeps strict llvm-dwarfdump --verify checks enabled; this PR does not weaken validation or change runtime semantics.

[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

5336d7074219 | 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)	556.875 ms	+35.1% (worse)	1.279 ms	-17.6% (better)
Linux	fmtprintf	5789416 B	+161.5% (worse)	6.605 s	+71.3% (worse)	2.656 ms	-19.2% (better)
Linux	println	687280 B	+845.9% (worse)	564.944 ms	+41.2% (worse)	1.618 ms	-3.3% (better)
macOS	cprintf	50784 B	-40.0% (better)	555.806 ms	+34.1% (worse)	3.477 ms	-3.1% (better)
macOS	fmtprintf	2745456 B	+17.0% (worse)	7.416 s	+88.2% (worse)	22.982 ms	-1.7% (better)
macOS	println	110336 B	-13.1% (better)	557.760 ms	+34.5% (worse)	5.246 ms	-27.7% (better)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs main
Linux	BenchmarkLookupPCRandom	13.420 ns/op	-0.4% (better)
Linux	BenchmarkMergeCompilerFlags	154.300 ns/op	-0.1% (better)
Linux	BenchmarkMergeLinkerFlags	96.170 ns/op	-5.1% (better)
Linux	BenchmarkChannelBuffered	33.380 ns/op	-4.0% (better)
Linux	BenchmarkChannelHandoff	28894 ns/op	-9.8% (better)
Linux	BenchmarkDefer	45.820 ns/op	-18.6% (better)
Linux	BenchmarkDirectCall	1.557 ns/op	-0.4% (better)
Linux	BenchmarkGlobalRead	1.869 ns/op	+19.8% (worse)
Linux	BenchmarkGlobalWrite	2.487 ns/op	-0.3% (better)
Linux	BenchmarkGoroutine	61095 ns/op	-19.2% (better)
Linux	BenchmarkInterfaceCall	9.349 ns/op	+20.0% (worse)
Linux	BenchmarkRuntimeGetG	2.180 ns/op	-58.9% (better)
macOS	BenchmarkLookupPCRandom	14.760 ns/op	+18.6% (worse)
macOS	BenchmarkMergeCompilerFlags	181.200 ns/op	+45.5% (worse)
macOS	BenchmarkMergeLinkerFlags	115.500 ns/op	+66.5% (worse)
macOS	BenchmarkChannelBuffered	29.160 ns/op	+5.2% (worse)
macOS	BenchmarkChannelHandoff	6765 ns/op	-3.5% (better)
macOS	BenchmarkDefer	41.380 ns/op	+10.6% (worse)
macOS	BenchmarkDirectCall	1.330 ns/op	-8.5% (better)
macOS	BenchmarkGlobalRead	1.184 ns/op	-2.2% (better)
macOS	BenchmarkGlobalWrite	1.266 ns/op	-14.9% (better)
macOS	BenchmarkGoroutine	45185 ns/op	+42.7% (worse)
macOS	BenchmarkInterfaceCall	6.860 ns/op	+27.1% (worse)
macOS	BenchmarkRuntimeGetG	2.897 ns/op	+12.4% (worse)

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

[cpunion]

按当前仓库范围和依赖整理策略，关闭 cpunion/llgo 内部 consolidated PR，避免后续 push 产生重复 CI。上游 xgo-dev#2143 及其独立依赖单元保留用于 review；本 PR 的组合/验收内容不作为新的上游合并单元。后续如需修改验证，基于现有分支重新打开或创建对应 PR。