cl: stabilize generic local ABI names across cache loads

[cpunion]

Summary

• keep synthesized generic-local TypeName objects positionless because their generated names already encode the complete local and generic identity
• prevent ABI naming from appending a loader-relative .p<token.Pos> suffix that differs between cached archives and a later synthetic root module
• add a cross-process cold/hot package-cache regression and assert the frontend position invariant

This preserves the existing position fallback for genuine detached local-scope objects while making generic-local cache hits link reliably.

Fixes #2297

Testing

• go test ./cl -run "^TestGenericLocalTypePatchHelpers$" -count=1
• go test ./cl -run "^TestRunAndTestFromTestgo$" -count=1
• go test -tags=dev ./internal/build -run "^TestGenericLocalTypeColdAndHotPackageCache$" -count=1
• go test ./ssa/abi -count=1
• issue reproducer: forced cold llgo test -a -p=1 ... followed by the immediate cache-hit llgo test -p=1 ...; both pass

[fennoai]

Review summary

The fix is correct and well-targeted. Replacing t.Obj().Pos() with token.NoPos for the detached TypeName in patchLocalGenericNamed (cl/compile.go) removes the source of cross-process non-determinism: scopeIndices in ssa/abi/abi.go only appends the loader-relative .p<pos> suffix when pos.IsValid(), so token.NoPos drops that varying suffix while type identity is fully preserved by the type-arg-encoded localNamedName. The explanatory comment is accurate, and the subprocess-based cold/hot cache test is the right shape — an in-process test could not reproduce #2297 since the offending token.Pos comes from a per-process FileSet.

Security, documentation-accuracy, and performance passes surfaced no issues (token.NoPos is in fact a net reduction in per-type work). One minor, optional test-robustness finding is left inline.

No blocking issues.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

LLGo baseline benchmarks

af9eb545153f | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	19288 B	0 B / +0.0%	334.317 ms	-2.748 ms / -0.8% (better)	1.361 ms	-19.95 us / -1.4% (better)
Linux	fmtprintf	1859320 B	0 B / +0.0%	2.551 s	-37.23 ms / -1.4% (better)	3.755 ms	+191.8 us / +5.4% (worse)
Linux	println	68776 B	0 B / +0.0%	327.389 ms	-7.706 ms / -2.3% (better)	1.649 ms	-72.58 us / -4.2% (better)
macOS	cprintf	84672 B	0 B / +0.0%	396.257 ms	+23.68 ms / +6.4% (worse)	3.110 ms	+568 us / +22.3% (worse)
macOS	fmtprintf	1892448 B	0 B / +0.0%	3.177 s	+950.7 ms / +42.7% (worse)	14.666 ms	+2.207 ms / +17.7% (worse)
macOS	println	121360 B	0 B / +0.0%	526.840 ms	+180.8 ms / +52.3% (worse)	4.700 ms	+867.3 us / +22.6% (worse)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	12.290 ns/op	-0.11 ns/op / -0.9% (better)
Linux	BenchmarkMergeCompilerFlags	146.700 ns/op	+3.2 ns/op / +2.2% (worse)
Linux	BenchmarkMergeLinkerFlags	96.120 ns/op	+2.6 ns/op / +2.8% (worse)
Linux	BenchmarkChannelBuffered	36.300 ns/op	+0.03 ns/op / +0.1% (worse)
Linux	BenchmarkChannelHandoff	24537 ns/op	-954 ns/op / -3.7% (better)
Linux	BenchmarkDefer	44.890 ns/op	+0.1 ns/op / +0.2% (worse)
Linux	BenchmarkDirectCall	1.759 ns/op	+0.001 ns/op / +0.1% (worse)
Linux	BenchmarkGlobalRead	1.760 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGlobalWrite	2.811 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGoroutine	30267 ns/op	-459 ns/op / -1.5% (better)
Linux	BenchmarkInterfaceCall	9.045 ns/op	-0.098 ns/op / -1.1% (better)
Linux	BenchmarkRuntimeGetG	2.111 ns/op	-0.001 ns/op / -0.04735% (better)
macOS	BenchmarkLookupPCRandom	13.260 ns/op	+1.62 ns/op / +13.9% (worse)
macOS	BenchmarkMergeCompilerFlags	187.600 ns/op	+81.5 ns/op / +76.8% (worse)
macOS	BenchmarkMergeLinkerFlags	103.400 ns/op	+35.39 ns/op / +52.0% (worse)
macOS	BenchmarkChannelBuffered	23.980 ns/op	+2.49 ns/op / +11.6% (worse)
macOS	BenchmarkChannelHandoff	7631 ns/op	-1424 ns/op / -15.7% (better)
macOS	BenchmarkDefer	33.380 ns/op	+1.27 ns/op / +4.0% (worse)
macOS	BenchmarkDirectCall	1.265 ns/op	+0.199 ns/op / +18.7% (worse)
macOS	BenchmarkGlobalRead	1.257 ns/op	+0.2673 ns/op / +27.0% (worse)
macOS	BenchmarkGlobalWrite	1.517 ns/op	+0.5193 ns/op / +52.0% (worse)
macOS	BenchmarkGoroutine	53491 ns/op	+27998 ns/op / +109.8% (worse)
macOS	BenchmarkInterfaceCall	5.589 ns/op	+1.162 ns/op / +26.2% (worse)
macOS	BenchmarkRuntimeGetG	2.960 ns/op	+0.576 ns/op / +24.2% (worse)

Compared with cf91afb98fca measured in the same runner job.