Fix nocopy values decoded from streams

[kelindar]

Summary

• allocate stable slices when decoding from generic streams
• add a regression test for adjacent short nocopy.String fields

Root cause

streamReader.Slice reused one 10-byte scratch buffer. nocopy.String and other retained nocopy codecs alias the returned slice, so decoding the next short value overwrote earlier values. Generic streams have no reusable backing buffer, so each retained slice must own its bytes. Slice-backed Unmarshal remains zero-copy.

Reproduction

Before the fix, decoding {First: "first", Second: "second"} through NewDecoder(bytes.NewReader(...)) produced {First: "secon", Second: "second"}.

Benchmarks

Apple M2, Go 1.25, 10 runs for the targeted benchmark. The pre-fix result is fast because it aliases mutable scratch memory and is incorrect.

benchmark	master	fix	delta
two short nocopy.String values via stream	39.42 ns/op, 0 B, 0 allocs	53.32 ns/op, 16 B, 2 allocs	+35.3%, +2 required allocations
repository binary/stream-dec	132.6 ns/op, 1 alloc	138.2 ns/op, 2 allocs	+4.2%, +1 allocation
slice-backed nocopy/str-dec	32.0 ns/op, 0 allocs	31.9 ns/op, 0 allocs	unchanged

Full repository benchmark suite: cd bench && go run . (100 samples, 10 ms/sample).

Checks

• go test ./...
• go vet ./...
• git diff --check

[github-actions]

████████████████████ 100.0% on changed lines

1 lines changed   0 uncovered   1 file

 📂 1 file changed

	File	Coverage	Uncovered Lines
🟢	reader.go	100.0%	—

 📋 Full diff-cover report

Diff Coverage

Diff: origin/master...HEAD, staged and unstaged changes

• reader.go (100%)

Summary

• Total: 1 line
• Missing: 0 lines
• Coverage: 100%

_{🛡️ diff-cover-action}