Performance

Zero-Allocation Hot Paths

The classification pipeline is designed for zero allocations on the hot path.

Hardware: AMD Ryzen 9 7950X, Go 1.26.4

Operation	Time	Allocs
Classify (built-in Error)	~9 ns	0
Classify (plain error)	~30 ns	0
IsRetryable	~9 ns	0
ExitCode	~9 ns	0
WithContext	~8 ns	0
ParseFamily	~12 ns	0
HandleError (with context)	~450 ns	5
Runner.Run (1 rule)	~420 ns	5

Interpretation

• Classify on a built-in *Error is ~9ns because it reads the family field directly (step 2: Classified interface).
• Classify on a plain error is ~30ns because it walks the sentinel map (step 4). At 50 registered sentinels, this is ~285ns/0 allocs.
• HandleError includes template resolution and stderr write. Use HandleErrorDetailed when you only need the structured result.

Registry Performance

The Registry uses atomic.Pointer[sentinelMap] for lock-free reads:

• Reads (the Classify hot path) load the pointer once and iterate lock-free and allocation-free
• Writes serialize under the write lock and publish a new snapshot via copy-on-write
• At 50 registered sentinels: ~285 ns/0 allocs (was ~1330 ns/3 allocs/1.8KB under the old RLock+copy approach)

Reproducing

GOEXPERIMENT=jsonv2 go test -bench=. -benchmem -count=1 ./...