Cache elapsedDays in a fixed-size atomic.Int32 array

[mjradwin]

Summary

Reintroduces the elapsedDays memoization that was removed in 4750e9d, but backed by a fixed-size atomic.Int32 array indexed directly by year rather than a mutex-protected map. Each slot is written at most once with a deterministic value, so atomic load/store is sufficient — and on x86/arm it lowers to the same MOV instructions as a plain load.

The cache covers Hebrew years 5000–6999 (~1240 CE to 3239 CE), which spans every realistic use of the library. Years outside the window fall through to recompute, so nothing breaks for callers operating on extreme dates. The cache cannot grow and contributes ~8 KB of static memory.

Why this shape

The previous mutex-protected map cache was a net slowdown on every realistic workload (a sync.RWMutex round-trip costs more than elapsedDays0 itself), which is what motivated 4750e9d. A year-indexed array avoids the lock entirely; using atomic.Int32 rather than a plain int32 slot keeps it race-free per the Go memory model at zero runtime cost on the platforms we run on.

Microbenchmarks (ns per elapsedDays call, Go 1.24, x86-64):

workload	nocache (today)	atomic-array (this PR)
hot (single year)	15	2
warm (~10 years)	19	3
cold (random in-range)	43	3
out-of-range year	15	15

That's a 5–15× speedup on every in-range workload; one ToRD call invokes elapsedDays ~5 times for the same year (via DaysInYear → LongCheshvan/ShortKislev), so the win compounds.

Test plan

• go test ./... — all existing tests pass, including coverage of out-of-range years (1, 2, 123, 1234, 3671, 3762) which exercise the fall-through path
• go test -race ./... — race detector clean
• go vet ./... clean
• staticcheck -checks=all ./... clean
• gofmt -l . clean
• Public API surface unchanged (go doc -all byte-identical to main)

---

Generated by Claude Code