Typed list destructuring — rt.SkyLenT / SkyElemT / SkyTailSliceT

2026-08-15. Follow-on to ../hof-dispatch-20260815/. That run removed the reflect.MakeFunc adapter from HOF callbacks and left 126 allocations per Std.Ui marker scan standing. This is where a large share of them were.

What was wrong

The Rust codegen's cons/list pattern binder (Pattern::Cons, Pattern::List in rust/crates/lower/src/lower.rs) emitted rt.SkyLen / rt.SkyElem / rt.SkyTailSlice. All three take x any and route through rt.AsList.

AsList fast-paths a []any. A []T for any other T misses that assertion and falls to the reflect arm, which allocates a fresh []any of length n and boxes every element into it (runtime-go/rt/rt.go, AsList).

So on a typed list each of those calls is O(n) allocation, not the single boxed slice header the signature suggests — and a cons loop rebuilt the entire list on every iteration, making an O(n) walk quadratic.

Measured directly (runtime-go/rt/sky_destructure_typed_test.go), on a 16-element []struct{string; int}:

18 = n + 2, i.e. the whole list re-boxed, per call.

The fix

When the subject's Go type is already a slice, emit the typed helpers. They compile to len(xs), xs[i] and xs[1:]; Go infers the type argument from the slice, so no call site carries one. The ABI is unchanged — no monomorphisation, one emit per definition, as with the eta-expansion.

The decision keys on subj.ty (the Go type of the expression being emitted), not on the subj_ty type-context threaded alongside it. Those can disagree, and it is subj.ty that decides whether the emitted Go type-checks: an over-conservative fall back to the any path is a missed optimisation, while the reverse is a go build failure.

Bounds guards are kept, though the pattern's length test has already run. That test is a claim about the caller; a helper that panics when the claim is wrong would convert a lowering bug into a runtime panic, and "no runtime panic from well-typed Sky" is not a property to rest on an invariant asserted in a comment. Out of range yields T's zero, mirroring the any versions' nil.

Measurement

Same harness, same app, same method as ../hof-dispatch-20260815/: examples/26-ui-showcase (384 elements), GOMAXPROCS=1, closed loop, 25 sessions, 20 s, warmup 5 s, ramp 3 s. Arms alternate per rep so thermal drift and burst-credit decay fall on both equally. The two binaries differ only by this change — verified by strings: SkyElemT appears 11× in the after binary and 0× in the before.

bench.sh is the script as run. The "before" arm here is the after arm of the HOF-dispatch run, i.e. eta-expansion already landed.

1.34× throughput. Ranges do not overlap. p50 latency 129.3 ms → 89.8 ms (−31%), p95 326.6 → 273.7 ms.

Both arms: error_rate: 0, 25/25 sessions established, generator CPU under 0.3% of the machine (generator_possibly_saturated: false), valid: true — so the faster arm is not winning by shedding work. Raw JSON for all six runs is committed beside this file.

Cumulative over both changes, against the original 135.47 interactions/sec baseline in ../hof-dispatch-20260815/: 1.80×.

Secondary effect, mechanically measured by xtask coerce-floor: because SkyElemT returns T and SkyTailSliceT returns []T, the narrowings that used to bridge their any / []any results disappear. The narrow class fell 9,986 → 9,479 (−507) across 26 projects, with the adapter class unchanged at 35 — which is the classified ratchet doing its job: it reported a tightening to be blessed rather than a widening to be investigated.

What these numbers do NOT say

• The per-call allocation ratio does not extrapolate to a request. 18 → 0 on one helper is not 1.34× because a large share of an interaction is network syscall and rendering no compiler change touches. The end-to-end number is the one to quote.
• The gain scales with list length and with how much destructuring an app does. 26-ui-showcase walks attribute lists constantly. An app that pattern- matches lists rarely will see less; this is not a uniform 1.34× for all Sky.
• Closed loop is the sensitive regime, not the typical one. -think 0 keeps the server saturated; with realistic think time the difference compresses.
• One host, one toolchain. Macmini, gomaxprocs: 8 reported but the app pinned to GOMAXPROCS=1. Not a cross-platform claim.
• []any subjects gain nothing measurable. They already hit AsList's fast path; the win is specifically on typed slices.