Render-pipeline optimization — Std.Ui renderNodeAs append-on-empty

Cuts per-element allocations in the Sky-side Element→Html pass (sky-stdlib/Std/Ui.sky), the largest remaining render allocation share (73% of render objects per the HEAD re-baseline in render-opt-htmltovnode-20260818/, after typed twins and the HtmlToVNode fast path landed). Output is byte-identical — an allocation/CPU change, not a behaviour change.

What the profile said (HEAD, before)

renderNodeAs traverses each element's attribute list and builds its attribute + child lists with ++. List.append (rt.List_appendT) always makes a fresh slice and copies both operands — even when one operand is []. The common element carries no extra attrs, no nearby children, and no pseudo/transition/animation attrs, so several of those ++ are appends onto [] that copy the other operand for nothing. In the HEAD flat profile (render-opt-htmltovnode-20260818/after-alloc-profile-showcase.txt), rt.List_appendT is a top-12 site at 3.67% of all render objects and rt.List_cons a further 2.92%.

Per common element, renderNodeAs paid:

The change (sky-stdlib/Std/Ui.sky)

Four guards, each preserving the emitted bytes exactly:

• allAttrs — if List.isEmpty extraAttrs then attrs else extraAttrs ++ attrs. [] ++ attrs == attrs.
• attrList — when pseudoAttrs/transitionAttrs/animationAttrs are all empty, the tail is htmlAttrs directly. [] ++ [] ++ [] ++ htmlAttrs == htmlAttrs.
• renderedChildren — if List.isEmpty renderedNearby then renderedMain else renderedMain ++ renderedNearby. xs ++ [] == xs.
• collectTransitions — case rules of [] -> ("", False); _ -> (String.join …, List.any …). renderNodeAs matches ("", _) and discards the Bool when the string is empty, so the empty branch is byte-identical in the emitted HTML.

The 2nd markerFlags propagatedAttrs was left in place: markerFlagStep lowers to a value-struct accumulator (_u := v_1; _u.Row = true), so it heap- allocates ~nothing, and removing it would be a CPU-only change against a deliberately-cautious comment.

Attribute order in attrList is not output-sensitive — Std.Html's renderer sorts an element's attributes alphabetically before emit (verified: a reordered-attrList mutant produced byte-identical HTML). The guards therefore preserve order trivially; the falsification below uses a content-dropping mutant instead.

The win (allocation, grounded; CPU secondary with spread)

Forum-shaped Element tree (Main_forumView, ~16 Std.Ui elements/post) at the re-baseline's view sizes, rendered through the full Std.Ui pass (Ui.layout [] el |> Html.render). BenchmarkRenderForum (render_bench_test.go), 3 repeats, one binary each for before/after (compiler re-embeds the stdlib); allocation reproduces to the object across all repeats.

The saving is a per-element fixed cost, so it scales with element count (−352 at 94 el, −3,484 at 974 el ≈ 3.6 allocs/element). CPU on this shared M1 moved −3.0% to −3.5% across these runs; treat it as secondary against the ±37% interaction-CPU swing the re-baseline documents — the allocation figure is the stable one.

Relating to the HEAD interaction

This is a −10% reduction of the Element→Html pass at both re-baseline view sizes. The re-baseline established that pass as 73% of render objects, so this is on the order of ~7% of total render objects per interaction — an estimate, because the harness isolates Element→Html (its Html.render serialisation stands in for the downstream HtmlToVNode, which this change does not touch). The absolute per-element saving (3.6 allocs/el) is the measured, app-independent quantity; it lands on every element on every first paint and every full-replace patch.

Byte-identical correctness gate

• Differential build diff (the primary gate). A corpus harness (harness-Main.sky, 11 elements exercising empty/one/many-style, html attrs, containers, fill-propagation, nearby, pseudo, transition, animation, a kitchen-sink combining all) rendered through Ui.layout [] el |> Html.render, built with the pre-change compiler and the post-change compiler, diffed: byte-identical, all 11 cases.
• Mutation-proven the gate bites. Injecting a dropped-attribute bug (attrList empty-extras branch → [] instead of htmlAttrs) turned the diff red on the html-attrs and container cases (the class attribute vanished). A reordering mutant was correctly not flagged — the renderer sorts attributes — confirming the gate keys on content, not incidental order.
• Committed regression. tests/Std/UiRenderConcatTest.sky (5 cases) pins the exact HTML of all four guarded code paths; discovered by scripts/sky-suites.sh and counted in SKY_SUITES_EXPECTED (393 → 398).
• Existing Std.Ui HTML suites (UiTransitionAnimationTest, UiPseudoClassTest, UiInputCheckboxTest, UiAspectGridTest, UiMediaQueryTest — 81 assertions) all green, including collectTransitions joins multiple AttrTransitions with ', '.
• coerce-floor unchanged — 00-standard-libs 186, 19-skyforum 256, 26-ui-showcase 365 (exact, no widening); adapter exact at 17.
• go test ./rt/... green; both examples build clean-slate.

Conditions