Render-pipeline optimization — Html→VNode reflect de-boxing

Removes the single largest allocation site in the Sky.Live render pipeline: asList reflect-boxing every attribute and child of every element inside HtmlToVNode (runtime-go/rt/live.go). Output is byte-identical — a runtime allocation/CPU change, not a behaviour change.

What the profile said (HEAD, before)

Allocation attribution of the full first-paint render of 26-ui-showcase (384 sky-id elements), Main_view → rt.HtmlRenderWithHandlers, -memprofile

• pprof -alloc_objects (before-alloc-profile-showcase.txt). Allocation is the stable observable (reproduces to 0.0–0.2%); CPU self-time is quoted with its spread.

Flat allocation sites, sorted — the #1 site was reflect.unsafe_New at 23.6% (11.06 M objects), and asList accounted for 22.24% cum of all objects:

11061458 23.62%  reflect.unsafe_New
 2142093  4.58%  rt.asList                         (make([]any, n))
 ...
        cum 22.24%  rt.asList  →  8.27M of the reflect.unsafe_New via rv.Index(i).Interface()

Root cause. Std_Html_Attributes_Attribute and Std_Html_Html are type aliases for rt.SkyADT, so an HElement's attribute and child lists arrive as []SkyADT boxed in Fields []any. The old HtmlToVNode routed both through asList, whose typed-slice branch does reflect.ValueOf(v) + rv.Index(i).Interface() per element — one reflect.unsafe_New heap box per attribute and per child, plus a throwaway []any per list.

The change

runtime-go/rt/live.go. HtmlToVNode gains a []SkyADT fast path (appendHtmlAttrs / appendHtmlChildren) that iterates the typed slice directly: no []any copy, no per-element reflect box, and Children is pre-sized (each child yields exactly one VNode). The []any and reflect branches are retained unchanged for the erased/mixed case (e.g. a Raw/VNode passthrough child). Shared shape functions (htmlShapeToVNode, applyHtmlAttrShape) keep the fast and slow paths behaviourally identical.

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

Html→VNode pass, ref-vs-new, one binary, at the re-baseline's view sizes

BenchmarkHtmlToVNode_{Before,After} in runtime-go/rt/html_to_vnode_diff_test.go runs the frozen pre-change converter (htmlToVNodeRef) and the new HtmlToVNode over a forum-shaped Html ADT (~16 elements/post). 3 repeats, allocation identical to the object across all (pass-ref-vs-new.txt):

CPU on this shared M1 moved −40% to −66% across runs (±37% swing — do not quote a point value); the allocation figure is the stable one.

Full first-paint interaction, same-binary controlled (26-ui-showcase, 384 el)

Rebuilt the app with only runtime-go/rt/live.go changed (compiler re-embeds the runtime), same benchmark:

Bytes 1,366,970 → 1,150,632 (−15.8%). The full-pipeline reduction equals the Html→VNode reduction exactly — the Sky-side pass is untouched. The after profile (after-alloc-profile-showcase.txt) confirms asList is gone from the render path and reflect.unsafe_New halved (11.06M → 6.27M); the residue is reflect.Value.call HOF dispatch inside the Sky-side pass, a different origin (the next lever, below).

The share of the interaction the win removes tracks how attribute-heavy the view is: 26-ui-showcase (charts, many styled elements) carried 22.5% of its render allocations in Html→VNode; the lighter 19-skyforum signed-in view is Sky-Element→Html-dominated (94% of its 7,815 render allocs/op), so the same pass is a smaller slice there. The pass itself is −64% everywhere it runs, on every first paint and full-replace patch.

Byte-identical correctness gate

• runtime-go/rt/html_to_vnode_diff_test.go — TestHtmlToVNodeDiff renders a corpus (every Html/Attribute ADT shape, []SkyADT and []any slices, a VNode passthrough child, class/style multi-value joining, URL-scheme neutralisation, empty/nested) through the frozen reference and the live converter and asserts the VNode tree is DeepEqual and the rendered HTML is byte-identical. TestHtmlToVNodeDiffGateIsFalsifiable proves the gate bites. Mutation-proven: injecting a dropped-child bug into the live appendHtmlChildren turned TestHtmlToVNodeDiff red on 4 corpus cases.
• App-level oracle: the full first-paint HTML of 26-ui-showcase (77,258 B) and 19-skyforum logged-out (18,099 B) + signed-in (18,354 B), captured before the change and diffed after the rebuild — byte-identical, all three.
• Full go test ./rt/ (incl. the render goldens, live_alloc_gate, live_render_deterministic) green; go test -race on the render/live subset green. No emit change → coerce-floor unaffected.

Next lever (identified, not taken here)

The Sky Element→Html pass (Main_view, 73% of render allocations) is the remaining prize, and its flat sites are now the profile top: the collectStyle accumulator-concat fold inside buildStyleStringWith (.7.17, 6.99% flat), String_fromInt, toAttrAttribute.func1 boxing, and List_cons. Closing those means editing sky-stdlib/Std/Ui.sky (compiled Sky) under the same byte-identical constraint — a higher-risk change than this runtime-Go one.

Conditions