v5.10.0

[5.10.0] - 2026-08-08

Hot-path performance (step 2 of the code audit)

Pure optimisation — no new features, no config changes, no visible behaviour change. Measured against v5.9.14 on a 3-room config (6 overlays + 5 zones + 8 icons + 8 labels + 4 gauges per room): 300× _update() went from 234 ms / 2 400 querySelector / 300 offsetWidth to 148 ms / 0 / 0 (−37 % time), and a 60-frame resize drag from 32 ms / 480 querySelector / 120 offsetWidth to 1.3 ms / 0 / 60 (−96 % time). jsdom timings aren't browser timings — the percentages are the signal, the DOM-op counts are exact.

_update() no longer touches the DOM to find nodes. Each icon's inner <ha-icon> was re-queried every tick — 20 icons = 20 querySelector calls per state change, for a node that cannot change between renders. Now cached at render time in _icoIconEls, the same pattern badges already used for _biconEls/_blabelEls.

Forced layout is now conditional. _update() read this.offsetWidth unconditionally to resolve %-based icon sizes and label font sizes, even for configs that use none. A per-render _needsCardWidth flag records whether any item actually sizes in % — including sizes hidden in per-profile overrides (portrait: {size: "12%"}) and legacy tier keys. Detection is deliberately conservative (a non-string value counts as "yes"): a false positive costs one reflow, a false negative would break sizing.

Geometry events no longer run a state pass. The ResizeObserver called the full _update() — every condition on every overlay, zone, badge, icon, label and gauge re-evaluated — once per frame during a window drag or an opening keyboard. Only three things in _update() depend on the box, so resize now runs a narrow _applyResizeStyles(). The one entangled case (day/night blind gauges, whose slat gradient derives from both their own height and the state value) still delegates to the full pass, gated on _hasDayNightGauge.

Layout passes coalesce. Several observers can fire in one frame, each previously running its own fit+stage sequence. All now funnel through _requestLayout(), collapsing a burst into one microtask pass — same pattern as the existing _requestPin(). Stage-only requests (.wrap resized but not the card) upgrade to a full fit if a full request lands in the same microtask, mirroring _schedule()'s nav-only/full upgrade.

Layout element refs cached. _layoutFitWrap(), _layoutStage() and _layoutRootHeight() each re-queried .wrap, .content and ha-card per call — 11 querySelector calls per observer callback. Now behind _elWrap()/_elContent()/_elCard(), which re-resolve only when the cached node has left the current shadow tree (getRootNode()!==this.shadowRoot — exactly what innerHTML replacement does to it), so no manual invalidation is needed.

New instance fields are declared in the constructor, restoring the "constructor documents the instance shape" contract the audit flagged.

Testing

tests/lifecycle.test.js → 35 assertions. The hot-path budget is now a hard ceiling: _update() must do zero querySelector calls and zero offsetWidth reads on a 40-item card, so an uncached query or forced read reintroduced in a per-item loop fails CI. The riskiest failure mode of this change — a % size silently ceasing to resolve — is covered behaviourally rather than structurally: a 10% icon on a 400 px card must compute to 40px and re-compute to 80px at 800 px. Coalescing (50 requests → 1 pass) and the absence of _update() from the resize path are asserted directly. Render (178) and smoke tiers unchanged and passing; render output diffed line-by-line against v5.9.14 and identical.