# 0009. Renderer-freeze exception for canvas allocation and indicator loading

> The 2026-10-10 runtime audit found two costs inside the frozen renderer area.

Source: https://kcq.nebutra.com/docs/architecture/adr/0009-renderer-freeze-perf-exception


* Status: Accepted
* Date: 2026-10-10
* Deciders: tseka (founder-authorized)
* Amends: ADR 0005, ADR 0008

Context [#context]

The 2026-10-10 runtime audit found two costs inside the frozen renderer area.

1. **Canvas memory.** Each chart allocated every canvas layer up front:

   * a drawing layer on every pane;
   * a left-axis pair even when no left axis is shown, sized to the full plot when the host had no left-axis layer;
   * a full-plot crosshair canvas.

   The landing hero held 9 canvases (38.6 MiB at DPR 2).
2. **Indicator loading.** `createChartController` loaded every indicator module before mounting. The workbench made 81 JS requests on first load.

Decision [#decision]

The founder authorizes changes to the frozen paths (`engine/**`, `controllers/**`), limited to two areas.

**Canvas allocation and sizing.** This covers:

* when a pane, axis, drawing or crosshair surface is created or released;
* the size of its backing store, which is the drawn area times DPR;
* where the surface sits in the DOM, without changing stacking order.

**Indicator loading.** This covers:

* the build-time static catalog;
* on-demand loading of indicator modules;
* adding newly loaded definitions to the calculation runtimes;
* the async boundaries for adding an indicator, restoring a layout and running Agent tools.

The layout-autosave scheduling change (coalescing viewport-driven checks) is also covered, because it lives in the same controllers path.

These stay frozen:

* drawing algorithms (`engine/renderers/**` paint code);
* `engine/scale/**`;
* `foundation/tokens/fonts.ts`;
* any change to what a frame draws.

Before and after builds must produce pixel-identical output.

Consequences [#consequences]

* PR 363045841/KLineChartQuant#309 ships the change.
* Screenshots of the hero and workbench harnesses are byte-identical before and after.
* Synchronous indicator methods accept only loaded definitions. Hosts call `loadIndicators` first, or opt into `indicatorLoading: 'all'`.
* Any other renderer change still needs its own ADR.