RenderFitGridSection class

Paints a rectangular block of grid cells as text, with real widgets layered over the cells that need them.

The point of the exercise: a 40 x 8 block of text is 320 widgets, elements and render objects in a conventional table, each with its own layout and paint. Here it is one RenderBox, a handful of cached TextPainters, and a single batched line draw for the grid rules. Widgets are spent only where they buy something.

Four things keep that from degrading:

• Windowing. Only rows in [firstVisibleRow, firstVisibleRow + visibleRowCount) and columns intersecting the horizontal viewport are touched, so cost tracks the viewport rather than the dataset.
• Painter caching. A cell's TextPainter survives across paints and is re-laid-out only when its spec or its column width changes. The cache is pruned to the window, so it stays bounded by what is on screen.
• Batched rules. Every divider in a band is one drawRawPoints(PointMode.lines) over a reused Float32List, not a drawLine per edge.
• Banded painting. Pinned columns are not a second render object. They are index ranges at either end of the same layout, painted into their own clip after the scrolling band, which is why freezing a column costs a clip rather than a parallel widget tree.

Painted cells are invisible to the accessibility tree unless someone puts them there, so this render object assembles its own: a table node holding one row per visible row and one cell per visible cell. That is the price of painting rather than building, and it is paid here, once, instead of being left to the caller.