knitr::spin() result not cached, called 3 times per .R file during index build

[cderv]

markdownFromKnitrSpinScript spawns a fresh Rscript.exe process each time it is called, and readBaseInputIndex triggers it 3 times for each .R spin script:

1. renderFormats → resolveFullMarkdownForFile → markdownForFile → spin
2. projectFileMetadata → markdownForFile → spin
3. engine.partitionedMarkdown → spin

quarto-cli/src/execute/rmd.ts

Lines 581 to 600 in 5b88d3d

	export async function markdownFromKnitrSpinScript(file: string) {
	// run spin to get .qmd and get markdown from .qmd
	// TODO: implement a caching system because spin is slow and it seems we call this twice for each run
	// 1. First as part of the target() call
	// 2. Second as part of renderProject() call to get `partitioned` information to get `resourcesFrom` with `resourceFilesFromRenderedFile()`
	// we need a temp dir for `CallR` to work but we don't have access to usual options.tempDir.
	const tempDir = quarto.system.tempContext().createDir();
	const result = await callR<string>(
	"spin",
	{ input: file },
	tempDir,
	undefined,
	true,
	);
	return result;
	}

There is already a TODO in the code:

quarto-cli/src/execute/rmd.ts

Lines 584 to 585 in 5b88d3d

	// TODO: implement a caching system because spin is slow and it seems we call this twice for each run
	// 1. First as part of the target() call

The spin result (markdown output) does not change between calls within the same render, so caching per file would eliminate 2 of the 3 R process spawns. In a project with many .R spin scripts, this adds up significantly (the reporter in #14225 measured 40s vs 3s).

Repro

https://github.com/byzheng/quarto-metadata-files-test

Context

Surfaced while investigating #14225. The spin caching issue is independent of the cache invalidation bug also reported from that investigation.