Make formatting work, and say which formatter is behind it

[thomasjm]

Went through every language server to see whether formatting works and what backs it.

Fixed where it didn't:

• Rust: rust-analyzer shells out to rustfmt from PATH, which wasn't there, so every format request failed with Failed to spawn "rustfmt". Added it (19 MB; rustc/cargo were already in the closure).
• Ruby: Solargraph gates formatting behind an initialization option that defaults off. Turned on via kernels.ruby.lsp.solargraph.formatting; it formats with rubocop, already a gem dependency with a config in this repo.

Added a setting where there's more than one formatter to choose from, all of them already built into the server except Python's:

Setting	Values	Default
kernels.go.lsp.gopls.formatter	gofmt, gofumpt	gofmt
kernels.haskell.lsp.haskell-language-server.formattingProvider	ormolu, fourmolu, stylish-haskell, none	ormolu
kernels.python3.lsp.python-lsp-server.formatter	autopep8, yapf, black, ruff, none	autopep8
exporters.typst.lsp.tinymist.formatter	typstyle, typstfmt, none	typstyle

clangd, clojure-lsp, LanguageServer.jl, R's languageserver and typescript-language-server needed nothing.

The feature matrix's Formatting column is derived, not hand-written: each language server declares passthru.formatters, and the column counts as supported only when the server advertises formatting and a formatter is named — which is exactly the bash/shfmt failure mode. The formatter's name shows in the SVG cell's tooltip and in the table in FEATURE-MATRIX.md.

Every formatter has a test that formats a misformatted document over LSP and checks the result.

Four languages reach their server through a notebook proxy, so those also get the same test against a main.ipynb cell — the only case where the server sees a bare .ipynb URI. Rust passes. Go passes for a cell of plain statements and is pending for a cell with an import. C++ and Haskell are pending: cpp-notebook-language-server forwards the notebook's URI so clangd answers "trying to format non-added document", and HLS rejects the extension outright ("ormolu does not support .ipynb filetypes"). plans/notebook-formatting-in-proxies.md has the details. The other nine languages set a non-empty notebook_suffix, so codedown hands their server main.ipynb.rb and friends — the same case the raw-file test already covers.

Rust also bumps rust-notebook-language-server to 0.2.5.0, which fixes formatting inside notebook cells (it was forwarding textDocument/formatting with the notebook's URI instead of the shadow file's).

Getting the R language server to run at all

R had no LSP test before this, and four separate things stopped one from working:

• It needs a /bin/sh at startup (processx's .onLoad calls system(), which goes through popen).
• parallel::detectCores() shells out to grep "^processor" /proc/cpuinfo | wc -l, and the hermetic PATH had no grep. Without it detectCores returns NA and the server dies sizing its session pool.
• Test directory names came from show code, so they contained backslashes; R.cache rewrites those to forward slashes and then can't write styler's cache. Formatting test names no longer embed the code.
• languageserver answers shutdown with result: [] where the spec says null. lsp-test couldn't decode it, the thread reading the server's output died, and the session hung in teardown forever. Fixed in lsp-test (lsp-test: don't let an undecodable message or a silent server hang the session lsp#1): undecodable messages are skipped and the shutdown wait is bounded.

The result: [] is a server bug either way, so languageserver now builds from codedownio/languageserver — upstream master plus the one-line fix. That also moves us off the 0.3.16 pin this repo has carried since 2024. The same branch is what goes upstream as a PR; the src here goes back to REditorSupport once it lands.

Unmerged pin

tests/stack.yaml points at codedownio/lsp-test-helpers branch kernel-piggyback-and-binsh (the /bin/sh), which needs merging and the pin moved.

Unrelated change in the regenerated matrix

Go's Semantic highlighting flips to "not available". gopls 0.22 leaves its semanticTokens setting off by default; the stored probe data was stale.