Module lsp_server 

Persistent server for editor/preview integration (latexml_oxide --server).

This is a JSON-RPC-over-stdio server using LSP message framing. It speaks a subset of LSP (initialize, textDocument/did{Open,Change,Close} → publishDiagnostics, shutdown, exit) plus a custom latexml/convert request that returns {html, log, diagnostics, sources, status, statusCode} — the response shape the ar5iv-editor client consumes for its live source↔preview loop (see docs/performance/SOURCE_PROVENANCE.md).

Performance model: the preamble (everything up to and including \begin{document}) is digested once and cached in this (parent) process. Each body conversion fork()s a child that inherits the warm post-preamble state via copy-on-write, digests only the body, builds the DOM, and writes the result back over a pipe before exiting. The child is a throwaway, so a body conversion can never pollute the cache, and a panicking/looping body can’t take down the server.

Concurrency model: a single thread drives everything. While a body child runs, the parent poll(2)s {stdin, child-pipe}; a newer latexml/convert for the same document SIGKILLs the in-flight child (a pid we still own — reaped here, so no PID-recycle race) and supersedes it. Keeping it single-threaded is also what makes the fork() safe: there is no second thread that could hold the allocator lock at fork time.

Functions

run_lsp_server

Run the server. timeout_secs is the per-conversion wall-clock budget (--timeout; 0 disables) and max_memory_mb the resident-memory ceiling (--max-memory; 0 disables). Both are applied fresh per conversion by the forked body child’s shared latexml_core::watchdog::Watchdog — so a child never runs against the parent’s stale warm-up deadline, and is reaped if it exceeds the RAM ceiling. The extension surfaces both as VSCode settings and passes them on the spawn.