[Juliusolsson05]

Fixes #921.

The bug

discardServer guarded the server registry deletion with object identity:

if (this.servers.get(server.key) === server) this.servers.delete(server.key)

…and then deleted docs and serverDocuments by server.key.

A server's key is derived from the language spec and the workspace root, so a replacement spawned after a crash reuses it. The registry comparison already knew that; the two cleanup loops did not. So a late callback from the dead server — its exit event, a destroyed-stream error, a rejected initialized — deleted the replacement server's documents.

From the editor: it is attached to a healthy new server, its documents vanish from both maps with diagnostics cleared, and the server still holds them open. Nothing re-opens them until the editor remounts, so completions and diagnostics simply stop for a file that is on screen and fine.

The fix

Each ServerRecord carries a generation, stamped on every document record it owns, and the two loops match on that instead of the reusable key.

What deliberately did not change:

• The dead server's own resources are still disposed unconditionally. Its connection and its process are its own, and leaving them around is a leak whoever holds the key now.
• A server dying with its own documents open still clears them and tells the editor — otherwise the file keeps stale diagnostics from a process that is gone.

Why nextServerGeneration is a named export

The whole fence collapses if a replacement can share its predecessor's value, and createServer cannot be driven without spawning a real language server. Without a named seam, the only thing standing between a working fence and a useless one would be that nobody replaced randomUUID() with a constant — and the suite could not tell. Mutating it to a constant now fails a test.

Tests

Driven against the real discardServer with two real record shapes, because the bug is precisely about telling two records with the same key apart.

mutation	result
docs fenced on the key again (the bug)	2 failed
serverDocuments fenced on the key	2 failed
the generation is shared by every server	1 failed
no cleanup of the server's own docs	2 failed
the dead server is not disposed	1 failed
restored	16 passed

npx tsc -b clean. Wider src/main/ run: 1648 passed, with one failure — src/main/workflows/control.system.test.ts — that fails identically on clean origin/main.

Scope

This is the first of four LSP issues filed together from the same planning pass. #922 (ordered document admission), #923 (post-synchronization revision tickets) and #924 (releasing stale read waits) are untouched here; they are ordering problems rather than identity ones, and they belong with each other rather than with this.

🤖 Generated with Claude Code