No latency contract for LSP requests: inlay hints/completion run per keystroke with no target and no ratchet

[abdushakoor12]

Summary

There is no latency contract for any LSP request, and no gate that would catch a
regression in one. [LSPARCH-FEATURES-INLAYHINTS] is two sentences and says
nothing about cost:

Hints cover inferred variable/return types, generic parameters, and call-site
parameter names. Configuration gates are applied before results are returned.

The benchmark ratchet in CLAUDE.md / [CHKARCH-TESTING-BENCH-RATCHET] governs
CLI check wall-clock only. Nothing measures textDocument/inlayHint,
textDocument/completion, hover, or semantic tokens, so their latency can move
in either direction without any signal.

This is filed as a missing contract, not as a performance bug — there is no
target to have regressed against.

What the numbers currently are

Measured against basilisk lsp over stdio on
JelleZijlstra/taxonomy @9b359fa,
taxonomy/shell.py (3,550 lines), sending a didChange before each request so
per-request work cannot be amortised:

Request	median	p90	max	result size
textDocument/inlayHint (full file)	25.0 ms	31.3 ms	49.2 ms	136 hints
textDocument/completion ("abc".)	11.6 ms	14.5 ms	28.3 ms	65 items

A real VS Code session on the same file recorded 90 ms for a single
inlayHint over a 72-line range, concurrent with a diagnostics publish — so
under genuine editor contention the cost is higher than an idle synthetic client
suggests.

Both requests fire on every keystroke. Neither figure is alarming on its own;
the point is that nothing pins them.

Why it matters now

Two recent changes moved per-request cost with no measurement either way:

• Build shared inference, narrowing, and subtyping foundations #377 made textDocument/completion run resolve_module_imports on every
  request (crates/basilisk-lsp/src/server/handlers/features.rs). Builtins are
  cached per snapshot, but the call is per-request.
• Inlay hints recompute a whole-file pass per keystroke, and (per the companion
  issue) throw the entire result away whenever the buffer does not parse.

Either could regress 5× and no test, ratchet, or CI job would notice.

Proposal

1. State a target per interactive request in [LSPARCH-FEATURES-*] — a
   budget of the order of one input frame (~16 ms) for hints/completion on a file
   of a stated size, or whatever the team judges right. The number matters less
   than its existence.
2. Measure it in the harness that already exists. The WebSocket E2E fixture
   (crates/basilisk-lsp/tests/lsp/ws_test_common.rs) already drives real
   JSON-RPC; timing an assertion there is a small addition.
3. Ratchet it the same way [CHKARCH-TESTING-BENCH-RATCHET] ratchets CLI
   times — read the committed baseline, fail on regression.

Worth deciding explicitly as part of this: whether the target is per-request
wall-clock, or per-keystroke total across the requests an edit triggers
(didChange currently fans out to inlayHint, codeLens, foldingRange, and
diagnostics).

Environment

• basilisk 0.0.0-PLACEHOLDER, clean cargo build --release of main @ e3e97d30
• macOS 15 (darwin 25.5.0), arm64, otherwise idle for the synthetic runs