Add unit tests for the LSP

[Rigidity]

Closes #63

Summary

• Library/binary split. rue-lsp is now a library (Backend) plus a small binary that runs it over stdio, so the server can be tested.
• Test harness (src/tests/mod.rs). It writes a temporary project to disk, opens each file in the server, and records the cursor from a $0 marker. Results are rendered as text and compared with expect-test snapshots.
• Tests (53). They cover:
  • hover: bindings, parameters, functions, constants, modules, aliases, builtin types, structs (including spread fields), fields
  • go to definition and references, including across files and file modules
  • completion: locals, partial identifiers, types, struct, pair and atom fields, struct initializers, module exports
  • diagnostics: updates on change, warnings, save reverting to the file on disk, errors reported across files
  • formatting, including document_end with CRLF and UTF-16, and fuzzy_match
  • an end-to-end JSON-RPC session through LspService

The snapshots record how the server behaves today, so later refactors will show up as snapshot diffs.

Fixes found by the tests

• Completion at the very start of a file panicked (index - 1 underflow).
• on_change panicked on non-file URIs, invalid Rue.toml files, and missing entrypoints. It now returns no diagnostics in those cases.
• Scope::symbol_names() / type_names() iterated a HashMap, so completion order was different on every run. They now use IndexMap (insertion order).

Behaviour the snapshots record but this PR doesn't change

• Builtin functions show {unresolved} as their type in completions.
• Functions and constants hover as let name: type.
• Generic arguments are dropped in hover, e.g. ...rest: List instead of List<Int>.
• No completions are offered in an empty type position (fn main(value: $0)).
• Go to definition on a declaration returns its references.
• Cursor positions are converted using character counts and byte slicing rather than UTF-16, so non-ASCII text before the cursor can give wrong positions.

Test plan

• cargo test --workspace --all-features
• cargo clippy --workspace --all-features --all-targets (no warnings)
• cargo fmt --check, cargo machete
• Ran the LSP tests 8 times in a row; no flaky results

---

Note

Low Risk
Mostly test infrastructure and defensive error handling in the LSP; no security or data-path changes, with small behavior fixes for panics and completion ordering.

Overview
Adds a broad LSP test suite by turning rue-lsp into a library + thin stdio binary: Backend and server logic live in lib.rs, and main.rs only wires up LspService.

Introduces a TestProject harness (temp Rue.toml project, $0 cursor markers, expect-test snapshots) plus JSON-RPC end-to-end coverage via LspService. Tests lock in hover, go-to-definition/references, completion, diagnostics, formatting, and protocol behavior.

Production fixes bundled with the tests: completion no longer uses index - 1 (avoids panics at BOF); on_change_impl returns empty diagnostics instead of panicking on non-file URIs, bad manifests, or failed compiles; Scope symbol/type name maps use IndexMap so completion order is stable across runs.

^{Reviewed by Cursor Bugbot for commit 7b97d26. Bugbot is set up for automated code reviews on this repo. Configure here.}