Feature/e2e tests

.github/workflows/ci.yml

@@ -35,15 +35,19 @@ jobs:
 
       - name: Check formatting
         if: matrix.rust == 'stable'
-        run: cargo fmt --check
+        run: make check-format
 
       - name: Check generated command reference
         if: matrix.rust == 'stable'
-        run: python3 scripts/update_readme_commands.py --check
+        run: make check-readme
 
       - name: Run tests
-        run: cargo test --locked
+        run: make check-tests
+
+      - name: Run real-server E2E smoke tests
+        if: matrix.rust == 'stable'
+        run: make test-real-server-e2e
 
       - name: Run clippy
         if: matrix.rust == 'stable'
-        run: cargo clippy --all-targets --all-features -- -D warnings
+        run: make check-clippy

.memory/unprocessed/ci-reuses-make-targets.md

@@ -0,0 +1,6 @@
+# CI reuses Makefile verification targets
+
+The initial E2E implementation duplicated formatting, Clippy, and live-server commands between
+`Makefile` and `.github/workflows/ci.yml`. The user explicitly rejected that structure. Keep the
+commands centralized as granular Make targets; CI should only choose which Make target runs for a
+toolchain or lane.

.memory/unprocessed/e2e-cases-are-split-by-language.md

@@ -0,0 +1,6 @@
+# E2E case manifests are split by language
+
+The E2E manifest is a directory rather than one `cases.yaml` file. Global command coverage lives
+in `cases/suite.yaml`; each `cases/<language>.yaml` owns exactly one project and all server pairs
+for that language. The `gowork` case intentionally contains only an isolated `go.work` file so its
+detection does not overlap `gomod` or `go`.

.memory/unprocessed/e2e-fake-lsp-is-a-rust-helper.md

@@ -0,0 +1,5 @@
+# The deterministic E2E LSP fixture is a Rust helper
+
+When offered Python, real rust-analyzer, or a Rust helper for deterministic command-path tests,
+the user selected the Rust helper. Keep the helper as an isolated, locked fixture package that
+reuses serde_json; do not replace it with a Python script merely to reduce fixture build code.

.memory/unprocessed/e2e-fixture-dedup-scope.md

@@ -0,0 +1,5 @@
+# E2E fixture deduplication applies to the Rust harness
+
+The project-phase deduplication item applies to repeated E2E harness setup, including validated
+manifest loading, isolated context construction, and command-strategy selection. It does not mean
+removing the intentionally equivalent domain concepts from the per-language playground sources.

.memory/unprocessed/e2e-servers-use-mason-latest.md

@@ -0,0 +1,9 @@
+# Real-server E2E provisioning uses Mason latest
+
+The user corrected the preferred-server E2E design: do not install LSP servers separately and do
+not pin a Mason registry release or package versions. Each real-server test must use `--download`
+against Mason latest. The preferred smoke pair must be derived from the first production preference
+in `data/lsp-cli.yaml`; do not duplicate that selection in the E2E case YAML.
+
+This deliberately trades reproducibility for immediate upstream compatibility coverage. Preserve
+the resolved Mason source ID in diagnostics so failures can still identify the installed version.

.memory/unprocessed/manual-e2e-verification-follows-provisioning.md

@@ -0,0 +1,6 @@
+# Manual E2E verification follows server provisioning
+
+The project plan originally required running every relevant LSP command against new playgrounds
+before selecting or provisioning their servers. Most new languages had no runnable compatible
+server, so the manual verification item was moved immediately after preferred-server provisioning
+and remains unchecked until those prerequisites exist.

.memory/unprocessed/new-playground-build-metadata.md

@@ -0,0 +1,6 @@
+# New playgrounds use dependency-free and relocatable metadata
+
+The Kotlin playground uses compiler-valid sources plus `settings.gradle.kts`, without adding a
+Gradle or Maven Kotlin plugin dependency. CUDA, Objective-C, and Objective-C++ use relative
+`compile_commands.json` working directories so their compilation metadata remains valid after the
+E2E harness copies a project into a randomized sandbox.

.memory/unprocessed/official-first-e2e-server-pins.md

@@ -0,0 +1,10 @@
+# Official-first E2E server pins expose launcher mismatches
+
+The user selected an official-first preferred E2E server matrix and explicitly accepted bespoke
+provisioning and current command/config gaps. In particular, use `kotlin_lsp` rather than the
+community `kotlin_language_server`, and `roslyn_ls` rather than OmniSharp.
+
+The pinned Mason registry exposes Kotlin LSP as `intellij-server`, although the data config starts
+`kotlin-lsp`. It exposes Roslyn as `roslyn-language-server`, although the data config starts `dotnet`
+with a literal `<my_folder>` DLL path. The preferred-server selection can be committed independently,
+but provisioning must resolve these mismatches before claiming either smoke case is runnable.

.memory/unprocessed/playground-validity-audit.md

@@ -0,0 +1,10 @@
+# Existing playgrounds are not all valid standalone projects
+
+`E2E_TESTS.md` requires every source-language playground to be valid, but the existing C++ fixture
+does not link because `main.cpp` calls undefined `f()` and `g()`. The Rust fixture also cannot be
+checked through its own manifest because Cargo treats the nested package as an undeclared member
+of the repository's root package workspace.
+
+The project audit must distinguish source presence from validated buildability. Repair these
+fixtures in the dedicated cleanup step rather than silently treating their files as a valid E2E
+baseline.

Cargo.toml

@@ -30,6 +30,11 @@ thiserror = "2.0.17"
 url = "2.5.7"
 zip = { version = "8.6.0", default-features = false, features = ["deflate"] }
 
+[dev-dependencies]
+command-group = "5.0.1"
+fs_extra = "1.3.0"
+wait-timeout = "0.2.1"
+
 [lints]
 workspace = true