[red knot] support EllipsisType as a singleton

This is a follow-up on #13746:

- Use `memmem::find` instead of rolling our own inferior version.
- Avoid `x.as_ref()` calls using `&**x`

.github/workflows/ci.yaml

@@ -148,7 +148,7 @@ jobs:
       # sync, not just public items. Eventually we should do this for all
       # crates; for now add crates here as they are warning-clean to prevent
       # regression.
-      - run: cargo doc --no-deps -p red_knot_python_semantic -p red_knot -p ruff_db --document-private-items
+      - run: cargo doc --no-deps -p red_knot_python_semantic -p red_knot -p red_knot_test -p ruff_db --document-private-items
         env:
           # Setting RUSTDOCFLAGS because `cargo doc --check` isn't yet implemented (https://github.com/rust-lang/cargo/issues/10025).
           RUSTDOCFLAGS: "-D warnings"
@@ -518,6 +518,8 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
+        with:
+          python-version: "3.13"
       - name: "Add SSH key"
         if: ${{ env.MKDOCS_INSIDERS_SSH_KEY_EXISTS == 'true' }}
         uses: webfactory/ssh-agent@v0.9.0
@@ -525,13 +527,15 @@ jobs:
           ssh-private-key: ${{ secrets.MKDOCS_INSIDERS_SSH_KEY }}
       - name: "Install Rust toolchain"
         run: rustup show
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
       - uses: Swatinem/rust-cache@v2
       - name: "Install Insiders dependencies"
         if: ${{ env.MKDOCS_INSIDERS_SSH_KEY_EXISTS == 'true' }}
-        run: pip install -r docs/requirements-insiders.txt
+        run: uv pip install -r docs/requirements-insiders.txt --system
       - name: "Install dependencies"
         if: ${{ env.MKDOCS_INSIDERS_SSH_KEY_EXISTS != 'true' }}
-        run: pip install -r docs/requirements.txt
+        run: uv pip install -r docs/requirements.txt --system
       - name: "Update README File"
         run: python scripts/transform_readme.py --target mkdocs
       - name: "Generate docs"
@@ -608,7 +612,7 @@ jobs:
           just test
 
   benchmarks:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-22.04
     needs: determine_changes
     if: ${{ github.repository == 'astral-sh/ruff' && (needs.determine_changes.outputs.code == 'true' || github.ref == 'refs/heads/main') }}
     timeout-minutes: 20

.github/workflows/release.yml

@@ -1,3 +1,5 @@
+# This file was autogenerated by cargo-dist: https://opensource.axo.dev/cargo-dist/
+#
 # Copyright 2022-2024, axodotdev
 # SPDX-License-Identifier: MIT or Apache-2.0
 #
@@ -64,7 +66,7 @@ jobs:
         # we specify bash to get pipefail; it guards against the `curl` command
         # failing. otherwise `sh` won't catch that `curl` returned non-0
         shell: bash
-        run: "curl --proto '=https' --tlsv1.2 -LsSf https://github.com/axodotdev/cargo-dist/releases/download/v0.18.0/cargo-dist-installer.sh | sh"
+        run: "curl --proto '=https' --tlsv1.2 -LsSf https://github.com/axodotdev/cargo-dist/releases/download/v0.22.1/cargo-dist-installer.sh | sh"
       - name: Cache cargo-dist
         uses: actions/upload-artifact@v4
         with:

.github/workflows/sync_typeshed.yaml

@@ -37,13 +37,13 @@ jobs:
       - name: Sync typeshed
         id: sync
         run: |
-          rm -rf ruff/crates/red_knot_python_semantic/vendor/typeshed
-          mkdir ruff/crates/red_knot_python_semantic/vendor/typeshed
-          cp typeshed/README.md ruff/crates/red_knot_python_semantic/vendor/typeshed
-          cp typeshed/LICENSE ruff/crates/red_knot_python_semantic/vendor/typeshed
-          cp -r typeshed/stdlib ruff/crates/red_knot_python_semantic/vendor/typeshed/stdlib
-          rm -rf ruff/crates/red_knot_python_semantic/vendor/typeshed/stdlib/@tests
-          git -C typeshed rev-parse HEAD > ruff/crates/red_knot_python_semantic/vendor/typeshed/source_commit.txt
+          rm -rf ruff/crates/red_knot_vendored/vendor/typeshed
+          mkdir ruff/crates/red_knot_vendored/vendor/typeshed
+          cp typeshed/README.md ruff/crates/red_knot_vendored/vendor/typeshed
+          cp typeshed/LICENSE ruff/crates/red_knot_vendored/vendor/typeshed
+          cp -r typeshed/stdlib ruff/crates/red_knot_vendored/vendor/typeshed/stdlib
+          rm -rf ruff/crates/red_knot_vendored/vendor/typeshed/stdlib/@tests
+          git -C typeshed rev-parse HEAD > ruff/crates/red_knot_vendored/vendor/typeshed/source_commit.txt
       - name: Commit the changes
         id: commit
         if: ${{ steps.sync.outcome == 'success' }}

.pre-commit-config.yaml

@@ -2,7 +2,7 @@ fail_fast: true
 
 exclude: |
   (?x)^(
-    crates/red_knot_python_semantic/vendor/.*|
+    crates/red_knot_vendored/vendor/.*|
     crates/red_knot_workspace/resources/.*|
     crates/ruff_linter/resources/.*|
     crates/ruff_linter/src/rules/.*/snapshots/.*|
@@ -17,7 +17,7 @@ exclude: |
 
 repos:
   - repo: https://github.com/abravalheri/validate-pyproject
-    rev: v0.19
+    rev: v0.20.2
     hooks:
       - id: validate-pyproject
 
@@ -28,6 +28,7 @@ repos:
         additional_dependencies:
           - mdformat-mkdocs
           - mdformat-admon
+          - mdformat-footnote
         exclude: |
           (?x)^(
             docs/formatter/black\.md
@@ -35,7 +36,7 @@ repos:
           )$
 
   - repo: https://github.com/igorshubovych/markdownlint-cli
-    rev: v0.41.0
+    rev: v0.42.0
     hooks:
       - id: markdownlint-fix
         exclude: |
@@ -45,7 +46,7 @@ repos:
           )$
 
   - repo: https://github.com/crate-ci/typos
-    rev: v1.24.5
+    rev: v1.25.0
     hooks:
       - id: typos
 
@@ -59,7 +60,7 @@ repos:
         pass_filenames: false # This makes it a lot faster
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.6.5
+    rev: v0.6.9
     hooks:
       - id: ruff-format
       - id: ruff
@@ -68,8 +69,8 @@ repos:
         require_serial: true
 
   # Prettier
-  - repo: https://github.com/pre-commit/mirrors-prettier
-    rev: v3.1.0
+  - repo: https://github.com/rbubley/mirrors-prettier
+    rev: v3.3.3
     hooks:
       - id: prettier
         types: [yaml]

CHANGELOG.md

@@ -1,5 +1,105 @@
 # Changelog
 
+## 0.6.9
+
+### Preview features
+
+- Fix codeblock dynamic line length calculation for indented docstring examples ([#13523](https://github.com/astral-sh/ruff/pull/13523))
+- \[`refurb`\] Mark `FURB118` fix as unsafe ([#13613](https://github.com/astral-sh/ruff/pull/13613))
+
+### Rule changes
+
+- \[`pydocstyle`\] Don't raise `D208` when last line is non-empty ([#13372](https://github.com/astral-sh/ruff/pull/13372))
+- \[`pylint`\] Preserve trivia (i.e. comments) in `PLR5501` autofix ([#13573](https://github.com/astral-sh/ruff/pull/13573))
+
+### Configuration
+
+- \[`pyflakes`\] Add `allow-unused-imports` setting for `unused-import` rule (`F401`) ([#13601](https://github.com/astral-sh/ruff/pull/13601))
+
+### Bug fixes
+
+- Support ruff discovery in pip build environments ([#13591](https://github.com/astral-sh/ruff/pull/13591))
+- \[`flake8-bugbear`\] Avoid short circuiting `B017` for multiple context managers ([#13609](https://github.com/astral-sh/ruff/pull/13609))
+- \[`pylint`\] Do not offer an invalid fix for `PLR1716` when the comparisons contain parenthesis ([#13527](https://github.com/astral-sh/ruff/pull/13527))
+- \[`pyupgrade`\] Fix `UP043` to apply to `collections.abc.Generator` and `collections.abc.AsyncGenerator` ([#13611](https://github.com/astral-sh/ruff/pull/13611))
+- \[`refurb`\] Fix handling of slices in tuples for `FURB118`, e.g., `x[:, 1]` ([#13518](https://github.com/astral-sh/ruff/pull/13518))
+
+### Documentation
+
+- Update GitHub Action link to `astral-sh/ruff-action` ([#13551](https://github.com/astral-sh/ruff/pull/13551))
+
+## 0.6.8
+
+### Preview features
+
+- Remove unnecessary parentheses around `match case` clauses ([#13510](https://github.com/astral-sh/ruff/pull/13510))
+- Parenthesize overlong `if` guards in `match..case` clauses ([#13513](https://github.com/astral-sh/ruff/pull/13513))
+- Detect basic wildcard imports in `ruff analyze graph` ([#13486](https://github.com/astral-sh/ruff/pull/13486))
+- \[`pylint`\] Implement `boolean-chained-comparison` (`R1716`) ([#13435](https://github.com/astral-sh/ruff/pull/13435))
+
+### Rule changes
+
+- \[`lake8-simplify`\] Detect `SIM910` when using variadic keyword arguments, i.e., `**kwargs` ([#13503](https://github.com/astral-sh/ruff/pull/13503))
+- \[`pyupgrade`\] Avoid false negatives with non-reference shadowed bindings of loop variables (`UP028`) ([#13504](https://github.com/astral-sh/ruff/pull/13504))
+
+### Bug fixes
+
+- Detect tuples bound to variadic positional arguments i.e. `*args` ([#13512](https://github.com/astral-sh/ruff/pull/13512))
+- Exit gracefully on broken pipe errors ([#13485](https://github.com/astral-sh/ruff/pull/13485))
+- Avoid panic when analyze graph hits broken pipe ([#13484](https://github.com/astral-sh/ruff/pull/13484))
+
+### Performance
+
+- Reuse `BTreeSets` in module resolver ([#13440](https://github.com/astral-sh/ruff/pull/13440))
+- Skip traversal for non-compound statements ([#13441](https://github.com/astral-sh/ruff/pull/13441))
+
+## 0.6.7
+
+### Preview features
+
+- Add Python version support to ruff analyze CLI ([#13426](https://github.com/astral-sh/ruff/pull/13426))
+- Add `exclude` support to `ruff analyze` ([#13425](https://github.com/astral-sh/ruff/pull/13425))
+- Fix parentheses around return type annotations ([#13381](https://github.com/astral-sh/ruff/pull/13381))
+
+### Rule changes
+
+- \[`pycodestyle`\] Fix: Don't autofix if the first line ends in a question mark? (D400) ([#13399](https://github.com/astral-sh/ruff/pull/13399))
+
+### Bug fixes
+
+- Respect `lint.exclude` in ruff check `--add-noqa` ([#13427](https://github.com/astral-sh/ruff/pull/13427))
+
+### Performance
+
+- Avoid tracking module resolver files in Salsa ([#13437](https://github.com/astral-sh/ruff/pull/13437))
+- Use `forget` for module resolver database ([#13438](https://github.com/astral-sh/ruff/pull/13438))
+
+## 0.6.6
+
+### Preview features
+
+- \[`refurb`\] Skip `slice-to-remove-prefix-or-suffix` (`FURB188`) when non-trivial slice steps are present ([#13405](https://github.com/astral-sh/ruff/pull/13405))
+- Add a subcommand to generate dependency graphs ([#13402](https://github.com/astral-sh/ruff/pull/13402))
+
+### Formatter
+
+- Fix placement of inline parameter comments ([#13379](https://github.com/astral-sh/ruff/pull/13379))
+
+### Server
+
+- Fix off-by one error in the `LineIndex::offset` calculation ([#13407](https://github.com/astral-sh/ruff/pull/13407))
+
+### Bug fixes
+
+- \[`fastapi`\] Respect FastAPI aliases in route definitions ([#13394](https://github.com/astral-sh/ruff/pull/13394))
+- \[`pydocstyle`\] Respect word boundaries when detecting function signature in docs ([#13388](https://github.com/astral-sh/ruff/pull/13388))
+
+### Documentation
+
+- Add backlinks to rule overview linter ([#13368](https://github.com/astral-sh/ruff/pull/13368))
+- Fix documentation for editor vim plugin ALE ([#13348](https://github.com/astral-sh/ruff/pull/13348))
+- Fix rendering of `FURB188` docs ([#13406](https://github.com/astral-sh/ruff/pull/13406))
+
 ## 0.6.5
 
 ### Preview features

CONTRIBUTING.md

@@ -29,16 +29,14 @@ You'll also need [Insta](https://insta.rs/docs/) to update snapshot tests:
 cargo install cargo-insta
 ```
 
-And you'll need pre-commit to run some validation checks:
-
-```shell
-pipx install pre-commit  # or `pip install pre-commit` if you have a virtualenv
-```
+You'll need [uv](https://docs.astral.sh/uv/getting-started/installation/) (or `pipx` and `pip`) to
+run Python utility commands.
 
 You can optionally install pre-commit hooks to automatically run the validation checks
 when making a commit:
 
 ```shell
+uv tool install pre-commit
 pre-commit install
 ```
 
@@ -66,7 +64,7 @@ and that it passes both the lint and test validation checks:
 ```shell
 cargo clippy --workspace --all-targets --all-features -- -D warnings  # Rust linting
 RUFF_UPDATE_SCHEMA=1 cargo test  # Rust testing and updating ruff.schema.json
-pre-commit run --all-files --show-diff-on-failure  # Rust and Python formatting, Markdown and Python linting, etc.
+uvx pre-commit run --all-files --show-diff-on-failure  # Rust and Python formatting, Markdown and Python linting, etc.
 ```
 
 These checks will run on GitHub Actions when you open your pull request, but running them locally
@@ -267,26 +265,20 @@ To preview any changes to the documentation locally:
 
 1. Install the [Rust toolchain](https://www.rust-lang.org/tools/install).
 
-1. Install MkDocs and Material for MkDocs with:
-
-    ```shell
-    pip install -r docs/requirements.txt
-    ```
-
 1. Generate the MkDocs site with:
 
     ```shell
-    python scripts/generate_mkdocs.py
+    uv run --no-project --isolated --with-requirements docs/requirements.txt scripts/generate_mkdocs.py
     ```
 
 1. Run the development server with:
 
     ```shell
     # For contributors.
-    mkdocs serve -f mkdocs.public.yml
+    uvx --with-requirements docs/requirements.txt -- mkdocs serve -f mkdocs.public.yml
 
     # For members of the Astral org, which has access to MkDocs Insiders via sponsorship.
-    mkdocs serve -f mkdocs.insiders.yml
+    uvx --with-requirements docs/requirements-insiders.txt -- mkdocs serve -f mkdocs.insiders.yml
     ```
 
 The documentation should then be available locally at
@@ -368,9 +360,8 @@ GitHub Actions will run your changes against a number of real-world projects fro
 report on any linter or formatter differences. You can also run those checks locally via:
 
 ```shell
-pip install -e ./python/ruff-ecosystem
-ruff-ecosystem check ruff "./target/debug/ruff"
-ruff-ecosystem format ruff "./target/debug/ruff"
+uvx --from ./python/ruff-ecosystem ruff-ecosystem check ruff "./target/debug/ruff"
+uvx --from ./python/ruff-ecosystem ruff-ecosystem format ruff "./target/debug/ruff"
 ```
 
 See the [ruff-ecosystem package](https://github.com/astral-sh/ruff/tree/main/python/ruff-ecosystem) for more details.

Cargo.toml

@@ -14,9 +14,10 @@ license = "MIT"
 [workspace.dependencies]
 ruff = { path = "crates/ruff" }
 ruff_cache = { path = "crates/ruff_cache" }
-ruff_db = { path = "crates/ruff_db" }
+ruff_db = { path = "crates/ruff_db", default-features = false }
 ruff_diagnostics = { path = "crates/ruff_diagnostics" }
 ruff_formatter = { path = "crates/ruff_formatter" }
+ruff_graph = { path = "crates/ruff_graph" }
 ruff_index = { path = "crates/ruff_index" }
 ruff_linter = { path = "crates/ruff_linter" }
 ruff_macros = { path = "crates/ruff_macros" }
@@ -33,15 +34,18 @@ ruff_python_trivia = { path = "crates/ruff_python_trivia" }
 ruff_server = { path = "crates/ruff_server" }
 ruff_source_file = { path = "crates/ruff_source_file" }
 ruff_text_size = { path = "crates/ruff_text_size" }
+red_knot_vendored = { path = "crates/red_knot_vendored" }
 ruff_workspace = { path = "crates/ruff_workspace" }
 
 red_knot_python_semantic = { path = "crates/red_knot_python_semantic" }
 red_knot_server = { path = "crates/red_knot_server" }
-red_knot_workspace = { path = "crates/red_knot_workspace" }
+red_knot_test = { path = "crates/red_knot_test" }
+red_knot_workspace = { path = "crates/red_knot_workspace", default-features = false }
 
 aho-corasick = { version = "1.1.3" }
 annotate-snippets = { version = "0.9.2", features = ["color"] }
 anyhow = { version = "1.0.80" }
+assert_fs = { version = "1.1.0" }
 argfile = { version = "0.2.0" }
 bincode = { version = "1.3.3" }
 bitflags = { version = "2.5.0" }
@@ -68,7 +72,11 @@ fern = { version = "0.6.1" }
 filetime = { version = "0.2.23" }
 glob = { version = "0.3.1" }
 globset = { version = "0.4.14" }
-hashbrown = "0.14.3"
+globwalk = { version = "0.9.1" }
+hashbrown = { version = "0.15.0", default-features = false, features = [
+    "raw-entry",
+    "inline-more",
+] }
 ignore = { version = "0.4.22" }
 imara-diff = { version = "0.1.5" }
 imperative = { version = "1.0.4" }
@@ -86,7 +94,7 @@ libcst = { version = "1.1.0", default-features = false }
 log = { version = "0.4.17" }
 lsp-server = { version = "0.7.6" }
 lsp-types = { git = "https://github.com/astral-sh/lsp-types.git", rev = "3512a9f", features = [
-  "proposed",
+    "proposed",
 ] }
 matchit = { version = "0.8.1" }
 memchr = { version = "2.7.1" }
@@ -107,8 +115,9 @@ quote = { version = "1.0.23" }
 rand = { version = "0.8.5" }
 rayon = { version = "1.10.0" }
 regex = { version = "1.10.2" }
+rstest = { version = "0.22.0", default-features = false }
 rustc-hash = { version = "2.0.0" }
-salsa = { git = "https://github.com/salsa-rs/salsa.git", rev = "f608ff8b24f07706492027199f51132244034f29" }
+salsa = { git = "https://github.com/salsa-rs/salsa.git", rev = "4a7c955255e707e64e43f3ce5eabb771ae067768" }
 schemars = { version = "0.8.16" }
 seahash = { version = "4.1.0" }
 serde = { version = "1.0.197", features = ["derive"] }
@@ -116,7 +125,7 @@ serde-wasm-bindgen = { version = "0.6.4" }
 serde_json = { version = "1.0.113" }
 serde_test = { version = "1.0.152" }
 serde_with = { version = "3.6.0", default-features = false, features = [
-  "macros",
+    "macros",
 ] }
 shellexpand = { version = "3.0.0" }
 similar = { version = "2.4.0", features = ["inline"] }
@@ -133,7 +142,10 @@ toml = { version = "0.8.11" }
 tracing = { version = "0.1.40" }
 tracing-flame = { version = "0.2.0" }
 tracing-indicatif = { version = "0.3.6" }
-tracing-subscriber = { version = "0.3.18", default-features = false, features = ["env-filter", "fmt"] }
+tracing-subscriber = { version = "0.3.18", default-features = false, features = [
+    "env-filter",
+    "fmt",
+] }
 tracing-tree = { version = "0.4.0" }
 typed-arena = { version = "2.0.2" }
 unic-ucd-category = { version = "0.9" }
@@ -144,10 +156,10 @@ unicode-normalization = { version = "0.1.23" }
 ureq = { version = "2.9.6" }
 url = { version = "2.5.0" }
 uuid = { version = "1.6.1", features = [
-  "v4",
-  "fast-rng",
-  "macro-diagnostics",
-  "js",
+    "v4",
+    "fast-rng",
+    "macro-diagnostics",
+    "js",
 ] }
 walkdir = { version = "2.3.2" }
 wasm-bindgen = { version = "0.2.92" }
@@ -158,7 +170,10 @@ zip = { version = "0.6.6", default-features = false }
 [workspace.lints.rust]
 unsafe_code = "warn"
 unreachable_pub = "warn"
-unexpected_cfgs = { level = "warn", check-cfg = ["cfg(fuzzing)", "cfg(codspeed)"] }
+unexpected_cfgs = { level = "warn", check-cfg = [
+    "cfg(fuzzing)",
+    "cfg(codspeed)",
+] }
 
 [workspace.lints.clippy]
 pedantic = { level = "warn", priority = -2 }
@@ -230,9 +245,9 @@ inherits = "release"
 # Config for 'cargo dist'
 [workspace.metadata.dist]
 # The preferred cargo-dist version to use in CI (Cargo.toml SemVer syntax)
-cargo-dist-version = "0.18.0"
+cargo-dist-version = "0.22.1"
 # CI backends to support
-ci = ["github"]
+ci = "github"
 # The installers to generate for each app
 installers = ["shell", "powershell"]
 # The archive format to use for windows builds (defaults .zip)
@@ -241,33 +256,33 @@ windows-archive = ".zip"
 unix-archive = ".tar.gz"
 # Target platforms to build apps for (Rust target-triple syntax)
 targets = [
-  "aarch64-apple-darwin",
-  "aarch64-pc-windows-msvc",
-  "aarch64-unknown-linux-gnu",
-  "aarch64-unknown-linux-musl",
-  "arm-unknown-linux-musleabihf",
-  "armv7-unknown-linux-gnueabihf",
-  "armv7-unknown-linux-musleabihf",
-  "i686-pc-windows-msvc",
-  "i686-unknown-linux-gnu",
-  "i686-unknown-linux-musl",
-  "powerpc64-unknown-linux-gnu",
-  "powerpc64le-unknown-linux-gnu",
-  "s390x-unknown-linux-gnu",
-  "x86_64-apple-darwin",
-  "x86_64-pc-windows-msvc",
-  "x86_64-unknown-linux-gnu",
-  "x86_64-unknown-linux-musl",
+    "aarch64-apple-darwin",
+    "aarch64-pc-windows-msvc",
+    "aarch64-unknown-linux-gnu",
+    "aarch64-unknown-linux-musl",
+    "arm-unknown-linux-musleabihf",
+    "armv7-unknown-linux-gnueabihf",
+    "armv7-unknown-linux-musleabihf",
+    "i686-pc-windows-msvc",
+    "i686-unknown-linux-gnu",
+    "i686-unknown-linux-musl",
+    "powerpc64-unknown-linux-gnu",
+    "powerpc64le-unknown-linux-gnu",
+    "s390x-unknown-linux-gnu",
+    "x86_64-apple-darwin",
+    "x86_64-pc-windows-msvc",
+    "x86_64-unknown-linux-gnu",
+    "x86_64-unknown-linux-musl",
 ]
 # Whether to auto-include files like READMEs, LICENSEs, and CHANGELOGs (default true)
 auto-includes = false
 # Whether cargo-dist should create a GitHub Release or use an existing draft
 create-release = true
-# Publish jobs to run in CI
+# Which actions to run on pull requests
 pr-run-mode = "skip"
 # Whether CI should trigger releases with dispatches instead of tag pushes
 dispatch-releases = true
-# The stage during which the GitHub Release should be created
+# Which phase cargo-dist should use to create the GitHub release
 github-release = "announce"
 # Whether CI should include auto-generated code to build local artifacts
 build-local-artifacts = false
@@ -275,9 +290,15 @@ build-local-artifacts = false
 local-artifacts-jobs = ["./build-binaries", "./build-docker"]
 # Publish jobs to run in CI
 publish-jobs = ["./publish-pypi", "./publish-wasm"]
-# Announcement jobs to run in CI
-post-announce-jobs = ["./notify-dependents", "./publish-docs", "./publish-playground"]
+# Post-announce jobs to run in CI
+post-announce-jobs = [
+    "./notify-dependents",
+    "./publish-docs",
+    "./publish-playground",
+]
 # Custom permissions for GitHub Jobs
 github-custom-job-permissions = { "build-docker" = { packages = "write", contents = "read" }, "publish-wasm" = { contents = "read", id-token = "write", packages = "write" } }
 # Whether to install an updater program
 install-updater = false
+# Path that installers should place binaries in
+install-path = "CARGO_HOME"

README.md

@@ -136,8 +136,8 @@ curl -LsSf https://astral.sh/ruff/install.sh | sh
 powershell -c "irm https://astral.sh/ruff/install.ps1 | iex"
 
 # For a specific version.
-curl -LsSf https://astral.sh/ruff/0.6.5/install.sh | sh
-powershell -c "irm https://astral.sh/ruff/0.6.5/install.ps1 | iex"
+curl -LsSf https://astral.sh/ruff/0.6.9/install.sh | sh
+powershell -c "irm https://astral.sh/ruff/0.6.9/install.ps1 | iex"
 ```
 
 You can also install Ruff via [Homebrew](https://formulae.brew.sh/formula/ruff), [Conda](https://anaconda.org/conda-forge/ruff),
@@ -170,7 +170,7 @@ Ruff can also be used as a [pre-commit](https://pre-commit.com/) hook via [`ruff
 ```yaml
 - repo: https://github.com/astral-sh/ruff-pre-commit
   # Ruff version.
-  rev: v0.6.5
+  rev: v0.6.9
   hooks:
     # Run the linter.
     - id: ruff
@@ -182,7 +182,7 @@ Ruff can also be used as a [pre-commit](https://pre-commit.com/) hook via [`ruff
 Ruff can also be used as a [VS Code extension](https://github.com/astral-sh/ruff-vscode) or with [various other editors](https://docs.astral.sh/ruff/editors/setup).
 
 Ruff can also be used as a [GitHub Action](https://github.com/features/actions) via
-[`ruff-action`](https://github.com/chartboost/ruff-action):
+[`ruff-action`](https://github.com/astral-sh/ruff-action):
 
 ```yaml
 name: Ruff
@@ -192,7 +192,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
-      - uses: chartboost/ruff-action@v1
+      - uses: astral-sh/ruff-action@v1
 ```
 
 ### Configuration<a id="configuration"></a>

_typos.toml

@@ -1,6 +1,6 @@
 [files]
 # https://github.com/crate-ci/typos/issues/868
-extend-exclude = ["crates/red_knot_python_semantic/vendor/**/*", "**/resources/**/*", "**/snapshots/**/*"]
+extend-exclude = ["crates/red_knot_vendored/vendor/**/*", "**/resources/**/*", "**/snapshots/**/*"]
 
 [default.extend-words]
 "arange" = "arange"  # e.g. `numpy.arange`
@@ -8,7 +8,7 @@ hel = "hel"
 whos = "whos"
 spawnve = "spawnve"
 ned = "ned"
-pn = "pn"  # `import panel as pd` is a thing
+pn = "pn"  # `import panel as pn` is a thing
 poit = "poit"
 BA = "BA" # acronym for "Bad Allowed", used in testing.
 jod = "jod" # e.g., `jod-thread`

crates/red_knot/Cargo.toml

@@ -13,9 +13,8 @@ license.workspace = true
 
 [dependencies]
 red_knot_python_semantic = { workspace = true }
-red_knot_workspace = { workspace = true }
+red_knot_workspace = { workspace = true, features = ["zstd"] }
 red_knot_server = { workspace = true }
-
 ruff_db = { workspace = true, features = ["os", "cache"] }
 
 anyhow = { workspace = true }

crates/red_knot/src/main.rs

@@ -160,7 +160,7 @@ fn run() -> anyhow::Result<ExitStatus> {
         SystemPathBuf::from_path_buf(cwd)
             .map_err(|path| {
                 anyhow!(
-                    "The current working directory '{}' contains non-unicode characters. Red Knot only supports unicode paths.",
+                    "The current working directory `{}` contains non-Unicode characters. Red Knot only supports Unicode paths.",
                     path.display()
                 )
             })?
@@ -174,7 +174,7 @@ fn run() -> anyhow::Result<ExitStatus> {
                 Ok(SystemPath::absolute(cwd, &cli_base_path))
             } else {
                 Err(anyhow!(
-                    "Provided current-directory path '{cwd}' is not a directory."
+                    "Provided current-directory path `{cwd}` is not a directory"
                 ))
             }
         })

crates/red_knot/tests/file_watching.rs

@@ -42,14 +42,14 @@ impl TestCase {
 
     fn stop_watch(&mut self) -> Vec<watch::ChangeEvent> {
         self.try_stop_watch(Duration::from_secs(10))
-            .expect("Expected watch changes but observed none.")
+            .expect("Expected watch changes but observed none")
     }
 
     fn try_stop_watch(&mut self, timeout: Duration) -> Option<Vec<watch::ChangeEvent>> {
         let watcher = self
             .watcher
             .take()
-            .expect("Cannot call `stop_watch` more than once.");
+            .expect("Cannot call `stop_watch` more than once");
 
         let mut all_events = self
             .changes_receiver
@@ -72,7 +72,7 @@ impl TestCase {
     #[cfg(unix)]
     fn take_watch_changes(&self) -> Vec<watch::ChangeEvent> {
         self.try_take_watch_changes(Duration::from_secs(10))
-            .expect("Expected watch changes but observed none.")
+            .expect("Expected watch changes but observed none")
     }
 
     fn try_take_watch_changes(&self, timeout: Duration) -> Option<Vec<watch::ChangeEvent>> {
@@ -150,14 +150,14 @@ where
             let absolute_path = workspace_path.join(relative_path);
             if let Some(parent) = absolute_path.parent() {
                 std::fs::create_dir_all(parent).with_context(|| {
-                    format!("Failed to create parent directory for file '{relative_path}'.",)
+                    format!("Failed to create parent directory for file `{relative_path}`")
                 })?;
             }
 
             let mut file = std::fs::File::create(absolute_path.as_std_path())
-                .with_context(|| format!("Failed to open file '{relative_path}'"))?;
+                .with_context(|| format!("Failed to open file `{relative_path}`"))?;
             file.write_all(content.as_bytes())
-                .with_context(|| format!("Failed to write to file '{relative_path}'"))?;
+                .with_context(|| format!("Failed to write to file `{relative_path}`"))?;
             file.sync_data()?;
         }
 
@@ -194,7 +194,7 @@ where
 
     let root_path = SystemPath::from_std_path(temp_dir.path()).ok_or_else(|| {
         anyhow!(
-            "Temp directory '{}' is not a valid UTF-8 path.",
+            "Temporary directory `{}` is not a valid UTF-8 path.",
             temp_dir.path().display()
         )
     })?;
@@ -209,7 +209,7 @@ where
     let workspace_path = root_path.join("workspace");
 
     std::fs::create_dir_all(workspace_path.as_std_path())
-        .with_context(|| format!("Failed to create workspace directory '{workspace_path}'",))?;
+        .with_context(|| format!("Failed to create workspace directory `{workspace_path}`"))?;
 
     setup_files
         .setup(&root_path, &workspace_path)
@@ -233,7 +233,7 @@ where
         }))
     {
         std::fs::create_dir_all(path.as_std_path())
-            .with_context(|| format!("Failed to create search path '{path}'"))?;
+            .with_context(|| format!("Failed to create search path `{path}`"))?;
     }
 
     let configuration = Configuration {
@@ -665,7 +665,7 @@ fn directory_deleted() -> anyhow::Result<()> {
 
     let bar = case.system_file(case.workspace_path("bar.py")).unwrap();
 
-    assert!(resolve_module(case.db().upcast(), ModuleName::new_static("sub.a").unwrap()).is_some(),);
+    assert!(resolve_module(case.db().upcast(), ModuleName::new_static("sub.a").unwrap()).is_some());
 
     let sub_path = case.workspace_path("sub");
 

crates/red_knot_python_semantic/Cargo.toml

@@ -24,7 +24,7 @@ bitflags = { workspace = true }
 camino = { workspace = true }
 compact_str = { workspace = true }
 countme = { workspace = true }
-once_cell = { workspace = true }
+itertools = { workspace = true}
 ordermap = { workspace = true }
 salsa = { workspace = true }
 thiserror = { workspace = true }
@@ -34,21 +34,18 @@ hashbrown = { workspace = true }
 smallvec = { workspace = true }
 static_assertions = { workspace = true }
 test-case = { workspace = true }
-
-[build-dependencies]
-path-slash = { workspace = true }
-walkdir = { workspace = true }
-zip = { workspace = true, features = ["zstd", "deflate"] }
+memchr = { workspace = true }
 
 [dev-dependencies]
 ruff_db = { workspace = true, features = ["os", "testing"] }
 ruff_python_parser = { workspace = true }
+red_knot_test = { workspace = true }
+red_knot_vendored = { workspace = true }
 
 anyhow = { workspace = true }
 insta = { workspace = true }
+rstest = { workspace = true }
 tempfile = { workspace = true }
-walkdir = { workspace = true }
-zip = { workspace = true }
 
 [lints]
 workspace = true

crates/red_knot_python_semantic/build.rs

@@ -1,87 +1,4 @@
-//! Build script to package our vendored typeshed files
-//! into a zip archive that can be included in the Ruff binary.
-//!
-//! This script should be automatically run at build time
-//! whenever the script itself changes, or whenever any files
-//! in `crates/red_knot_python_semantic/vendor/typeshed` change.
-
-use std::fs::File;
-use std::path::Path;
-
-use path_slash::PathExt;
-use zip::result::ZipResult;
-use zip::write::{FileOptions, ZipWriter};
-use zip::CompressionMethod;
-
-const TYPESHED_SOURCE_DIR: &str = "vendor/typeshed";
-const TYPESHED_ZIP_LOCATION: &str = "/zipped_typeshed.zip";
-
-/// Recursively zip the contents of an entire directory.
-///
-/// This routine is adapted from a recipe at
-/// <https://github.com/zip-rs/zip-old/blob/5d0f198124946b7be4e5969719a7f29f363118cd/examples/write_dir.rs>
-fn zip_dir(directory_path: &str, writer: File) -> ZipResult<File> {
-    let mut zip = ZipWriter::new(writer);
-
-    // Use deflated compression for WASM builds because compiling `zstd-sys` requires clang
-    // [source](https://github.com/gyscos/zstd-rs/wiki/Compile-for-WASM) which complicates the build
-    // by a lot. Deflated compression is slower but it shouldn't matter much for the WASM use case
-    // (WASM itself is already slower than a native build for a specific platform).
-    // We can't use `#[cfg(...)]` here because the target-arch in a build script is the
-    // architecture of the system running the build script and not the architecture of the build-target.
-    // That's why we use the `TARGET` environment variable here.
-    let method = if std::env::var("TARGET").unwrap().contains("wasm32") {
-        CompressionMethod::Deflated
-    } else {
-        CompressionMethod::Zstd
-    };
-
-    let options = FileOptions::default()
-        .compression_method(method)
-        .unix_permissions(0o644);
-
-    for entry in walkdir::WalkDir::new(directory_path) {
-        let dir_entry = entry.unwrap();
-        let absolute_path = dir_entry.path();
-        let normalized_relative_path = absolute_path
-            .strip_prefix(Path::new(directory_path))
-            .unwrap()
-            .to_slash()
-            .expect("Unexpected non-utf8 typeshed path!");
-
-        // Write file or directory explicitly
-        // Some unzip tools unzip files with directory paths correctly, some do not!
-        if absolute_path.is_file() {
-            println!("adding file {absolute_path:?} as {normalized_relative_path:?} ...");
-            zip.start_file(normalized_relative_path, options)?;
-            let mut f = File::open(absolute_path)?;
-            std::io::copy(&mut f, &mut zip).unwrap();
-        } else if !normalized_relative_path.is_empty() {
-            // Only if not root! Avoids path spec / warning
-            // and mapname conversion failed error on unzip
-            println!("adding dir {absolute_path:?} as {normalized_relative_path:?} ...");
-            zip.add_directory(normalized_relative_path, options)?;
-        }
-    }
-    zip.finish()
-}
-
-fn main() {
-    println!("cargo:rerun-if-changed={TYPESHED_SOURCE_DIR}");
-    assert!(
-        Path::new(TYPESHED_SOURCE_DIR).is_dir(),
-        "Where is typeshed?"
-    );
-    let out_dir = std::env::var("OUT_DIR").unwrap();
-
-    // N.B. Deliberately using `format!()` instead of `Path::join()` here,
-    // so that we use `/` as a path separator on all platforms.
-    // That enables us to load the typeshed zip at compile time in `module.rs`
-    // (otherwise we'd have to dynamically determine the exact path to the typeshed zip
-    // based on the default path separator for the specific platform we're on,
-    // which can't be done at compile time.)
-    let zipped_typeshed_location = format!("{out_dir}{TYPESHED_ZIP_LOCATION}");
-
-    let zipped_typeshed = File::create(zipped_typeshed_location).unwrap();
-    zip_dir(TYPESHED_SOURCE_DIR, zipped_typeshed).unwrap();
+/// Rebuild the crate if a test file is added or removed from
+pub fn main() {
+    println!("cargo:rerun-if-changed=resources/mdtest");
 }

crates/red_knot_python_semantic/resources/README.md

@@ -0,0 +1,4 @@
+Markdown files within the `mdtest/` subdirectory are tests of type inference and type checking;
+executed by the `tests/mdtest.rs` integration test.
+
+See `crates/red_knot_test/README.md` for documentation of this test format.

crates/red_knot_python_semantic/resources/mdtest/comparison/byte_literals.md

@@ -0,0 +1,43 @@
+### Comparison: Byte literals
+
+These tests assert that we infer precise `Literal` types for comparisons between objects
+inferred as having `Literal` bytes types:
+
+```py
+reveal_type(b"abc" == b"abc")  # revealed: Literal[True]
+reveal_type(b"abc" == b"ab")   # revealed: Literal[False]
+
+reveal_type(b"abc" != b"abc")  # revealed: Literal[False]
+reveal_type(b"abc" != b"ab")   # revealed: Literal[True]
+
+reveal_type(b"abc" < b"abd")   # revealed: Literal[True]
+reveal_type(b"abc" < b"abb")   # revealed: Literal[False]
+
+reveal_type(b"abc" <= b"abc")  # revealed: Literal[True]
+reveal_type(b"abc" <= b"abb")  # revealed: Literal[False]
+
+reveal_type(b"abc" > b"abd")   # revealed: Literal[False]
+reveal_type(b"abc" > b"abb")   # revealed: Literal[True]
+
+reveal_type(b"abc" >= b"abc")  # revealed: Literal[True]
+reveal_type(b"abc" >= b"abd")  # revealed: Literal[False]
+
+reveal_type(b"" in b"")                      # revealed: Literal[True]
+reveal_type(b"" in b"abc")                   # revealed: Literal[True]
+reveal_type(b"abc" in b"")                   # revealed: Literal[False]
+reveal_type(b"ab" in b"abc")                 # revealed: Literal[True]
+reveal_type(b"abc" in b"abc")                # revealed: Literal[True]
+reveal_type(b"d" in b"abc")                  # revealed: Literal[False]
+reveal_type(b"ac" in b"abc")                 # revealed: Literal[False]
+reveal_type(b"\x81\x82" in b"\x80\x81\x82")  # revealed: Literal[True]
+reveal_type(b"\x82\x83" in b"\x80\x81\x82")  # revealed: Literal[False]
+
+reveal_type(b"ab" not in b"abc")             # revealed: Literal[False]
+reveal_type(b"ac" not in b"abc")             # revealed: Literal[True]
+
+reveal_type(b"abc" is b"abc")      # revealed: bool
+reveal_type(b"abc" is b"ab")       # revealed: Literal[False]
+
+reveal_type(b"abc" is not b"abc")  # revealed: bool
+reveal_type(b"abc" is not b"ab")   # revealed: Literal[True]
+```

crates/red_knot_python_semantic/resources/mdtest/narrow/conditionals_is.md

@@ -0,0 +1,29 @@
+# Narrowing for `is` conditionals
+
+## `is None`
+
+```py
+x = None if flag else 1
+
+if x is None:
+    # TODO the following should be simplified to 'None'
+    reveal_type(x)  # revealed: None | Literal[1] & None
+
+reveal_type(x)  # revealed: None | Literal[1]
+```
+
+## `is` for other types
+
+```py
+class A:
+    ...
+
+x = A()
+y = x if flag else None
+
+if y is x:
+    # TODO the following should be simplified to 'A'
+    reveal_type(y)  # revealed: A | None & A
+
+reveal_type(y)  # revealed: A | None
+```

crates/red_knot_python_semantic/resources/mdtest/narrow/conditionals_is_not.md

@@ -0,0 +1,40 @@
+# Narrowing for `is not` conditionals
+
+## `is not None`
+
+The type guard removes `None` from the union type:
+
+```py
+x = None if flag else 1
+
+if x is not None:
+    reveal_type(x)  # revealed: Literal[1]
+
+reveal_type(x)  # revealed: None | Literal[1]
+```
+
+## `is not` for other singleton types
+
+```py
+x = True if flag else False
+reveal_type(x)  # revealed: bool
+
+if x is not False:
+    # TODO the following should be `Literal[True]`
+    reveal_type(x)  # revealed: bool & ~Literal[False]
+```
+
+## `is not` for non-singleton types
+
+Non-singleton types should *not* narrow the type: two instances of a
+non-singleton class may occupy different addresses in memory even if
+they compare equal.
+
+```py
+x = [1]
+y = [1]
+
+if x is not y:
+    # TODO: should include type parameter: list[int]
+    reveal_type(x)  # revealed: list
+```

crates/red_knot_python_semantic/resources/mdtest/narrow/match.md

@@ -0,0 +1,17 @@
+# Narrowing for `match` statements
+
+## Single `match` pattern
+
+```py
+x = None if flag else 1
+reveal_type(x)  # revealed: None | Literal[1]
+
+y = 0
+
+match x:
+    case None:
+        y = x
+
+# TODO intersection simplification: should be just Literal[0] | None
+reveal_type(y)  # revealed: Literal[0] | None | Literal[1] & None
+```

crates/red_knot_python_semantic/resources/mdtest/numbers.md

@@ -0,0 +1,35 @@
+# Numbers
+
+## Integers
+
+### Literals
+
+We can infer an integer literal type:
+
+```py
+reveal_type(1)  # revealed: Literal[1]
+```
+
+### Overflow
+
+We only track integer literals within the range of an i64:
+
+```py
+reveal_type(9223372036854775808)  # revealed: int
+```
+
+## Floats
+
+There aren't literal float types, but we infer the general float type:
+
+```py
+reveal_type(1.0)  # revealed: float
+```
+
+## Complex
+
+Same for complex:
+
+```py
+reveal_type(2j)  # revealed: complex
+```

crates/red_knot_python_semantic/src/db.rs

@@ -11,7 +11,6 @@ pub trait Db: SourceDb + Upcast<dyn SourceDb> {
 pub(crate) mod tests {
     use std::sync::Arc;
 
-    use crate::module_resolver::vendored_typeshed_stubs;
     use ruff_db::files::{File, Files};
     use ruff_db::system::{DbWithTestSystem, System, TestSystem};
     use ruff_db::vendored::VendoredFileSystem;
@@ -33,7 +32,7 @@ pub(crate) mod tests {
             Self {
                 storage: salsa::Storage::default(),
                 system: TestSystem::default(),
-                vendored: vendored_typeshed_stubs().clone(),
+                vendored: red_knot_vendored::file_system().clone(),
                 events: std::sync::Arc::default(),
                 files: Files::default(),
             }

crates/red_knot_python_semantic/src/lib.rs

@@ -4,7 +4,7 @@ use rustc_hash::FxHasher;
 
 pub use db::Db;
 pub use module_name::ModuleName;
-pub use module_resolver::{resolve_module, system_module_search_paths, vendored_typeshed_stubs};
+pub use module_resolver::{resolve_module, system_module_search_paths, Module};
 pub use program::{Program, ProgramSettings, SearchPathSettings, SitePackages};
 pub use python_version::PythonVersion;
 pub use semantic_model::{HasTy, SemanticModel};

crates/red_knot_python_semantic/src/module_resolver/mod.rs

@@ -1,10 +1,9 @@
 use std::iter::FusedIterator;
 
-pub(crate) use module::Module;
+pub use module::Module;
 pub use resolver::resolve_module;
 pub(crate) use resolver::{file_to_module, SearchPaths};
 use ruff_db::system::SystemPath;
-pub use typeshed::vendored_typeshed_stubs;
 
 use crate::module_resolver::resolver::search_paths;
 use crate::Db;

crates/red_knot_python_semantic/src/module_resolver/resolver.rs

@@ -1,20 +1,21 @@
-use rustc_hash::{FxBuildHasher, FxHashSet};
 use std::borrow::Cow;
 use std::iter::FusedIterator;
-use std::ops::Deref;
+
+use rustc_hash::{FxBuildHasher, FxHashSet};
 
 use ruff_db::files::{File, FilePath, FileRootKind};
 use ruff_db::system::{DirectoryEntry, System, SystemPath, SystemPathBuf};
 use ruff_db::vendored::{VendoredFileSystem, VendoredPath};
 
-use super::module::{Module, ModuleKind};
-use super::path::{ModulePath, SearchPath, SearchPathValidationError};
 use crate::db::Db;
 use crate::module_name::ModuleName;
 use crate::module_resolver::typeshed::{vendored_typeshed_versions, TypeshedVersions};
 use crate::site_packages::VirtualEnvironment;
 use crate::{Program, PythonVersion, SearchPathSettings, SitePackages};
 
+use super::module::{Module, ModuleKind};
+use super::path::{ModulePath, SearchPath, SearchPathValidationError};
+
 /// Resolves a module name to a module.
 pub fn resolve_module(db: &dyn Db, module_name: ModuleName) -> Option<Module> {
     let interned_name = ModuleNameIngredient::new(db, module_name);
@@ -35,14 +36,14 @@ pub(crate) fn resolve_module_query<'db>(
     let _span = tracing::trace_span!("resolve_module", %name).entered();
 
     let Some((search_path, module_file, kind)) = resolve_name(db, name) else {
-        tracing::debug!("Module '{name}' not found in the search paths.");
+        tracing::debug!("Module `{name}` not found in search paths");
         return None;
     };
 
     let module = Module::new(name.clone(), kind, search_path, module_file);
 
     tracing::trace!(
-        "Resolved module '{name}' to '{path}'.",
+        "Resolved module `{name}` to `{path}`",
         path = module_file.path(db)
     );
 
@@ -136,7 +137,7 @@ pub(crate) struct SearchPaths {
     /// for the first `site-packages` path
     site_packages: Vec<SearchPath>,
 
-    typeshed_versions: ResolvedTypeshedVersions,
+    typeshed_versions: TypeshedVersions,
 }
 
 impl SearchPaths {
@@ -202,11 +203,11 @@ impl SearchPaths {
 
             let search_path = SearchPath::custom_stdlib(db, &custom_typeshed)?;
 
-            (ResolvedTypeshedVersions::Custom(parsed), search_path)
+            (parsed, search_path)
         } else {
             tracing::debug!("Using vendored stdlib");
             (
-                ResolvedTypeshedVersions::Vendored(vendored_typeshed_versions()),
+                vendored_typeshed_versions(db),
                 SearchPath::vendored_stdlib(),
             )
         };
@@ -279,23 +280,6 @@ impl SearchPaths {
     }
 }
 
-#[derive(Debug, PartialEq, Eq)]
-enum ResolvedTypeshedVersions {
-    Vendored(&'static TypeshedVersions),
-    Custom(TypeshedVersions),
-}
-
-impl Deref for ResolvedTypeshedVersions {
-    type Target = TypeshedVersions;
-
-    fn deref(&self) -> &Self::Target {
-        match self {
-            ResolvedTypeshedVersions::Vendored(versions) => versions,
-            ResolvedTypeshedVersions::Custom(versions) => versions,
-        }
-    }
-}
-
 /// Collect all dynamic search paths. For each `site-packages` path:
 /// - Collect that `site-packages` path
 /// - Collect any search paths listed in `.pth` files in that `site-packages` directory
@@ -340,7 +324,7 @@ pub(crate) fn dynamic_resolution_paths(db: &dyn Db) -> Vec<SearchPath> {
 
         let site_packages_root = files
             .root(db.upcast(), site_packages_dir)
-            .expect("Site-package root to have been created.");
+            .expect("Site-package root to have been created");
 
         // This query needs to be re-executed each time a `.pth` file
         // is added, modified or removed from the `site-packages` directory.

crates/red_knot_python_semantic/src/module_resolver/typeshed.rs

@@ -4,25 +4,19 @@ use std::num::{NonZeroU16, NonZeroUsize};
 use std::ops::{RangeFrom, RangeInclusive};
 use std::str::FromStr;
 
-use once_cell::sync::Lazy;
 use rustc_hash::FxHashMap;
 
-use super::vendored::vendored_typeshed_stubs;
 use crate::db::Db;
 use crate::module_name::ModuleName;
 use crate::{Program, PythonVersion};
 
-static VENDORED_VERSIONS: Lazy<TypeshedVersions> = Lazy::new(|| {
+pub(in crate::module_resolver) fn vendored_typeshed_versions(db: &dyn Db) -> TypeshedVersions {
     TypeshedVersions::from_str(
-        &vendored_typeshed_stubs()
+        &db.vendored()
             .read_to_string("stdlib/VERSIONS")
-            .unwrap(),
+            .expect("The vendored typeshed stubs should contain a VERSIONS file"),
     )
-    .unwrap()
-});
-
-pub(crate) fn vendored_typeshed_versions() -> &'static TypeshedVersions {
-    &VENDORED_VERSIONS
+    .expect("The VERSIONS file in the vendored typeshed stubs should be well-formed")
 }
 
 pub(crate) fn typeshed_versions(db: &dyn Db) -> &TypeshedVersions {
@@ -332,6 +326,8 @@ mod tests {
 
     use insta::assert_snapshot;
 
+    use crate::db::tests::TestDb;
+
     use super::*;
 
     const TYPESHED_STDLIB_DIR: &str = "stdlib";
@@ -353,12 +349,9 @@ mod tests {
 
     #[test]
     fn can_parse_vendored_versions_file() {
-        let versions_data = include_str!(concat!(
-            env!("CARGO_MANIFEST_DIR"),
-            "/vendor/typeshed/stdlib/VERSIONS"
-        ));
+        let db = TestDb::new();
 
-        let versions = TypeshedVersions::from_str(versions_data).unwrap();
+        let versions = vendored_typeshed_versions(&db);
         assert!(versions.len() > 100);
         assert!(versions.len() < 1000);
 
@@ -395,9 +388,10 @@ mod tests {
 
     #[test]
     fn typeshed_versions_consistent_with_vendored_stubs() {
-        const VERSIONS_DATA: &str = include_str!("../../../vendor/typeshed/stdlib/VERSIONS");
-        let vendored_typeshed_dir = Path::new("vendor/typeshed").canonicalize().unwrap();
-        let vendored_typeshed_versions = TypeshedVersions::from_str(VERSIONS_DATA).unwrap();
+        let db = TestDb::new();
+        let vendored_typeshed_versions = vendored_typeshed_versions(&db);
+        let vendored_typeshed_dir =
+            Path::new(env!("CARGO_MANIFEST_DIR")).join("../red_knot_vendored/vendor/typeshed");
 
         let mut empty_iterator = true;
 

crates/red_knot_python_semantic/src/module_resolver/typeshed/mod.rs

@@ -1,8 +0,0 @@
-pub use self::vendored::vendored_typeshed_stubs;
-pub(super) use self::versions::{
-    typeshed_versions, vendored_typeshed_versions, TypeshedVersions, TypeshedVersionsParseError,
-    TypeshedVersionsQueryResult,
-};
-
-mod vendored;
-mod versions;

crates/red_knot_python_semantic/src/python_version.rs

@@ -54,6 +54,13 @@ impl TryFrom<(&str, &str)> for PythonVersion {
     }
 }
 
+impl From<(u8, u8)> for PythonVersion {
+    fn from(value: (u8, u8)) -> Self {
+        let (major, minor) = value;
+        Self { major, minor }
+    }
+}
+
 impl fmt::Display for PythonVersion {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         let PythonVersion { major, minor } = self;

crates/red_knot_python_semantic/src/semantic_index.rs

@@ -1,7 +1,7 @@
 use std::iter::FusedIterator;
 use std::sync::Arc;
 
-use rustc_hash::FxHashMap;
+use rustc_hash::{FxBuildHasher, FxHashMap};
 use salsa::plumbing::AsId;
 
 use ruff_db::files::File;
@@ -31,7 +31,7 @@ pub(crate) use self::use_def::{
     BindingWithConstraints, BindingWithConstraintsIterator, DeclarationsIterator,
 };
 
-type SymbolMap = hashbrown::HashMap<ScopedSymbolId, (), ()>;
+type SymbolMap = hashbrown::HashMap<ScopedSymbolId, (), FxBuildHasher>;
 
 /// Returns the semantic index for `file`.
 ///
@@ -115,6 +115,9 @@ pub(crate) struct SemanticIndex<'db> {
     /// Note: We should not depend on this map when analysing other files or
     /// changing a file invalidates all dependents.
     ast_ids: IndexVec<FileScopeId, AstIds>,
+
+    /// Flags about the global scope (code usage impacting inference)
+    has_future_annotations: bool,
 }
 
 impl<'db> SemanticIndex<'db> {
@@ -215,6 +218,12 @@ impl<'db> SemanticIndex<'db> {
     pub(crate) fn node_scope(&self, node: NodeWithScopeRef) -> FileScopeId {
         self.scopes_by_node[&node.node_key()]
     }
+
+    /// Checks if there is an import of `__future__.annotations` in the global scope, which affects
+    /// the logic for type inference.
+    pub(super) fn has_future_annotations(&self) -> bool {
+        self.has_future_annotations
+    }
 }
 
 pub struct AncestorsIter<'a> {
@@ -1027,7 +1036,7 @@ class C[T]:
         }
 
         let TestCase { db, file } = test_case(
-            r#"
+            r"
 class Test:
     def foo():
         def bar():
@@ -1036,7 +1045,7 @@ class Test:
         pass
 
 def x():
-    pass"#,
+    pass",
         );
 
         let index = semantic_index(&db, file);

crates/red_knot_python_semantic/src/semantic_index/builder.rs

@@ -45,6 +45,9 @@ pub(super) struct SemanticIndexBuilder<'db> {
     /// Flow states at each `break` in the current loop.
     loop_break_states: Vec<FlowSnapshot>,
 
+    /// Flags about the file's global scope
+    has_future_annotations: bool,
+
     // Semantic Index fields
     scopes: IndexVec<FileScopeId, Scope>,
     scope_ids_by_scope: IndexVec<FileScopeId, ScopeId<'db>>,
@@ -68,6 +71,8 @@ impl<'db> SemanticIndexBuilder<'db> {
             current_match_case: None,
             loop_break_states: vec![],
 
+            has_future_annotations: false,
+
             scopes: IndexVec::new(),
             symbol_tables: IndexVec::new(),
             ast_ids: IndexVec::new(),
@@ -450,6 +455,7 @@ impl<'db> SemanticIndexBuilder<'db> {
             scopes_by_expression: self.scopes_by_expression,
             scopes_by_node: self.scopes_by_node,
             use_def_maps,
+            has_future_annotations: self.has_future_annotations,
         }
     }
 }
@@ -543,7 +549,16 @@ where
                         &alias.name.id
                     };
 
+                    // Look for imports `from __future__ import annotations`, ignore `as ...`
+                    // We intentionally don't enforce the rules about location of `__future__`
+                    // imports here, we assume the user's intent was to apply the `__future__`
+                    // import, so we still check using it (and will also emit a diagnostic about a
+                    // miss-placed `__future__` import.)
+                    self.has_future_annotations |= alias.name.id == "annotations"
+                        && node.module.as_deref() == Some("__future__");
+
                     let symbol = self.add_symbol(symbol_name.clone());
+
                     self.add_definition(symbol, ImportFromDefinitionNodeRef { node, alias_index });
                 }
             }

crates/red_knot_python_semantic/src/semantic_index/definition.rs

@@ -3,6 +3,7 @@ use ruff_db::parsed::ParsedModule;
 use ruff_python_ast as ast;
 
 use crate::ast_node_ref::AstNodeRef;
+use crate::module_resolver::file_to_module;
 use crate::node_key::NodeKey;
 use crate::semantic_index::symbol::{FileScopeId, ScopeId, ScopedSymbolId};
 use crate::Db;
@@ -45,6 +46,14 @@ impl<'db> Definition<'db> {
     pub(crate) fn is_binding(self, db: &'db dyn Db) -> bool {
         self.kind(db).category().is_binding()
     }
+
+    /// Return true if this is a symbol was defined in the `typing` or `typing_extensions` modules
+    pub(crate) fn is_typing_definition(self, db: &'db dyn Db) -> bool {
+        file_to_module(db, self.file(db)).is_some_and(|module| {
+            module.search_path().is_standard_library()
+                && matches!(&**module.name(), "typing" | "typing_extensions")
+        })
+    }
 }
 
 #[derive(Copy, Clone, Debug)]
@@ -330,7 +339,7 @@ impl DefinitionCategory {
     /// If so, any assignments reached by this definition are in error if they assign a value of a
     /// type not assignable to the declared type.
     ///
-    /// Annotations establish a declared type. So do function and class definition.
+    /// Annotations establish a declared type. So do function and class definitions, and imports.
     pub(crate) fn is_declaration(self) -> bool {
         matches!(
             self,
@@ -371,10 +380,11 @@ pub enum DefinitionKind {
 impl DefinitionKind {
     pub(crate) fn category(&self) -> DefinitionCategory {
         match self {
-            // functions and classes always bind a value, and we always consider them declarations
-            DefinitionKind::Function(_) | DefinitionKind::Class(_) => {
-                DefinitionCategory::DeclarationAndBinding
-            }
+            // functions, classes, and imports always bind, and we consider them declarations
+            DefinitionKind::Function(_)
+            | DefinitionKind::Class(_)
+            | DefinitionKind::Import(_)
+            | DefinitionKind::ImportFrom(_) => DefinitionCategory::DeclarationAndBinding,
             // a parameter always binds a value, but is only a declaration if annotated
             DefinitionKind::Parameter(parameter) => {
                 if parameter.annotation.is_some() {
@@ -400,9 +410,7 @@ impl DefinitionKind {
                 }
             }
             // all of these bind values without declaring a type
-            DefinitionKind::Import(_)
-            | DefinitionKind::ImportFrom(_)
-            | DefinitionKind::NamedExpression(_)
+            DefinitionKind::NamedExpression(_)
             | DefinitionKind::Assignment(_)
             | DefinitionKind::AugmentedAssignment(_)
             | DefinitionKind::For(_)

crates/red_knot_python_semantic/src/site_packages.rs

@@ -192,7 +192,7 @@ impl VirtualEnvironment {
             } else {
                 tracing::warn!(
                     "Failed to resolve `sys.prefix` of the system Python installation \
-from the `home` value in the `pyvenv.cfg` file at '{}'. \
+from the `home` value in the `pyvenv.cfg` file at `{}`. \
 System site-packages will not be used for module resolution.",
                     venv_path.join("pyvenv.cfg")
                 );
@@ -426,7 +426,7 @@ impl Deref for SysPrefixPath {
 
 impl fmt::Display for SysPrefixPath {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        write!(f, "`sys.prefix` path '{}'", self.0)
+        write!(f, "`sys.prefix` path `{}`", self.0)
     }
 }
 
@@ -483,7 +483,7 @@ impl Deref for PythonHomePath {
 
 impl fmt::Display for PythonHomePath {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        write!(f, "`home` location '{}'", self.0)
+        write!(f, "`home` location `{}`", self.0)
     }
 }
 

crates/red_knot_python_semantic/src/stdlib.rs

@@ -11,6 +11,7 @@ enum CoreStdlibModule {
     Builtins,
     Types,
     Typeshed,
+    TypingExtensions,
 }
 
 impl CoreStdlibModule {
@@ -19,6 +20,7 @@ impl CoreStdlibModule {
             Self::Builtins => "builtins",
             Self::Types => "types",
             Self::Typeshed => "_typeshed",
+            Self::TypingExtensions => "typing_extensions",
         };
         ModuleName::new_static(module_name)
             .unwrap_or_else(|| panic!("{module_name} should be a valid module name!"))
@@ -62,6 +64,14 @@ pub(crate) fn typeshed_symbol_ty<'db>(db: &'db dyn Db, symbol: &str) -> Type<'db
     core_module_symbol_ty(db, CoreStdlibModule::Typeshed, symbol)
 }
 
+/// Lookup the type of `symbol` in the `typing_extensions` module namespace.
+///
+/// Returns `Unbound` if the `typing_extensions` module isn't available for some reason.
+#[inline]
+pub(crate) fn typing_extensions_symbol_ty<'db>(db: &'db dyn Db, symbol: &str) -> Type<'db> {
+    core_module_symbol_ty(db, CoreStdlibModule::TypingExtensions, symbol)
+}
+
 /// Get the scope of a core stdlib module.
 ///
 /// Can return `None` if a custom typeshed is used that is missing the core module in question.

crates/red_knot_python_semantic/src/types/builder.rs

@@ -25,12 +25,14 @@
 //!   * No type in an intersection can be a supertype of any other type in the intersection (just
 //!     eliminate the supertype from the intersection).
 //!   * An intersection containing two non-overlapping types should simplify to [`Type::Never`].
-use crate::types::{builtins_symbol_ty, IntersectionType, Type, UnionType};
+use crate::types::{IntersectionType, Type, UnionType};
 use crate::{Db, FxOrderSet};
-use ordermap::set::MutableValues;
+use smallvec::SmallVec;
+
+use super::KnownClass;
 
 pub(crate) struct UnionBuilder<'db> {
-    elements: FxOrderSet<Type<'db>>,
+    elements: Vec<Type<'db>>,
     db: &'db dyn Db,
 }
 
@@ -38,7 +40,7 @@ impl<'db> UnionBuilder<'db> {
     pub(crate) fn new(db: &'db dyn Db) -> Self {
         Self {
             db,
-            elements: FxOrderSet::default(),
+            elements: vec![],
         }
     }
 
@@ -46,47 +48,70 @@ impl<'db> UnionBuilder<'db> {
     pub(crate) fn add(mut self, ty: Type<'db>) -> Self {
         match ty {
             Type::Union(union) => {
-                self.elements.extend(union.elements(self.db));
+                let new_elements = union.elements(self.db);
+                self.elements.reserve(new_elements.len());
+                for element in new_elements {
+                    self = self.add(*element);
+                }
             }
             Type::Never => {}
             _ => {
-                self.elements.insert(ty);
+                let bool_pair = if let Type::BooleanLiteral(b) = ty {
+                    Some(Type::BooleanLiteral(!b))
+                } else {
+                    None
+                };
+
+                let mut to_add = ty;
+                let mut to_remove = SmallVec::<[usize; 2]>::new();
+                for (index, element) in self.elements.iter().enumerate() {
+                    if Some(*element) == bool_pair {
+                        to_add = KnownClass::Bool.to_instance(self.db);
+                        to_remove.push(index);
+                        // The type we are adding is a BooleanLiteral, which doesn't have any
+                        // subtypes. And we just found that the union already contained our
+                        // mirror-image BooleanLiteral, so it can't also contain bool or any
+                        // supertype of bool. Therefore, we are done.
+                        break;
+                    }
+                    if ty.is_subtype_of(self.db, *element) {
+                        return self;
+                    } else if element.is_subtype_of(self.db, ty) {
+                        to_remove.push(index);
+                    }
+                }
+
+                match to_remove[..] {
+                    [] => self.elements.push(to_add),
+                    [index] => self.elements[index] = to_add,
+                    _ => {
+                        let mut current_index = 0;
+                        let mut to_remove = to_remove.into_iter();
+                        let mut next_to_remove_index = to_remove.next();
+                        self.elements.retain(|_| {
+                            let retain = if Some(current_index) == next_to_remove_index {
+                                next_to_remove_index = to_remove.next();
+                                false
+                            } else {
+                                true
+                            };
+                            current_index += 1;
+                            retain
+                        });
+                        self.elements.push(to_add);
+                    }
+                }
             }
         }
 
         self
     }
 
-    /// Performs the following normalizations:
-    ///     - Replaces `Literal[True,False]` with `bool`.
-    ///     - TODO For enums `E` with members `X1`,...,`Xn`, replaces
-    ///     `Literal[E.X1,...,E.Xn]` with `E`.
-    fn simplify(&mut self) {
-        if let Some(true_index) = self.elements.get_index_of(&Type::BooleanLiteral(true)) {
-            if self.elements.contains(&Type::BooleanLiteral(false)) {
-                *self.elements.get_index_mut2(true_index).unwrap() =
-                    builtins_symbol_ty(self.db, "bool");
-                self.elements.remove(&Type::BooleanLiteral(false));
-            }
-        }
-    }
-
-    pub(crate) fn build(mut self) -> Type<'db> {
+    pub(crate) fn build(self) -> Type<'db> {
         match self.elements.len() {
             0 => Type::Never,
             1 => self.elements[0],
-            _ => {
-                self.simplify();
-
-                match self.elements.len() {
-                    0 => Type::Never,
-                    1 => self.elements[0],
-                    _ => {
-                        self.elements.shrink_to_fit();
-                        Type::Union(UnionType::new(self.db, self.elements))
-                    }
-                }
-            }
+            _ => Type::Union(UnionType::new(self.db, self.elements.into())),
         }
     }
 }
@@ -192,6 +217,7 @@ impl<'db> InnerIntersectionBuilder<'db> {
 
     /// Adds a positive type to this intersection.
     fn add_positive(&mut self, db: &'db dyn Db, ty: Type<'db>) {
+        // TODO `Any`/`Unknown`/`Todo` actually should not self-cancel
         match ty {
             Type::Intersection(inter) => {
                 let pos = inter.positive(db);
@@ -211,7 +237,7 @@ impl<'db> InnerIntersectionBuilder<'db> {
 
     /// Adds a negative type to this intersection.
     fn add_negative(&mut self, db: &'db dyn Db, ty: Type<'db>) {
-        // TODO Any/Unknown actually should not self-cancel
+        // TODO `Any`/`Unknown`/`Todo` actually should not self-cancel
         match ty {
             Type::Intersection(intersection) => {
                 let pos = intersection.negative(db);
@@ -276,16 +302,10 @@ mod tests {
     use crate::db::tests::TestDb;
     use crate::program::{Program, SearchPathSettings};
     use crate::python_version::PythonVersion;
-    use crate::types::{builtins_symbol_ty, UnionBuilder};
+    use crate::types::{KnownClass, UnionBuilder};
     use crate::ProgramSettings;
     use ruff_db::system::{DbWithTestSystem, SystemPathBuf};
 
-    impl<'db> UnionType<'db> {
-        fn elements_vec(self, db: &'db TestDb) -> Vec<Type<'db>> {
-            self.elements(db).into_iter().copied().collect()
-        }
-    }
-
     fn setup_db() -> TestDb {
         let db = TestDb::new();
 
@@ -313,7 +333,7 @@ mod tests {
         let t1 = Type::IntLiteral(1);
         let union = UnionType::from_elements(&db, [t0, t1]).expect_union();
 
-        assert_eq!(union.elements_vec(&db), &[t0, t1]);
+        assert_eq!(union.elements(&db), &[t0, t1]);
     }
 
     #[test]
@@ -342,7 +362,7 @@ mod tests {
     #[test]
     fn build_union_bool() {
         let db = setup_db();
-        let bool_ty = builtins_symbol_ty(&db, "bool");
+        let bool_instance_ty = KnownClass::Bool.to_instance(&db);
 
         let t0 = Type::BooleanLiteral(true);
         let t1 = Type::BooleanLiteral(true);
@@ -350,10 +370,10 @@ mod tests {
         let t3 = Type::IntLiteral(17);
 
         let union = UnionType::from_elements(&db, [t0, t1, t3]).expect_union();
-        assert_eq!(union.elements_vec(&db), &[t0, t3]);
+        assert_eq!(union.elements(&db), &[t0, t3]);
 
         let union = UnionType::from_elements(&db, [t0, t1, t2, t3]).expect_union();
-        assert_eq!(union.elements_vec(&db), &[bool_ty, t3]);
+        assert_eq!(union.elements(&db), &[bool_instance_ty, t3]);
     }
 
     #[test]
@@ -365,7 +385,44 @@ mod tests {
         let u1 = UnionType::from_elements(&db, [t0, t1]);
         let union = UnionType::from_elements(&db, [u1, t2]).expect_union();
 
-        assert_eq!(union.elements_vec(&db), &[t0, t1, t2]);
+        assert_eq!(union.elements(&db), &[t0, t1, t2]);
+    }
+
+    #[test]
+    fn build_union_simplify_subtype() {
+        let db = setup_db();
+        let t0 = KnownClass::Str.to_instance(&db);
+        let t1 = Type::LiteralString;
+        let u0 = UnionType::from_elements(&db, [t0, t1]);
+        let u1 = UnionType::from_elements(&db, [t1, t0]);
+
+        assert_eq!(u0, t0);
+        assert_eq!(u1, t0);
+    }
+
+    #[test]
+    fn build_union_no_simplify_unknown() {
+        let db = setup_db();
+        let t0 = KnownClass::Str.to_instance(&db);
+        let t1 = Type::Unknown;
+        let u0 = UnionType::from_elements(&db, [t0, t1]);
+        let u1 = UnionType::from_elements(&db, [t1, t0]);
+
+        assert_eq!(u0.expect_union().elements(&db), &[t0, t1]);
+        assert_eq!(u1.expect_union().elements(&db), &[t1, t0]);
+    }
+
+    #[test]
+    fn build_union_subsume_multiple() {
+        let db = setup_db();
+        let str_ty = KnownClass::Str.to_instance(&db);
+        let int_ty = KnownClass::Int.to_instance(&db);
+        let object_ty = KnownClass::Object.to_instance(&db);
+        let unknown_ty = Type::Unknown;
+
+        let u0 = UnionType::from_elements(&db, [str_ty, unknown_ty, int_ty, object_ty]);
+
+        assert_eq!(u0.expect_union().elements(&db), &[unknown_ty, object_ty]);
     }
 
     impl<'db> IntersectionType<'db> {
@@ -446,7 +503,7 @@ mod tests {
             .add_positive(u0)
             .build()
             .expect_union();
-        let [Type::Intersection(i0), Type::Intersection(i1)] = union.elements_vec(&db)[..] else {
+        let [Type::Intersection(i0), Type::Intersection(i1)] = union.elements(&db)[..] else {
             panic!("expected a union of two intersections");
         };
         assert_eq!(i0.pos_vec(&db), &[ta, t0]);

crates/red_knot_python_semantic/src/types/display.rs

@@ -37,7 +37,7 @@ impl Display for DisplayType<'_> {
                 | Type::Class(_)
                 | Type::Function(_)
         ) {
-            write!(f, "Literal[{representation}]",)
+            write!(f, "Literal[{representation}]")
         } else {
             representation.fmt(f)
         }
@@ -66,6 +66,9 @@ impl Display for DisplayRepresentation<'_> {
             Type::Unknown => f.write_str("Unknown"),
             Type::Unbound => f.write_str("Unbound"),
             Type::None => f.write_str("None"),
+            // `[Type::Todo]`'s display should be explicit that is not a valid display of
+            // any other type
+            Type::Todo => f.write_str("@Todo"),
             Type::Module(file) => {
                 write!(f, "<module '{:?}'>", file.path(self.db))
             }
@@ -332,7 +335,7 @@ mod tests {
             class B: ...
             ",
         )?;
-        let mod_file = system_path_to_file(&db, "src/main.py").expect("Expected file to exist.");
+        let mod_file = system_path_to_file(&db, "src/main.py").expect("file to exist");
 
         let union_elements = &[
             Type::Unknown,

crates/red_knot_python_semantic/src/types/narrow.rs

@@ -155,13 +155,24 @@ impl<'db> NarrowingConstraintsBuilder<'db> {
             let inference = infer_expression_types(self.db, expression);
             for (op, comparator) in std::iter::zip(&**ops, &**comparators) {
                 let comp_ty = inference.expression_ty(comparator.scoped_ast_id(self.db, scope));
-                if matches!(op, ast::CmpOp::IsNot) {
-                    let ty = IntersectionBuilder::new(self.db)
-                        .add_negative(comp_ty)
-                        .build();
-                    self.constraints.insert(symbol, ty);
-                };
-                // TODO other comparison types
+                match op {
+                    ast::CmpOp::IsNot => {
+                        if comp_ty.is_singleton(self.db) {
+                            let ty = IntersectionBuilder::new(self.db)
+                                .add_negative(comp_ty)
+                                .build();
+                            self.constraints.insert(symbol, ty);
+                        } else {
+                            // Non-singletons cannot be safely narrowed using `is not`
+                        }
+                    }
+                    ast::CmpOp::Is => {
+                        self.constraints.insert(symbol, comp_ty);
+                    }
+                    _ => {
+                        // TODO other comparison types
+                    }
+                }
             }
         }
     }

crates/red_knot_python_semantic/tests/mdtest.rs

@@ -0,0 +1,14 @@
+use red_knot_test::run;
+use std::path::PathBuf;
+
+/// See `crates/red_knot_test/README.md` for documentation on these tests.
+#[rstest::rstest]
+fn mdtest(#[files("resources/mdtest/**/*.md")] path: PathBuf) {
+    let crate_dir = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+        .join("resources")
+        .join("mdtest")
+        .canonicalize()
+        .unwrap();
+    let title = path.strip_prefix(crate_dir).unwrap();
+    run(&path, title.as_os_str().to_str().unwrap());
+}

crates/red_knot_python_semantic/vendor/typeshed/source_commit.txt

@@ -1 +0,0 @@
-9e506eb5e8fc2823db8c60ad561b1145ff114947

crates/red_knot_python_semantic/vendor/typeshed/stdlib/importlib/resources/__init__.pyi

@@ -1,46 +0,0 @@
-import os
-import sys
-from collections.abc import Iterator
-from contextlib import AbstractContextManager
-from pathlib import Path
-from types import ModuleType
-from typing import Any, BinaryIO, TextIO
-from typing_extensions import TypeAlias
-
-if sys.version_info >= (3, 9):
-    from importlib.abc import Traversable
-
-__all__ = ["Package", "Resource", "contents", "is_resource", "open_binary", "open_text", "path", "read_binary", "read_text"]
-
-if sys.version_info >= (3, 9):
-    __all__ += ["as_file", "files"]
-
-if sys.version_info >= (3, 10):
-    __all__ += ["ResourceReader"]
-
-Package: TypeAlias = str | ModuleType
-
-if sys.version_info >= (3, 11):
-    Resource: TypeAlias = str
-else:
-    Resource: TypeAlias = str | os.PathLike[Any]
-
-def open_binary(package: Package, resource: Resource) -> BinaryIO: ...
-def open_text(package: Package, resource: Resource, encoding: str = "utf-8", errors: str = "strict") -> TextIO: ...
-def read_binary(package: Package, resource: Resource) -> bytes: ...
-def read_text(package: Package, resource: Resource, encoding: str = "utf-8", errors: str = "strict") -> str: ...
-def path(package: Package, resource: Resource) -> AbstractContextManager[Path]: ...
-def is_resource(package: Package, name: str) -> bool: ...
-def contents(package: Package) -> Iterator[str]: ...
-
-if sys.version_info >= (3, 9):
-    def as_file(path: Traversable) -> AbstractContextManager[Path]: ...
-
-if sys.version_info >= (3, 12):
-    def files(anchor: Package | None = ...) -> Traversable: ...
-
-elif sys.version_info >= (3, 9):
-    def files(package: Package) -> Traversable: ...
-
-if sys.version_info >= (3, 10):
-    from importlib.abc import ResourceReader as ResourceReader

crates/red_knot_server/src/server/api/requests/diagnostic.rs

@@ -47,7 +47,7 @@ impl BackgroundDocumentRequestHandler for DocumentDiagnosticRequestHandler {
 fn compute_diagnostics(snapshot: &DocumentSnapshot, db: &RootDatabase) -> Vec<Diagnostic> {
     let Some(file) = snapshot.file(db) else {
         tracing::info!(
-            "No file found for snapshot for '{}'",
+            "No file found for snapshot for `{}`",
             snapshot.query().file_url()
         );
         return vec![];

crates/red_knot_test/Cargo.toml

@@ -0,0 +1,33 @@
+[package]
+name = "red_knot_test"
+version = "0.0.0"
+publish = false
+edition.workspace = true
+rust-version.workspace = true
+homepage.workspace = true
+documentation.workspace = true
+repository.workspace = true
+authors.workspace = true
+license.workspace = true
+
+[dependencies]
+red_knot_python_semantic = { workspace = true }
+red_knot_vendored = { workspace = true }
+ruff_db = { workspace = true }
+ruff_index = { workspace = true }
+ruff_python_trivia = { workspace = true }
+ruff_source_file = { workspace = true }
+ruff_text_size = { workspace = true }
+
+anyhow = { workspace = true }
+colored = { workspace = true }
+once_cell = { workspace = true }
+regex = { workspace = true }
+rustc-hash = { workspace = true }
+salsa = { workspace = true }
+smallvec = { workspace = true }
+
+[dev-dependencies]
+
+[lints]
+workspace = true

crates/red_knot_test/README.md

@@ -0,0 +1,470 @@
+# Writing type-checking / type-inference tests
+
+Any Markdown file can be a test suite.
+
+In order for it to be run as one, `red_knot_test::run` must be called with its path; see
+`crates/red_knot_python_semantic/tests/mdtest.rs` for an example that treats all Markdown files
+under a certain directory as test suites.
+
+A Markdown test suite can contain any number of tests. A test consists of one or more embedded
+"files", each defined by a triple-backticks fenced code block. The code block must have a tag string
+specifying its language; currently only `py` (Python files) and `pyi` (type stub files) are
+supported.
+
+The simplest possible test suite consists of just a single test, with a single embedded file:
+
+````markdown
+```py
+reveal_type(1)  # revealed: Literal[1]
+```
+````
+
+When running this test, the mdtest framework will write a file with these contents to the default
+file path (`/src/test.py`) in its in-memory file system, run a type check on that file, and then
+match the resulting diagnostics with the assertions in the test. Assertions are in the form of
+Python comments. If all diagnostics and all assertions are matched, the test passes; otherwise, it
+fails.
+
+<!---
+(If you are reading this document in raw Markdown source rather than rendered Markdown, note that
+the quadruple-backtick-fenced "markdown" language code block above is NOT itself part of the mdtest
+syntax, it's just how this README embeds an example mdtest Markdown document.)
+--->
+
+See actual example mdtest suites in
+[`crates/red_knot_python_semantic/resources/mdtest`](https://github.com/astral-sh/ruff/tree/main/crates/red_knot_python_semantic/resources/mdtest).
+
+> ℹ️ Note: If you use `rstest` to generate a separate test for all Markdown files in a certain directory,
+> as with the example in `crates/red_knot_python_semantic/tests/mdtest.rs`,
+> you will likely want to also make sure that the crate the tests are in is rebuilt every time a
+> Markdown file is added or removed from the directory. See
+> [`crates/red_knot_python_semantic/build.rs`](https://github.com/astral-sh/ruff/tree/main/crates/red_knot_python_semantic/build.rs)
+> for an example of how to do this.
+>
+> This is because `rstest` generates its tests at build time rather than at runtime.
+> Without the `build.rs` file to force a rebuild when a Markdown file is added or removed,
+> a new Markdown test suite might not be run unless some other change in the crate caused a rebuild
+> following the addition of the new test file.
+
+## Assertions
+
+Two kinds of assertions are supported: `# revealed:` (shown above) and `# error:`.
+
+### Assertion kinds
+
+#### revealed
+
+A `# revealed:` assertion should always be paired with a call to the `reveal_type` utility, which
+reveals (via a diagnostic) the inferred type of its argument (which can be any expression). The text
+after `# revealed:` must match exactly with the displayed form of the revealed type of that
+expression.
+
+The `reveal_type` function can be imported from the `typing` standard library module (or, for older
+Python versions, from the `typing_extensions` pseudo-standard-library module[^extensions]):
+
+```py
+from typing import reveal_type
+
+reveal_type("foo")  # revealed: Literal["foo"]
+```
+
+For convenience, type checkers also pretend that `reveal_type` is a built-in, so that this import is
+not required. Using `reveal_type` without importing it issues a diagnostic warning that it was used
+without importing it, in addition to the diagnostic revealing the type of the expression.
+
+The `# revealed:` assertion must always match a revealed-type diagnostic, and will also match the
+undefined-reveal diagnostic, if present, so it's safe to use `reveal_type` in tests either with or
+without importing it. (Style preference is to not import it in tests, unless specifically testing
+something about the behavior of importing it.)
+
+#### error
+
+A comment beginning with `# error:` is an assertion that a type checker diagnostic will be emitted,
+with text span starting on that line. The matching can be narrowed in three ways:
+
+- `# error: [invalid-assignment]` requires that the matched diagnostic have the rule code
+    `invalid-assignment`. (The square brackets are required.)
+- `# error: "Some text"` requires that the diagnostic's full message contain the text `Some text`.
+    (The double quotes are required in the assertion comment; they are not part of the matched text.)
+- `# error: 8 [rule-code]` or `# error: 8 "Some text"` additionally requires that the matched
+    diagnostic's text span begins on column 8 (one-indexed) of this line.
+
+Assertions must contain either a rule code or a contains-text, or both, and may optionally also
+include a column number. They must come in order: first column, if present; then rule code, if
+present; then contains-text, if present. For example, an assertion using all three would look like
+`# error: 8 [invalid-assignment] "Some text"`.
+
+Error assertions in tests intended to test type checker semantics should primarily use rule-code
+assertions, with occasional contains-text assertions where needed to disambiguate or validate some
+details of the diagnostic message.
+
+### Assertion locations
+
+An assertion comment may be a line-trailing comment, in which case it applies to the line it is on:
+
+```py
+x: str = 1  # error: [invalid-assignment]
+```
+
+Or it may be a comment on its own line, in which case it applies to the next line that does not
+contain an assertion comment:
+
+```py
+# error: [invalid-assignment]
+x: str = 1
+```
+
+Multiple assertions applying to the same line may be stacked:
+
+```py
+# error: [invalid-assignment]
+# revealed: Literal[1]
+x: str = reveal_type(1)
+```
+
+Intervening empty lines or non-assertion comments are not allowed; an assertion stack must be one
+assertion per line, immediately following each other, with the line immediately following the last
+assertion as the line of source code on which the matched diagnostics are emitted.
+
+## Multi-file tests
+
+Some tests require multiple files, with imports from one file into another. Multiple fenced code
+blocks represent multiple embedded files. Since files must have unique names, at most one file can
+use the default name of `/src/test.py`. Other files must explicitly specify their file name:
+
+````markdown
+```py
+from b import C
+reveal_type(C)  # revealed: Literal[C]
+```
+
+```py path=b.py
+class C: pass
+```
+````
+
+Relative file names are always relative to the "workspace root", which is also an import root (that
+is, the equivalent of a runtime entry on `sys.path`).
+
+The default workspace root is `/src/`. Currently it is not possible to customize this in a test, but
+this is a feature we will want to add in the future.
+
+So the above test creates two files, `/src/test.py` and `/src/b.py`, and sets the workspace root to
+`/src/`, allowing `test.py` to import from `b.py` using the module name `b`.
+
+## Multi-test suites
+
+A single test suite (Markdown file) can contain multiple tests, by demarcating them using Markdown
+header lines:
+
+````markdown
+# Same-file invalid assignment
+
+```py
+x: int = "foo"  # error: [invalid-assignment]
+```
+
+# Cross-file invalid assignment
+
+```py
+from b import y
+x: int = y  # error: [invalid-assignment]
+```
+
+```py path=b.py
+y = "foo"
+```
+````
+
+This test suite contains two tests, one named "Same-file invalid assignment" and the other named
+"Cross-file invalid assignment". The first test involves only a single embedded file, and the second
+test involves two embedded files.
+
+The tests are run independently, in independent in-memory file systems and with new red-knot
+[Salsa](https://github.com/salsa-rs/salsa) databases. This means that each is a from-scratch run of
+the type checker, with no data persisting from any previous test.
+
+Due to `cargo test` limitations, an entire test suite (Markdown file) is run as a single Rust test,
+so it's not possible to select individual tests within it to run.
+
+## Structured test suites
+
+Markdown headers can also be used to group related tests within a suite:
+
+````markdown
+# Literals
+
+## Numbers
+
+### Integer
+
+```py
+reveal_type(1)  # revealed: Literal[1]
+```
+
+### Float
+
+```py
+reveal_type(1.0)  # revealed: float
+```
+
+## Strings
+
+```py
+reveal_type("foo")  # revealed: Literal["foo"]
+```
+````
+
+This test suite contains three tests, named "Literals - Numbers - Integer", "Literals - Numbers -
+Float", and "Literals - Strings".
+
+A header-demarcated section must either be a test or a grouping header; it cannot be both. That is,
+a header section can either contain embedded files (making it a test), or it can contain more
+deeply-nested headers (headers with more `#`), but it cannot contain both.
+
+## Documentation of tests
+
+Arbitrary Markdown syntax (including of course normal prose paragraphs) is permitted (and ignored by
+the test framework) between fenced code blocks. This permits natural documentation of
+why a test exists, and what it intends to assert:
+
+````markdown
+Assigning a string to a variable annotated as `int` is not permitted:
+
+```py
+x: int = "foo"  # error: [invalid-assignment]
+```
+````
+
+## Planned features
+
+There are some designed features that we intend for the test framework to have, but have not yet
+implemented:
+
+### Multi-line diagnostic assertions
+
+We may want to be able to assert that a diagnostic spans multiple lines, and to assert the columns it
+begins and/or ends on. The planned syntax for this will use `<<<` and `>>>` to mark the start and end lines for
+an assertion:
+
+```py
+(3  # error: 2 [unsupported-operands] <<<
+  +
+ "foo")  # error: 6 >>>
+```
+
+The column assertion `6` on the ending line should be optional.
+
+In cases of overlapping such assertions, resolve ambiguity using more angle brackets: `<<<<` begins
+an assertion ended by `>>>>`, etc.
+
+### Non-Python files
+
+Some tests may need to specify non-Python embedded files: typeshed `stdlib/VERSIONS`, `pth` files,
+`py.typed` files, `pyvenv.cfg` files...
+
+We will allow specifying any of these using the `text` language in the code block tag string:
+
+````markdown
+```text path=/third-party/foo/py.typed
+partial
+```
+````
+
+We may want to also support testing Jupyter notebooks as embedded files; exact syntax for this is
+yet to be determined.
+
+Of course, red-knot is only run directly on `py` and `pyi` files, and assertion comments are only
+possible in these files.
+
+A fenced code block with no language will always be an error.
+
+### Configuration
+
+We will add the ability to specify non-default red-knot configurations to use in tests, by including
+a TOML code block:
+
+````markdown
+```toml
+[tool.knot]
+warn-on-any = true
+```
+
+```py
+from typing import Any
+
+def f(x: Any):  # error: [use-of-any]
+    pass
+```
+````
+
+It should be possible to include a TOML code block in a single test (as shown), or in a grouping
+section, in which case it applies to all nested tests within that grouping section. Configurations
+at multiple level are allowed and merged, with the most-nested (closest to the test) taking
+precedence.
+
+### Running just a single test from a suite
+
+Having each test in a suite always run as a distinct Rust test would require writing our own test
+runner or code-generating tests in a build script; neither of these is planned.
+
+We could still allow running just a single test from a suite, for debugging purposes, either via
+some "focus" syntax that could be easily temporarily added to a test, or via an environment
+variable.
+
+### Configuring search paths and kinds
+
+The red-knot TOML configuration format hasn't been designed yet, and we may want to implement
+support in the test framework for configuring search paths before it is designed. If so, we can
+define some configuration options for now under the `[tool.knot.tests]` namespace. In the future,
+perhaps some of these can be replaced by real red-knot configuration options; some or all may also
+be kept long-term as test-specific options.
+
+Some configuration options we will want to provide:
+
+- We should be able to configure the default workspace root to something other than `/src/` using a
+    `workspace-root` configuration option.
+
+- We should be able to add a third-party root using the `third-party-root` configuration option.
+
+- We may want to add additional configuration options for setting additional search path kinds.
+
+Paths for `workspace-root` and `third-party-root` must be absolute.
+
+Relative embedded-file paths are relative to the workspace root, even if it is explicitly set to a
+non-default value using the `workspace-root` config.
+
+### Specifying a custom typeshed
+
+Some tests will need to override the default typeshed with custom files. The `[tool.knot.tests]`
+configuration option `typeshed-root` should be usable for this:
+
+````markdown
+```toml
+[tool.knot.tests]
+typeshed-root = "/typeshed"
+```
+
+This file is importable as part of our custom typeshed, because it is within `/typeshed`, which we
+configured above as our custom typeshed root:
+
+```py path=/typeshed/stdlib/builtins.pyi
+I_AM_THE_ONLY_BUILTIN = 1
+```
+
+This file is written to `/src/test.py`, because the default workspace root is `/src/ and the default
+file path is `test.py`:
+
+```py
+reveal_type(I_AM_THE_ONLY_BUILTIN)  # revealed: Literal[1]
+```
+
+````
+
+A fenced code block with language `text` can be used to provide a `stdlib/VERSIONS` file in the
+custom typeshed root. If no such file is created explicitly, one should be created implicitly
+including entries enabling all specified `<typeshed-root>/stdlib` files for all supported Python
+versions.
+
+### I/O errors
+
+We could use an `error=` configuration option in the tag string to make an embedded file cause an
+I/O error on read.
+
+### Asserting on full diagnostic output
+
+The inline comment diagnostic assertions are useful for making quick, readable assertions about
+diagnostics in a particular location. But sometimes we will want to assert on the full diagnostic
+output of checking an embedded Python file. Or sometimes (see “incremental tests” below) we will
+want to assert on diagnostics in a file, without impacting the contents of that file by changing a
+comment in it. In these cases, a Python code block in a test could be followed by a fenced code
+block with language `output`; this would contain the full diagnostic output for the preceding test
+file:
+
+````markdown
+# full output
+
+```py
+x = 1
+reveal_type(x)
+```
+
+This is just an example, not a proposal that red-knot would ever actually output diagnostics in
+precisely this format:
+
+```output
+test.py, line 1, col 1: revealed type is 'Literal[1]'
+```
+````
+
+We will want to build tooling to automatically capture and update these “full diagnostic output”
+blocks, when tests are run in an update-output mode (probably specified by an environment variable.)
+
+By default, an `output` block will specify diagnostic output for the file `<workspace-root>/test.py`.
+An `output` block can have a `path=` option, to explicitly specify the Python file for which it
+asserts diagnostic output, and a `stage=` option, to specify which stage of an incremental test it
+specifies diagnostic output at. (See “incremental tests” below.)
+
+It is an error for an `output` block to exist, if there is no `py` or `python` block in the same
+test for the same file path.
+
+### Incremental tests
+
+Some tests should validate incremental checking, by initially creating some files, checking them,
+and then modifying/adding/deleting files and checking again.
+
+We should add the capability to create an incremental test by using the `stage=` option on some
+fenced code blocks in the test:
+
+````markdown
+# Incremental
+
+## modify a file
+
+Initial version of `test.py` and `b.py`:
+
+```py
+from b import x
+reveal_type(x)
+```
+
+```py path=b.py
+x = 1
+```
+
+Initial expected output for `test.py`:
+
+```output
+/src/test.py, line 1, col 1: revealed type is 'Literal[1]'
+```
+
+Now in our first incremental stage, modify the contents of `b.py`:
+
+```py path=b.py stage=1
+# b.py
+x = 2
+```
+
+And this is our updated expected output for `test.py` at stage 1:
+
+```output stage=1
+/src/test.py, line 1, col 1: revealed type is 'Literal[2]'
+```
+
+(One reason to use full-diagnostic-output blocks in this test is that updating
+inline-comment diagnostic assertions for `test.py` would require specifying new
+contents for `test.py` in stage 1, which we don't want to do in this test.)
+````
+
+It will be possible to provide any number of stages in an incremental test. If a stage re-specifies
+a filename that was specified in a previous stage (or the initial stage), that file is modified. A
+new filename appearing for the first time in a new stage will create a new file. To delete a
+previously created file, specify that file with the tag `delete` in its tag string (in this case, it
+is an error to provide non-empty contents). Any previously-created files that are not re-specified
+in a later stage continue to exist with their previously-specified contents, and are not "touched".
+
+All stages should be run in order, incrementally, and then the final state should also be re-checked
+cold, to validate equivalence of cold and incremental check results.
+
+[^extensions]: `typing-extensions` is a third-party module, but typeshed, and thus type checkers
+    also, treat it as part of the standard library.

crates/red_knot_test/src/assertion.rs

@@ -0,0 +1,621 @@
+//! Parse type and type-error assertions in Python comment form.
+//!
+//! Parses comments of the form `# revealed: SomeType` and `# error: 8 [rule-code] "message text"`.
+//! In the latter case, the `8` is a column number, and `"message text"` asserts that the full
+//! diagnostic message contains the text `"message text"`; all three are optional (`# error:` will
+//! match any error.)
+//!
+//! Assertion comments may be placed at end-of-line:
+//!
+//! ```py
+//! x: int = "foo"  # error: [invalid-assignment]
+//! ```
+//!
+//! Or as a full-line comment on the preceding line:
+//!
+//! ```py
+//! # error: [invalid-assignment]
+//! x: int = "foo"
+//! ```
+//!
+//! Multiple assertion comments may apply to the same line; in this case all (or all but the last)
+//! must be full-line comments:
+//!
+//! ```py
+//! # error: [unbound-name]
+//! reveal_type(x)  # revealed: Unbound
+//! ```
+//!
+//! or
+//!
+//! ```py
+//! # error: [unbound-name]
+//! # revealed: Unbound
+//! reveal_type(x)
+//! ```
+
+use crate::db::Db;
+use once_cell::sync::Lazy;
+use regex::Regex;
+use ruff_db::files::File;
+use ruff_db::parsed::parsed_module;
+use ruff_db::source::{line_index, source_text, SourceText};
+use ruff_python_trivia::CommentRanges;
+use ruff_source_file::{LineIndex, Locator, OneIndexed};
+use ruff_text_size::{Ranged, TextRange};
+use smallvec::SmallVec;
+use std::ops::Deref;
+
+/// Diagnostic assertion comments in a single embedded file.
+#[derive(Debug)]
+pub(crate) struct InlineFileAssertions {
+    comment_ranges: CommentRanges,
+    source: SourceText,
+    lines: LineIndex,
+}
+
+impl InlineFileAssertions {
+    pub(crate) fn from_file(db: &Db, file: File) -> Self {
+        let source = source_text(db, file);
+        let lines = line_index(db, file);
+        let parsed = parsed_module(db, file);
+        let comment_ranges = CommentRanges::from(parsed.tokens());
+        Self {
+            comment_ranges,
+            source,
+            lines,
+        }
+    }
+
+    fn locator(&self) -> Locator {
+        Locator::with_index(&self.source, self.lines.clone())
+    }
+
+    fn line_number(&self, range: &impl Ranged) -> OneIndexed {
+        self.lines.line_index(range.start())
+    }
+
+    fn is_own_line_comment(&self, ranged_assertion: &AssertionWithRange) -> bool {
+        CommentRanges::is_own_line(ranged_assertion.start(), &self.locator())
+    }
+}
+
+impl<'a> IntoIterator for &'a InlineFileAssertions {
+    type Item = LineAssertions<'a>;
+    type IntoIter = LineAssertionsIterator<'a>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        Self::IntoIter {
+            file_assertions: self,
+            inner: AssertionWithRangeIterator {
+                file_assertions: self,
+                inner: self.comment_ranges.into_iter(),
+            }
+            .peekable(),
+        }
+    }
+}
+
+/// An [`Assertion`] with the [`TextRange`] of its original inline comment.
+#[derive(Debug)]
+struct AssertionWithRange<'a>(Assertion<'a>, TextRange);
+
+impl<'a> Deref for AssertionWithRange<'a> {
+    type Target = Assertion<'a>;
+
+    fn deref(&self) -> &Self::Target {
+        &self.0
+    }
+}
+
+impl Ranged for AssertionWithRange<'_> {
+    fn range(&self) -> TextRange {
+        self.1
+    }
+}
+
+impl<'a> From<AssertionWithRange<'a>> for Assertion<'a> {
+    fn from(value: AssertionWithRange<'a>) -> Self {
+        value.0
+    }
+}
+
+/// Iterator that yields all assertions within a single embedded Python file.
+#[derive(Debug)]
+struct AssertionWithRangeIterator<'a> {
+    file_assertions: &'a InlineFileAssertions,
+    inner: std::iter::Copied<std::slice::Iter<'a, TextRange>>,
+}
+
+impl<'a> Iterator for AssertionWithRangeIterator<'a> {
+    type Item = AssertionWithRange<'a>;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let locator = self.file_assertions.locator();
+        loop {
+            let inner_next = self.inner.next()?;
+            let comment = locator.slice(inner_next);
+            if let Some(assertion) = Assertion::from_comment(comment) {
+                return Some(AssertionWithRange(assertion, inner_next));
+            };
+        }
+    }
+}
+
+impl std::iter::FusedIterator for AssertionWithRangeIterator<'_> {}
+
+/// A vector of [`Assertion`]s belonging to a single line.
+///
+/// Most lines will have zero or one assertion, so we use a [`SmallVec`] optimized for a single
+/// element to avoid most heap vector allocations.
+type AssertionVec<'a> = SmallVec<[Assertion<'a>; 1]>;
+
+#[derive(Debug)]
+pub(crate) struct LineAssertionsIterator<'a> {
+    file_assertions: &'a InlineFileAssertions,
+    inner: std::iter::Peekable<AssertionWithRangeIterator<'a>>,
+}
+
+impl<'a> Iterator for LineAssertionsIterator<'a> {
+    type Item = LineAssertions<'a>;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let file = self.file_assertions;
+        let ranged_assertion = self.inner.next()?;
+        let mut collector = AssertionVec::new();
+        let mut line_number = file.line_number(&ranged_assertion);
+        // Collect all own-line comments on consecutive lines; these all apply to the same line of
+        // code. For example:
+        //
+        // ```py
+        // # error: [unbound-name]
+        // # revealed: Unbound
+        // reveal_type(x)
+        // ```
+        //
+        if file.is_own_line_comment(&ranged_assertion) {
+            collector.push(ranged_assertion.into());
+            let mut only_own_line = true;
+            while let Some(ranged_assertion) = self.inner.peek() {
+                let next_line_number = line_number.saturating_add(1);
+                if file.line_number(ranged_assertion) == next_line_number {
+                    if !file.is_own_line_comment(ranged_assertion) {
+                        only_own_line = false;
+                    }
+                    line_number = next_line_number;
+                    collector.push(self.inner.next().unwrap().into());
+                    // If we see an end-of-line comment, it has to be the end of the stack,
+                    // otherwise we'd botch this case, attributing all three errors to the `bar`
+                    // line:
+                    //
+                    // ```py
+                    // # error:
+                    // foo  # error:
+                    // bar  # error:
+                    // ```
+                    //
+                    if !only_own_line {
+                        break;
+                    }
+                } else {
+                    break;
+                }
+            }
+            if only_own_line {
+                // The collected comments apply to the _next_ line in the code.
+                line_number = line_number.saturating_add(1);
+            }
+        } else {
+            // We have a line-trailing comment; it applies to its own line, and is not grouped.
+            collector.push(ranged_assertion.into());
+        }
+        Some(LineAssertions {
+            line_number,
+            assertions: collector,
+        })
+    }
+}
+
+impl std::iter::FusedIterator for LineAssertionsIterator<'_> {}
+
+/// One or more assertions referring to the same line of code.
+#[derive(Debug)]
+pub(crate) struct LineAssertions<'a> {
+    /// The line these assertions refer to.
+    ///
+    /// Not necessarily the same line the assertion comment is located on; for an own-line comment,
+    /// it's the next non-assertion line.
+    pub(crate) line_number: OneIndexed,
+
+    /// The assertions referring to this line.
+    pub(crate) assertions: AssertionVec<'a>,
+}
+
+impl<'a> Deref for LineAssertions<'a> {
+    type Target = [Assertion<'a>];
+
+    fn deref(&self) -> &Self::Target {
+        &self.assertions
+    }
+}
+
+static TYPE_RE: Lazy<Regex> =
+    Lazy::new(|| Regex::new(r"^#\s*revealed:\s*(?<ty_display>.+?)\s*$").unwrap());
+
+static ERROR_RE: Lazy<Regex> = Lazy::new(|| {
+    Regex::new(
+        r#"^#\s*error:(\s*(?<column>\d+))?(\s*\[(?<rule>.+?)\])?(\s*"(?<message>.+?)")?\s*$"#,
+    )
+    .unwrap()
+});
+
+/// A single diagnostic assertion comment.
+#[derive(Debug)]
+pub(crate) enum Assertion<'a> {
+    /// A `revealed: ` assertion.
+    Revealed(&'a str),
+
+    /// An `error: ` assertion.
+    Error(ErrorAssertion<'a>),
+}
+
+impl<'a> Assertion<'a> {
+    fn from_comment(comment: &'a str) -> Option<Self> {
+        if let Some(caps) = TYPE_RE.captures(comment) {
+            Some(Self::Revealed(caps.name("ty_display").unwrap().as_str()))
+        } else {
+            ERROR_RE.captures(comment).map(|caps| {
+                Self::Error(ErrorAssertion {
+                    rule: caps.name("rule").map(|m| m.as_str()),
+                    column: caps.name("column").and_then(|m| m.as_str().parse().ok()),
+                    message_contains: caps.name("message").map(|m| m.as_str()),
+                })
+            })
+        }
+    }
+}
+
+impl std::fmt::Display for Assertion<'_> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Revealed(expected_type) => write!(f, "revealed: {expected_type}"),
+            Self::Error(assertion) => assertion.fmt(f),
+        }
+    }
+}
+
+/// An `error: ` assertion comment.
+#[derive(Debug)]
+pub(crate) struct ErrorAssertion<'a> {
+    /// The diagnostic rule code we expect.
+    pub(crate) rule: Option<&'a str>,
+
+    /// The column we expect the diagnostic range to start at.
+    pub(crate) column: Option<OneIndexed>,
+
+    /// A string we expect to be contained in the diagnostic message.
+    pub(crate) message_contains: Option<&'a str>,
+}
+
+impl std::fmt::Display for ErrorAssertion<'_> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("error:")?;
+        if let Some(column) = self.column {
+            write!(f, " {column}")?;
+        }
+        if let Some(rule) = self.rule {
+            write!(f, " [{rule}]")?;
+        }
+        if let Some(message) = self.message_contains {
+            write!(f, r#" "{message}""#)?;
+        }
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{Assertion, InlineFileAssertions, LineAssertions};
+    use ruff_db::files::system_path_to_file;
+    use ruff_db::system::{DbWithTestSystem, SystemPathBuf};
+    use ruff_python_trivia::textwrap::dedent;
+    use ruff_source_file::OneIndexed;
+
+    fn get_assertions(source: &str) -> InlineFileAssertions {
+        let mut db = crate::db::Db::setup(SystemPathBuf::from("/src"));
+        db.write_file("/src/test.py", source).unwrap();
+        let file = system_path_to_file(&db, "/src/test.py").unwrap();
+        InlineFileAssertions::from_file(&db, file)
+    }
+
+    fn as_vec(assertions: &InlineFileAssertions) -> Vec<LineAssertions> {
+        assertions.into_iter().collect()
+    }
+
+    #[test]
+    fn ty_display() {
+        let assertions = get_assertions(&dedent(
+            "
+            reveal_type(1)  # revealed: Literal[1]
+            ",
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(1));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(format!("{assert}"), "revealed: Literal[1]");
+    }
+
+    #[test]
+    fn error() {
+        let assertions = get_assertions(&dedent(
+            "
+            x  # error:
+            ",
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(1));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(format!("{assert}"), "error:");
+    }
+
+    #[test]
+    fn prior_line() {
+        let assertions = get_assertions(&dedent(
+            "
+            # revealed: Literal[1]
+            reveal_type(1)
+            ",
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(2));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(format!("{assert}"), "revealed: Literal[1]");
+    }
+
+    #[test]
+    fn stacked_prior_line() {
+        let assertions = get_assertions(&dedent(
+            "
+            # revealed: Unbound
+            # error: [unbound-name]
+            reveal_type(x)
+            ",
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(3));
+
+        let [assert1, assert2] = &line.assertions[..] else {
+            panic!("expected two assertions");
+        };
+
+        assert_eq!(format!("{assert1}"), "revealed: Unbound");
+        assert_eq!(format!("{assert2}"), "error: [unbound-name]");
+    }
+
+    #[test]
+    fn stacked_mixed() {
+        let assertions = get_assertions(&dedent(
+            "
+            # revealed: Unbound
+            reveal_type(x) # error: [unbound-name]
+            ",
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(2));
+
+        let [assert1, assert2] = &line.assertions[..] else {
+            panic!("expected two assertions");
+        };
+
+        assert_eq!(format!("{assert1}"), "revealed: Unbound");
+        assert_eq!(format!("{assert2}"), "error: [unbound-name]");
+    }
+
+    #[test]
+    fn multiple_lines() {
+        let assertions = get_assertions(&dedent(
+            r#"
+            # error: [invalid-assignment]
+            x: int = "foo"
+            y  # error: [unbound-name]
+            "#,
+        ));
+
+        let [line1, line2] = &as_vec(&assertions)[..] else {
+            panic!("expected two lines");
+        };
+
+        assert_eq!(line1.line_number, OneIndexed::from_zero_indexed(2));
+        assert_eq!(line2.line_number, OneIndexed::from_zero_indexed(3));
+
+        let [Assertion::Error(error1)] = &line1.assertions[..] else {
+            panic!("expected one error assertion");
+        };
+
+        assert_eq!(error1.rule, Some("invalid-assignment"));
+
+        let [Assertion::Error(error2)] = &line2.assertions[..] else {
+            panic!("expected one error assertion");
+        };
+
+        assert_eq!(error2.rule, Some("unbound-name"));
+    }
+
+    #[test]
+    fn multiple_lines_mixed_stack() {
+        let assertions = get_assertions(&dedent(
+            r#"
+            # error: [invalid-assignment]
+            x: int = reveal_type("foo")  # revealed: str
+            y  # error: [unbound-name]
+            "#,
+        ));
+
+        let [line1, line2] = &as_vec(&assertions)[..] else {
+            panic!("expected two lines");
+        };
+
+        assert_eq!(line1.line_number, OneIndexed::from_zero_indexed(2));
+        assert_eq!(line2.line_number, OneIndexed::from_zero_indexed(3));
+
+        let [Assertion::Error(error1), Assertion::Revealed(expected_ty)] = &line1.assertions[..]
+        else {
+            panic!("expected one error assertion and one Revealed assertion");
+        };
+
+        assert_eq!(error1.rule, Some("invalid-assignment"));
+        assert_eq!(*expected_ty, "str");
+
+        let [Assertion::Error(error2)] = &line2.assertions[..] else {
+            panic!("expected one error assertion");
+        };
+
+        assert_eq!(error2.rule, Some("unbound-name"));
+    }
+
+    #[test]
+    fn error_with_rule() {
+        let assertions = get_assertions(&dedent(
+            "
+            x  # error: [unbound-name]
+            ",
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(1));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(format!("{assert}"), "error: [unbound-name]");
+    }
+
+    #[test]
+    fn error_with_rule_and_column() {
+        let assertions = get_assertions(&dedent(
+            "
+            x  # error: 1 [unbound-name]
+            ",
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(1));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(format!("{assert}"), "error: 1 [unbound-name]");
+    }
+
+    #[test]
+    fn error_with_rule_and_message() {
+        let assertions = get_assertions(&dedent(
+            r#"
+            # error: [unbound-name] "`x` is unbound"
+            x
+            "#,
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(2));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(
+            format!("{assert}"),
+            r#"error: [unbound-name] "`x` is unbound""#
+        );
+    }
+
+    #[test]
+    fn error_with_message_and_column() {
+        let assertions = get_assertions(&dedent(
+            r#"
+            # error: 1 "`x` is unbound"
+            x
+            "#,
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(2));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(format!("{assert}"), r#"error: 1 "`x` is unbound""#);
+    }
+
+    #[test]
+    fn error_with_rule_and_message_and_column() {
+        let assertions = get_assertions(&dedent(
+            r#"
+            # error: 1 [unbound-name] "`x` is unbound"
+            x
+            "#,
+        ));
+
+        let [line] = &as_vec(&assertions)[..] else {
+            panic!("expected one line");
+        };
+
+        assert_eq!(line.line_number, OneIndexed::from_zero_indexed(2));
+
+        let [assert] = &line.assertions[..] else {
+            panic!("expected one assertion");
+        };
+
+        assert_eq!(
+            format!("{assert}"),
+            r#"error: 1 [unbound-name] "`x` is unbound""#
+        );
+    }
+}

crates/red_knot_test/src/db.rs

@@ -0,0 +1,88 @@
+use red_knot_python_semantic::{
+    Db as SemanticDb, Program, ProgramSettings, PythonVersion, SearchPathSettings,
+};
+use ruff_db::files::{File, Files};
+use ruff_db::system::SystemPathBuf;
+use ruff_db::system::{DbWithTestSystem, System, TestSystem};
+use ruff_db::vendored::VendoredFileSystem;
+use ruff_db::{Db as SourceDb, Upcast};
+
+#[salsa::db]
+pub(crate) struct Db {
+    storage: salsa::Storage<Self>,
+    files: Files,
+    system: TestSystem,
+    vendored: VendoredFileSystem,
+}
+
+impl Db {
+    pub(crate) fn setup(workspace_root: SystemPathBuf) -> Self {
+        let db = Self {
+            storage: salsa::Storage::default(),
+            system: TestSystem::default(),
+            vendored: red_knot_vendored::file_system().clone(),
+            files: Files::default(),
+        };
+
+        db.memory_file_system()
+            .create_directory_all(&workspace_root)
+            .unwrap();
+
+        Program::from_settings(
+            &db,
+            &ProgramSettings {
+                target_version: PythonVersion::default(),
+                search_paths: SearchPathSettings::new(workspace_root),
+            },
+        )
+        .expect("Invalid search path settings");
+
+        db
+    }
+}
+
+impl DbWithTestSystem for Db {
+    fn test_system(&self) -> &TestSystem {
+        &self.system
+    }
+
+    fn test_system_mut(&mut self) -> &mut TestSystem {
+        &mut self.system
+    }
+}
+
+#[salsa::db]
+impl SourceDb for Db {
+    fn vendored(&self) -> &VendoredFileSystem {
+        &self.vendored
+    }
+
+    fn system(&self) -> &dyn System {
+        &self.system
+    }
+
+    fn files(&self) -> &Files {
+        &self.files
+    }
+}
+
+impl Upcast<dyn SourceDb> for Db {
+    fn upcast(&self) -> &(dyn SourceDb + 'static) {
+        self
+    }
+    fn upcast_mut(&mut self) -> &mut (dyn SourceDb + 'static) {
+        self
+    }
+}
+
+#[salsa::db]
+impl SemanticDb for Db {
+    fn is_file_open(&self, file: File) -> bool {
+        !file.path(self).is_vendored_path()
+    }
+}
+
+#[salsa::db]
+impl salsa::Database for Db {
+    fn salsa_event(&self, _event: &dyn Fn() -> salsa::Event) {}
+}

crates/red_knot_test/src/diagnostic.rs

@@ -0,0 +1,173 @@
+//! Sort and group diagnostics by line number, so they can be correlated with assertions.
+//!
+//! We don't assume that we will get the diagnostics in source order.
+
+use ruff_source_file::{LineIndex, OneIndexed};
+use ruff_text_size::Ranged;
+use std::ops::{Deref, Range};
+
+/// All diagnostics for one embedded Python file, sorted and grouped by start line number.
+///
+/// The diagnostics are kept in a flat vector, sorted by line number. A separate vector of
+/// [`LineDiagnosticRange`] has one entry for each contiguous slice of the diagnostics vector
+/// containing diagnostics which all start on the same line.
+#[derive(Debug)]
+pub(crate) struct SortedDiagnostics<T> {
+    diagnostics: Vec<T>,
+    line_ranges: Vec<LineDiagnosticRange>,
+}
+
+impl<T> SortedDiagnostics<T>
+where
+    T: Ranged + Clone,
+{
+    pub(crate) fn new(diagnostics: impl IntoIterator<Item = T>, line_index: &LineIndex) -> Self {
+        let mut diagnostics: Vec<_> = diagnostics
+            .into_iter()
+            .map(|diagnostic| DiagnosticWithLine {
+                line_number: line_index.line_index(diagnostic.start()),
+                diagnostic,
+            })
+            .collect();
+        diagnostics.sort_unstable_by_key(|diagnostic_with_line| diagnostic_with_line.line_number);
+
+        let mut diags = Self {
+            diagnostics: Vec::with_capacity(diagnostics.len()),
+            line_ranges: vec![],
+        };
+
+        let mut current_line_number = None;
+        let mut start = 0;
+        for DiagnosticWithLine {
+            line_number,
+            diagnostic,
+        } in diagnostics
+        {
+            match current_line_number {
+                None => {
+                    current_line_number = Some(line_number);
+                }
+                Some(current) => {
+                    if line_number != current {
+                        let end = diags.diagnostics.len();
+                        diags.line_ranges.push(LineDiagnosticRange {
+                            line_number: current,
+                            diagnostic_index_range: start..end,
+                        });
+                        start = end;
+                        current_line_number = Some(line_number);
+                    }
+                }
+            }
+            diags.diagnostics.push(diagnostic);
+        }
+        if let Some(line_number) = current_line_number {
+            diags.line_ranges.push(LineDiagnosticRange {
+                line_number,
+                diagnostic_index_range: start..diags.diagnostics.len(),
+            });
+        }
+
+        diags
+    }
+
+    pub(crate) fn iter_lines(&self) -> LineDiagnosticsIterator<T> {
+        LineDiagnosticsIterator {
+            diagnostics: self.diagnostics.as_slice(),
+            inner: self.line_ranges.iter(),
+        }
+    }
+}
+
+/// Range delineating diagnostics in [`SortedDiagnostics`] that begin on a single line.
+#[derive(Debug)]
+struct LineDiagnosticRange {
+    line_number: OneIndexed,
+    diagnostic_index_range: Range<usize>,
+}
+
+/// Iterator to group sorted diagnostics by line.
+pub(crate) struct LineDiagnosticsIterator<'a, T> {
+    diagnostics: &'a [T],
+    inner: std::slice::Iter<'a, LineDiagnosticRange>,
+}
+
+impl<'a, T> Iterator for LineDiagnosticsIterator<'a, T>
+where
+    T: Ranged + Clone,
+{
+    type Item = LineDiagnostics<'a, T>;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let LineDiagnosticRange {
+            line_number,
+            diagnostic_index_range,
+        } = self.inner.next()?;
+        Some(LineDiagnostics {
+            line_number: *line_number,
+            diagnostics: &self.diagnostics[diagnostic_index_range.clone()],
+        })
+    }
+}
+
+impl<T> std::iter::FusedIterator for LineDiagnosticsIterator<'_, T> where T: Clone + Ranged {}
+
+/// All diagnostics that start on a single line of source code in one embedded Python file.
+#[derive(Debug)]
+pub(crate) struct LineDiagnostics<'a, T> {
+    /// Line number on which these diagnostics start.
+    pub(crate) line_number: OneIndexed,
+
+    /// Diagnostics starting on this line.
+    pub(crate) diagnostics: &'a [T],
+}
+
+impl<T> Deref for LineDiagnostics<'_, T> {
+    type Target = [T];
+
+    fn deref(&self) -> &Self::Target {
+        self.diagnostics
+    }
+}
+
+#[derive(Debug)]
+struct DiagnosticWithLine<T> {
+    line_number: OneIndexed,
+    diagnostic: T,
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::db::Db;
+    use ruff_db::files::system_path_to_file;
+    use ruff_db::source::line_index;
+    use ruff_db::system::{DbWithTestSystem, SystemPathBuf};
+    use ruff_source_file::OneIndexed;
+    use ruff_text_size::{TextRange, TextSize};
+
+    #[test]
+    fn sort_and_group() {
+        let mut db = Db::setup(SystemPathBuf::from("/src"));
+        db.write_file("/src/test.py", "one\ntwo\n").unwrap();
+        let file = system_path_to_file(&db, "/src/test.py").unwrap();
+        let lines = line_index(&db, file);
+
+        let ranges = vec![
+            TextRange::new(TextSize::new(0), TextSize::new(1)),
+            TextRange::new(TextSize::new(5), TextSize::new(10)),
+            TextRange::new(TextSize::new(1), TextSize::new(7)),
+        ];
+
+        let sorted = super::SortedDiagnostics::new(&ranges, &lines);
+        let grouped = sorted.iter_lines().collect::<Vec<_>>();
+
+        let [line1, line2] = &grouped[..] else {
+            panic!("expected two lines");
+        };
+
+        assert_eq!(line1.line_number, OneIndexed::from_zero_indexed(0));
+        assert_eq!(line1.diagnostics.len(), 2);
+        assert_eq!(line2.line_number, OneIndexed::from_zero_indexed(1));
+        assert_eq!(line2.diagnostics.len(), 1);
+    }
+}

crates/red_knot_test/src/lib.rs

@@ -0,0 +1,95 @@
+use colored::Colorize;
+use parser as test_parser;
+use red_knot_python_semantic::types::check_types;
+use ruff_db::files::system_path_to_file;
+use ruff_db::parsed::parsed_module;
+use ruff_db::system::{DbWithTestSystem, SystemPathBuf};
+use std::collections::BTreeMap;
+use std::path::PathBuf;
+
+type Failures = BTreeMap<SystemPathBuf, matcher::FailuresByLine>;
+
+mod assertion;
+mod db;
+mod diagnostic;
+mod matcher;
+mod parser;
+
+/// Run `path` as a markdown test suite with given `title`.
+///
+/// Panic on test failure, and print failure details.
+#[allow(clippy::print_stdout)]
+pub fn run(path: &PathBuf, title: &str) {
+    let source = std::fs::read_to_string(path).unwrap();
+    let suite = match test_parser::parse(title, &source) {
+        Ok(suite) => suite,
+        Err(err) => {
+            panic!("Error parsing `{}`: {err}", path.to_str().unwrap())
+        }
+    };
+
+    let mut any_failures = false;
+    for test in suite.tests() {
+        if let Err(failures) = run_test(&test) {
+            any_failures = true;
+            println!("\n{}\n", test.name().bold().underline());
+
+            for (path, by_line) in failures {
+                println!("{}", path.as_str().bold());
+                for (line_number, failures) in by_line.iter() {
+                    for failure in failures {
+                        let line_info = format!("line {line_number}:").cyan();
+                        println!("    {line_info} {failure}");
+                    }
+                }
+                println!();
+            }
+        }
+    }
+
+    println!("{}\n", "-".repeat(50));
+
+    assert!(!any_failures, "Some tests failed.");
+}
+
+fn run_test(test: &parser::MarkdownTest) -> Result<(), Failures> {
+    let workspace_root = SystemPathBuf::from("/src");
+    let mut db = db::Db::setup(workspace_root.clone());
+
+    let mut system_paths = vec![];
+
+    for file in test.files() {
+        assert!(
+            matches!(file.lang, "py" | "pyi"),
+            "Non-Python files not supported yet."
+        );
+        let full_path = workspace_root.join(file.path);
+        db.write_file(&full_path, file.code).unwrap();
+        system_paths.push(full_path);
+    }
+
+    let mut failures = BTreeMap::default();
+
+    for path in system_paths {
+        let file = system_path_to_file(&db, path.clone()).unwrap();
+        let parsed = parsed_module(&db, file);
+
+        // TODO allow testing against code with syntax errors
+        assert!(
+            parsed.errors().is_empty(),
+            "Python syntax errors in {}, {:?}: {:?}",
+            test.name(),
+            path,
+            parsed.errors()
+        );
+
+        matcher::match_file(&db, file, check_types(&db, file)).unwrap_or_else(|line_failures| {
+            failures.insert(path, line_failures);
+        });
+    }
+    if failures.is_empty() {
+        Ok(())
+    } else {
+        Err(failures)
+    }
+}

crates/red_knot_test/src/matcher.rs

@@ -0,0 +1,929 @@
+//! Match [`TypeCheckDiagnostic`]s against [`Assertion`]s and produce test failure messages for any
+//! mismatches.
+use crate::assertion::{Assertion, ErrorAssertion, InlineFileAssertions};
+use crate::db::Db;
+use crate::diagnostic::SortedDiagnostics;
+use colored::Colorize;
+use red_knot_python_semantic::types::TypeCheckDiagnostic;
+use ruff_db::files::File;
+use ruff_db::source::{line_index, source_text, SourceText};
+use ruff_source_file::{LineIndex, OneIndexed};
+use ruff_text_size::Ranged;
+use std::cmp::Ordering;
+use std::ops::Range;
+use std::sync::Arc;
+
+#[derive(Debug, Default)]
+pub(super) struct FailuresByLine {
+    failures: Vec<String>,
+    lines: Vec<LineFailures>,
+}
+
+impl FailuresByLine {
+    pub(super) fn iter(&self) -> impl Iterator<Item = (OneIndexed, &[String])> {
+        self.lines.iter().map(|line_failures| {
+            (
+                line_failures.line_number,
+                &self.failures[line_failures.range.clone()],
+            )
+        })
+    }
+
+    fn push(&mut self, line_number: OneIndexed, messages: Vec<String>) {
+        let start = self.failures.len();
+        self.failures.extend(messages);
+        self.lines.push(LineFailures {
+            line_number,
+            range: start..self.failures.len(),
+        });
+    }
+
+    fn is_empty(&self) -> bool {
+        self.lines.is_empty()
+    }
+}
+
+#[derive(Debug)]
+struct LineFailures {
+    line_number: OneIndexed,
+    range: Range<usize>,
+}
+
+pub(super) fn match_file<T>(
+    db: &Db,
+    file: File,
+    diagnostics: impl IntoIterator<Item = T>,
+) -> Result<(), FailuresByLine>
+where
+    T: Diagnostic + Clone,
+{
+    // Parse assertions from comments in the file, and get diagnostics from the file; both
+    // ordered by line number.
+    let assertions = InlineFileAssertions::from_file(db, file);
+    let diagnostics = SortedDiagnostics::new(diagnostics, &line_index(db, file));
+
+    // Get iterators over assertions and diagnostics grouped by line, in ascending line order.
+    let mut line_assertions = assertions.into_iter();
+    let mut line_diagnostics = diagnostics.iter_lines();
+
+    let mut current_assertions = line_assertions.next();
+    let mut current_diagnostics = line_diagnostics.next();
+
+    let matcher = Matcher::from_file(db, file);
+    let mut failures = FailuresByLine::default();
+
+    loop {
+        match (&current_assertions, &current_diagnostics) {
+            (Some(assertions), Some(diagnostics)) => {
+                match assertions.line_number.cmp(&diagnostics.line_number) {
+                    Ordering::Equal => {
+                        // We have assertions and diagnostics on the same line; check for
+                        // matches and error on any that don't match, then advance both
+                        // iterators.
+                        matcher
+                            .match_line(diagnostics, assertions)
+                            .unwrap_or_else(|messages| {
+                                failures.push(assertions.line_number, messages);
+                            });
+                        current_assertions = line_assertions.next();
+                        current_diagnostics = line_diagnostics.next();
+                    }
+                    Ordering::Less => {
+                        // We have assertions on an earlier line than diagnostics; report these
+                        // assertions as all unmatched, and advance the assertions iterator.
+                        failures.push(assertions.line_number, unmatched(assertions));
+                        current_assertions = line_assertions.next();
+                    }
+                    Ordering::Greater => {
+                        // We have diagnostics on an earlier line than assertions; report these
+                        // diagnostics as all unmatched, and advance the diagnostics iterator.
+                        failures.push(diagnostics.line_number, unmatched(diagnostics));
+                        current_diagnostics = line_diagnostics.next();
+                    }
+                }
+            }
+            (Some(assertions), None) => {
+                // We've exhausted diagnostics but still have assertions; report these assertions
+                // as unmatched and advance the assertions iterator.
+                failures.push(assertions.line_number, unmatched(assertions));
+                current_assertions = line_assertions.next();
+            }
+            (None, Some(diagnostics)) => {
+                // We've exhausted assertions but still have diagnostics; report these
+                // diagnostics as unmatched and advance the diagnostics iterator.
+                failures.push(diagnostics.line_number, unmatched(diagnostics));
+                current_diagnostics = line_diagnostics.next();
+            }
+            // When we've exhausted both diagnostics and assertions, break.
+            (None, None) => break,
+        }
+    }
+
+    if failures.is_empty() {
+        Ok(())
+    } else {
+        Err(failures)
+    }
+}
+
+pub(super) trait Diagnostic: Ranged {
+    fn rule(&self) -> &str;
+
+    fn message(&self) -> &str;
+}
+
+impl Diagnostic for Arc<TypeCheckDiagnostic> {
+    fn rule(&self) -> &str {
+        self.as_ref().rule()
+    }
+
+    fn message(&self) -> &str {
+        self.as_ref().message()
+    }
+}
+
+trait Unmatched {
+    fn unmatched(&self) -> String;
+}
+
+fn unmatched<'a, T: Unmatched + 'a>(unmatched: &'a [T]) -> Vec<String> {
+    unmatched.iter().map(Unmatched::unmatched).collect()
+}
+
+trait UnmatchedWithColumn {
+    fn unmatched_with_column(&self, column: OneIndexed) -> String;
+}
+
+impl Unmatched for Assertion<'_> {
+    fn unmatched(&self) -> String {
+        format!("{} {self}", "unmatched assertion:".red())
+    }
+}
+
+fn maybe_add_undefined_reveal_clarification<T: Diagnostic>(
+    diagnostic: &T,
+    original: std::fmt::Arguments,
+) -> String {
+    if diagnostic.rule() == "undefined-reveal" {
+        format!(
+            "{} add a `# revealed` assertion on this line (original diagnostic: {original})",
+            "used built-in `reveal_type`:".yellow()
+        )
+    } else {
+        format!("{} {original}", "unexpected error:".red())
+    }
+}
+
+impl<T> Unmatched for T
+where
+    T: Diagnostic,
+{
+    fn unmatched(&self) -> String {
+        maybe_add_undefined_reveal_clarification(
+            self,
+            format_args!(r#"[{}] "{}""#, self.rule(), self.message()),
+        )
+    }
+}
+
+impl<T> UnmatchedWithColumn for T
+where
+    T: Diagnostic,
+{
+    fn unmatched_with_column(&self, column: OneIndexed) -> String {
+        maybe_add_undefined_reveal_clarification(
+            self,
+            format_args!(r#"{column} [{}] "{}""#, self.rule(), self.message()),
+        )
+    }
+}
+
+struct Matcher {
+    line_index: LineIndex,
+    source: SourceText,
+}
+
+impl Matcher {
+    fn from_file(db: &Db, file: File) -> Self {
+        Self {
+            line_index: line_index(db, file),
+            source: source_text(db, file),
+        }
+    }
+
+    /// Check a slice of [`Diagnostic`]s against a slice of [`Assertion`]s.
+    ///
+    /// Return vector of [`Unmatched`] for any unmatched diagnostics or assertions.
+    fn match_line<'a, 'b, T: Diagnostic + 'a>(
+        &self,
+        diagnostics: &'a [T],
+        assertions: &'a [Assertion<'b>],
+    ) -> Result<(), Vec<String>>
+    where
+        'b: 'a,
+    {
+        let mut failures = vec![];
+        let mut unmatched: Vec<_> = diagnostics.iter().collect();
+        for assertion in assertions {
+            if matches!(
+                assertion,
+                Assertion::Error(ErrorAssertion {
+                    rule: None,
+                    message_contains: None,
+                    ..
+                })
+            ) {
+                failures.push(format!(
+                    "{} no rule or message text",
+                    "invalid assertion:".red()
+                ));
+                continue;
+            }
+            if !self.matches(assertion, &mut unmatched) {
+                failures.push(assertion.unmatched());
+            }
+        }
+        for diagnostic in unmatched {
+            failures.push(diagnostic.unmatched_with_column(self.column(diagnostic)));
+        }
+        if failures.is_empty() {
+            Ok(())
+        } else {
+            Err(failures)
+        }
+    }
+
+    fn column<T: Ranged>(&self, ranged: &T) -> OneIndexed {
+        self.line_index
+            .source_location(ranged.start(), &self.source)
+            .column
+    }
+
+    /// Check if `assertion` matches any [`Diagnostic`]s in `unmatched`.
+    ///
+    /// If so, return `true` and remove the matched diagnostics from `unmatched`. Otherwise, return
+    /// `false`.
+    ///
+    /// An `Error` assertion can only match one diagnostic; even if it could match more than one,
+    /// we short-circuit after the first match.
+    ///
+    /// A `Revealed` assertion must match a revealed-type diagnostic, and may also match an
+    /// undefined-reveal diagnostic, if present.
+    fn matches<T: Diagnostic>(&self, assertion: &Assertion, unmatched: &mut Vec<&T>) -> bool {
+        match assertion {
+            Assertion::Error(error) => {
+                let position = unmatched.iter().position(|diagnostic| {
+                    !error.rule.is_some_and(|rule| rule != diagnostic.rule())
+                        && !error
+                            .column
+                            .is_some_and(|col| col != self.column(*diagnostic))
+                        && !error
+                            .message_contains
+                            .is_some_and(|needle| !diagnostic.message().contains(needle))
+                });
+                if let Some(position) = position {
+                    unmatched.swap_remove(position);
+                    true
+                } else {
+                    false
+                }
+            }
+            Assertion::Revealed(expected_type) => {
+                let mut matched_revealed_type = None;
+                let mut matched_undefined_reveal = None;
+                let expected_reveal_type_message = format!("Revealed type is `{expected_type}`");
+                for (index, diagnostic) in unmatched.iter().enumerate() {
+                    if matched_revealed_type.is_none()
+                        && diagnostic.rule() == "revealed-type"
+                        && diagnostic.message() == expected_reveal_type_message
+                    {
+                        matched_revealed_type = Some(index);
+                    } else if matched_undefined_reveal.is_none()
+                        && diagnostic.rule() == "undefined-reveal"
+                    {
+                        matched_undefined_reveal = Some(index);
+                    }
+                    if matched_revealed_type.is_some() && matched_undefined_reveal.is_some() {
+                        break;
+                    }
+                }
+                let mut idx = 0;
+                unmatched.retain(|_| {
+                    let retain =
+                        Some(idx) != matched_revealed_type && Some(idx) != matched_undefined_reveal;
+                    idx += 1;
+                    retain
+                });
+                matched_revealed_type.is_some()
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::FailuresByLine;
+    use ruff_db::files::system_path_to_file;
+    use ruff_db::system::{DbWithTestSystem, SystemPathBuf};
+    use ruff_python_trivia::textwrap::dedent;
+    use ruff_source_file::OneIndexed;
+    use ruff_text_size::{Ranged, TextRange};
+
+    #[derive(Clone, Debug)]
+    struct TestDiagnostic {
+        rule: &'static str,
+        message: &'static str,
+        range: TextRange,
+    }
+
+    impl TestDiagnostic {
+        fn new(rule: &'static str, message: &'static str, offset: usize) -> Self {
+            let offset: u32 = offset.try_into().unwrap();
+            Self {
+                rule,
+                message,
+                range: TextRange::new(offset.into(), (offset + 1).into()),
+            }
+        }
+    }
+
+    impl super::Diagnostic for TestDiagnostic {
+        fn rule(&self) -> &str {
+            self.rule
+        }
+
+        fn message(&self) -> &str {
+            self.message
+        }
+    }
+
+    impl Ranged for TestDiagnostic {
+        fn range(&self) -> ruff_text_size::TextRange {
+            self.range
+        }
+    }
+
+    fn get_result(source: &str, diagnostics: Vec<TestDiagnostic>) -> Result<(), FailuresByLine> {
+        colored::control::set_override(false);
+
+        let mut db = crate::db::Db::setup(SystemPathBuf::from("/src"));
+        db.write_file("/src/test.py", source).unwrap();
+        let file = system_path_to_file(&db, "/src/test.py").unwrap();
+
+        super::match_file(&db, file, diagnostics)
+    }
+
+    fn assert_fail(result: Result<(), FailuresByLine>, messages: &[(usize, &[&str])]) {
+        let Err(failures) = result else {
+            panic!("expected a failure");
+        };
+
+        let expected: Vec<(OneIndexed, Vec<String>)> = messages
+            .iter()
+            .map(|(idx, msgs)| {
+                (
+                    OneIndexed::from_zero_indexed(*idx),
+                    msgs.iter().map(ToString::to_string).collect(),
+                )
+            })
+            .collect();
+        let failures: Vec<(OneIndexed, Vec<String>)> = failures
+            .iter()
+            .map(|(idx, msgs)| (idx, msgs.to_vec()))
+            .collect();
+
+        assert_eq!(failures, expected);
+    }
+
+    fn assert_ok(result: &Result<(), FailuresByLine>) {
+        assert!(result.is_ok(), "{result:?}");
+    }
+
+    #[test]
+    fn revealed_match() {
+        let result = get_result(
+            "x # revealed: Foo",
+            vec![TestDiagnostic::new(
+                "revealed-type",
+                "Revealed type is `Foo`",
+                0,
+            )],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn revealed_wrong_rule() {
+        let result = get_result(
+            "x # revealed: Foo",
+            vec![TestDiagnostic::new(
+                "not-revealed-type",
+                "Revealed type is `Foo`",
+                0,
+            )],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "unmatched assertion: revealed: Foo",
+                    r#"unexpected error: 1 [not-revealed-type] "Revealed type is `Foo`""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn revealed_wrong_message() {
+        let result = get_result(
+            "x # revealed: Foo",
+            vec![TestDiagnostic::new("revealed-type", "Something else", 0)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "unmatched assertion: revealed: Foo",
+                    r#"unexpected error: 1 [revealed-type] "Something else""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn revealed_unmatched() {
+        let result = get_result("x # revealed: Foo", vec![]);
+
+        assert_fail(result, &[(0, &["unmatched assertion: revealed: Foo"])]);
+    }
+
+    #[test]
+    fn revealed_match_with_undefined() {
+        let result = get_result(
+            "x # revealed: Foo",
+            vec![
+                TestDiagnostic::new("revealed-type", "Revealed type is `Foo`", 0),
+                TestDiagnostic::new("undefined-reveal", "Doesn't matter", 0),
+            ],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn revealed_match_with_only_undefined() {
+        let result = get_result(
+            "x # revealed: Foo",
+            vec![TestDiagnostic::new("undefined-reveal", "Doesn't matter", 0)],
+        );
+
+        assert_fail(result, &[(0, &["unmatched assertion: revealed: Foo"])]);
+    }
+
+    #[test]
+    fn revealed_mismatch_with_undefined() {
+        let result = get_result(
+            "x # revealed: Foo",
+            vec![
+                TestDiagnostic::new("revealed-type", "Revealed type is `Bar`", 0),
+                TestDiagnostic::new("undefined-reveal", "Doesn't matter", 0),
+            ],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "unmatched assertion: revealed: Foo",
+                    r#"unexpected error: 1 [revealed-type] "Revealed type is `Bar`""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn undefined_reveal_type_unmatched() {
+        let result = get_result(
+            "reveal_type(1)",
+            vec![
+                TestDiagnostic::new("undefined-reveal", "undefined reveal message", 0),
+                TestDiagnostic::new("revealed-type", "Revealed type is `Literal[1]`", 12),
+            ],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "used built-in `reveal_type`: add a `# revealed` assertion on this line (\
+                    original diagnostic: [undefined-reveal] \"undefined reveal message\")",
+                    r#"unexpected error: [revealed-type] "Revealed type is `Literal[1]`""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn undefined_reveal_type_mismatched() {
+        let result = get_result(
+            "reveal_type(1) # error: [something-else]",
+            vec![
+                TestDiagnostic::new("undefined-reveal", "undefined reveal message", 0),
+                TestDiagnostic::new("revealed-type", "Revealed type is `Literal[1]`", 12),
+            ],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "unmatched assertion: error: [something-else]",
+                    "used built-in `reveal_type`: add a `# revealed` assertion on this line (\
+                    original diagnostic: 1 [undefined-reveal] \"undefined reveal message\")",
+                    r#"unexpected error: 13 [revealed-type] "Revealed type is `Literal[1]`""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn error_unmatched() {
+        let result = get_result("x # error: [rule]", vec![]);
+
+        assert_fail(result, &[(0, &["unmatched assertion: error: [rule]"])]);
+    }
+
+    #[test]
+    fn error_match_rule() {
+        let result = get_result(
+            "x # error: [some-rule]",
+            vec![TestDiagnostic::new("some-rule", "Any message", 0)],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn error_wrong_rule() {
+        let result = get_result(
+            "x # error: [some-rule]",
+            vec![TestDiagnostic::new("anything", "Any message", 0)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "unmatched assertion: error: [some-rule]",
+                    r#"unexpected error: 1 [anything] "Any message""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn error_match_message() {
+        let result = get_result(
+            r#"x # error: "contains this""#,
+            vec![TestDiagnostic::new("anything", "message contains this", 0)],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn error_wrong_message() {
+        let result = get_result(
+            r#"x # error: "contains this""#,
+            vec![TestDiagnostic::new("anything", "Any message", 0)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    r#"unmatched assertion: error: "contains this""#,
+                    r#"unexpected error: 1 [anything] "Any message""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn error_match_column_and_rule() {
+        let result = get_result(
+            "x # error: 1 [some-rule]",
+            vec![TestDiagnostic::new("some-rule", "Any message", 0)],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn error_wrong_column() {
+        let result = get_result(
+            "x # error: 2 [rule]",
+            vec![TestDiagnostic::new("rule", "Any message", 0)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "unmatched assertion: error: 2 [rule]",
+                    r#"unexpected error: 1 [rule] "Any message""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn error_match_column_and_message() {
+        let result = get_result(
+            r#"x # error: 1 "contains this""#,
+            vec![TestDiagnostic::new("anything", "message contains this", 0)],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn error_match_rule_and_message() {
+        let result = get_result(
+            r#"x # error: [a-rule] "contains this""#,
+            vec![TestDiagnostic::new("a-rule", "message contains this", 0)],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn error_match_all() {
+        let result = get_result(
+            r#"x # error: 1 [a-rule] "contains this""#,
+            vec![TestDiagnostic::new("a-rule", "message contains this", 0)],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn error_match_all_wrong_column() {
+        let result = get_result(
+            r#"x # error: 2 [some-rule] "contains this""#,
+            vec![TestDiagnostic::new("some-rule", "message contains this", 0)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    r#"unmatched assertion: error: 2 [some-rule] "contains this""#,
+                    r#"unexpected error: 1 [some-rule] "message contains this""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn error_match_all_wrong_rule() {
+        let result = get_result(
+            r#"x # error: 1 [some-rule] "contains this""#,
+            vec![TestDiagnostic::new(
+                "other-rule",
+                "message contains this",
+                0,
+            )],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    r#"unmatched assertion: error: 1 [some-rule] "contains this""#,
+                    r#"unexpected error: 1 [other-rule] "message contains this""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn error_match_all_wrong_message() {
+        let result = get_result(
+            r#"x # error: 1 [some-rule] "contains this""#,
+            vec![TestDiagnostic::new("some-rule", "Any message", 0)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    r#"unmatched assertion: error: 1 [some-rule] "contains this""#,
+                    r#"unexpected error: 1 [some-rule] "Any message""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn interspersed_matches_and_mismatches() {
+        let source = dedent(
+            r#"
+            1 # error: [line-one]
+            2
+            3 # error: [line-three]
+            4 # error: [line-four]
+            5
+            6: # error: [line-six]
+            "#,
+        );
+        let two = source.find('2').unwrap();
+        let three = source.find('3').unwrap();
+        let five = source.find('5').unwrap();
+        let result = get_result(
+            &source,
+            vec![
+                TestDiagnostic::new("line-two", "msg", two),
+                TestDiagnostic::new("line-three", "msg", three),
+                TestDiagnostic::new("line-five", "msg", five),
+            ],
+        );
+
+        assert_fail(
+            result,
+            &[
+                (1, &["unmatched assertion: error: [line-one]"]),
+                (2, &[r#"unexpected error: [line-two] "msg""#]),
+                (4, &["unmatched assertion: error: [line-four]"]),
+                (5, &[r#"unexpected error: [line-five] "msg""#]),
+                (6, &["unmatched assertion: error: [line-six]"]),
+            ],
+        );
+    }
+
+    #[test]
+    fn more_diagnostics_than_assertions() {
+        let source = dedent(
+            r#"
+            1 # error: [line-one]
+            2
+            "#,
+        );
+        let one = source.find('1').unwrap();
+        let two = source.find('2').unwrap();
+        let result = get_result(
+            &source,
+            vec![
+                TestDiagnostic::new("line-one", "msg", one),
+                TestDiagnostic::new("line-two", "msg", two),
+            ],
+        );
+
+        assert_fail(result, &[(2, &[r#"unexpected error: [line-two] "msg""#])]);
+    }
+
+    #[test]
+    fn multiple_assertions_and_diagnostics_same_line() {
+        let source = dedent(
+            "
+            # error: [one-rule]
+            # error: [other-rule]
+            x
+            ",
+        );
+        let x = source.find('x').unwrap();
+        let result = get_result(
+            &source,
+            vec![
+                TestDiagnostic::new("one-rule", "msg", x),
+                TestDiagnostic::new("other-rule", "msg", x),
+            ],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn multiple_assertions_and_diagnostics_same_line_all_same() {
+        let source = dedent(
+            "
+            # error: [one-rule]
+            # error: [one-rule]
+            x
+            ",
+        );
+        let x = source.find('x').unwrap();
+        let result = get_result(
+            &source,
+            vec![
+                TestDiagnostic::new("one-rule", "msg", x),
+                TestDiagnostic::new("one-rule", "msg", x),
+            ],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn multiple_assertions_and_diagnostics_same_line_mismatch() {
+        let source = dedent(
+            "
+            # error: [one-rule]
+            # error: [other-rule]
+            x
+            ",
+        );
+        let x = source.find('x').unwrap();
+        let result = get_result(
+            &source,
+            vec![
+                TestDiagnostic::new("one-rule", "msg", x),
+                TestDiagnostic::new("other-rule", "msg", x),
+                TestDiagnostic::new("third-rule", "msg", x),
+            ],
+        );
+
+        assert_fail(
+            result,
+            &[(3, &[r#"unexpected error: 1 [third-rule] "msg""#])],
+        );
+    }
+
+    #[test]
+    fn parenthesized_expression() {
+        let source = dedent(
+            "
+            a = b + (
+                error: [undefined-reveal]
+                reveal_type(5)  # revealed: Literal[5]
+            )
+            ",
+        );
+        let reveal = source.find("reveal_type").unwrap();
+        let result = get_result(
+            &source,
+            vec![
+                TestDiagnostic::new("undefined-reveal", "msg", reveal),
+                TestDiagnostic::new("revealed-type", "Revealed type is `Literal[5]`", reveal),
+            ],
+        );
+
+        assert_ok(&result);
+    }
+
+    #[test]
+    fn bare_error_assertion_not_allowed() {
+        let source = "x  # error:";
+        let x = source.find('x').unwrap();
+        let result = get_result(
+            source,
+            vec![TestDiagnostic::new("some-rule", "some message", x)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "invalid assertion: no rule or message text",
+                    r#"unexpected error: 1 [some-rule] "some message""#,
+                ],
+            )],
+        );
+    }
+
+    #[test]
+    fn column_only_error_assertion_not_allowed() {
+        let source = "x  # error: 1";
+        let x = source.find('x').unwrap();
+        let result = get_result(
+            source,
+            vec![TestDiagnostic::new("some-rule", "some message", x)],
+        );
+
+        assert_fail(
+            result,
+            &[(
+                0,
+                &[
+                    "invalid assertion: no rule or message text",
+                    r#"unexpected error: 1 [some-rule] "some message""#,
+                ],
+            )],
+        );
+    }
+}

crates/red_knot_test/src/parser.rs

@@ -0,0 +1,576 @@
+use once_cell::sync::Lazy;
+use regex::{Captures, Regex};
+use ruff_index::{newtype_index, IndexVec};
+use rustc_hash::{FxHashMap, FxHashSet};
+
+/// Parse the Markdown `source` as a test suite with given `title`.
+pub(crate) fn parse<'s>(title: &'s str, source: &'s str) -> anyhow::Result<MarkdownTestSuite<'s>> {
+    let parser = Parser::new(title, source);
+    parser.parse()
+}
+
+/// A parsed markdown file containing tests.
+///
+/// Borrows from the source string and filepath it was created from.
+#[derive(Debug)]
+pub(crate) struct MarkdownTestSuite<'s> {
+    /// Header sections.
+    sections: IndexVec<SectionId, Section<'s>>,
+
+    /// Test files embedded within the Markdown file.
+    files: IndexVec<EmbeddedFileId, EmbeddedFile<'s>>,
+}
+
+impl<'s> MarkdownTestSuite<'s> {
+    pub(crate) fn tests(&self) -> MarkdownTestIterator<'_, 's> {
+        MarkdownTestIterator {
+            suite: self,
+            current_file_index: 0,
+        }
+    }
+}
+
+/// A single test inside a [`MarkdownTestSuite`].
+///
+/// A test is a single header section (or the implicit root section, if there are no Markdown
+/// headers in the file), containing one or more embedded Python files as fenced code blocks, and
+/// containing no nested header subsections.
+#[derive(Debug)]
+pub(crate) struct MarkdownTest<'m, 's> {
+    suite: &'m MarkdownTestSuite<'s>,
+    section: &'m Section<'s>,
+    files: &'m [EmbeddedFile<'s>],
+}
+
+impl<'m, 's> MarkdownTest<'m, 's> {
+    pub(crate) fn name(&self) -> String {
+        let mut name = String::new();
+        let mut parent_id = self.section.parent_id;
+        while let Some(next_id) = parent_id {
+            let parent = &self.suite.sections[next_id];
+            parent_id = parent.parent_id;
+            if !name.is_empty() {
+                name.insert_str(0, " - ");
+            }
+            name.insert_str(0, parent.title);
+        }
+        if !name.is_empty() {
+            name.push_str(" - ");
+        }
+        name.push_str(self.section.title);
+        name
+    }
+
+    pub(crate) fn files(&self) -> impl Iterator<Item = &'m EmbeddedFile<'s>> {
+        self.files.iter()
+    }
+}
+
+/// Iterator yielding all [`MarkdownTest`]s in a [`MarkdownTestSuite`].
+#[derive(Debug)]
+pub(crate) struct MarkdownTestIterator<'m, 's> {
+    suite: &'m MarkdownTestSuite<'s>,
+    current_file_index: usize,
+}
+
+impl<'m, 's> Iterator for MarkdownTestIterator<'m, 's> {
+    type Item = MarkdownTest<'m, 's>;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let mut current_file_index = self.current_file_index;
+        let mut file = self.suite.files.get(current_file_index.into());
+        let section_id = file?.section;
+        while file.is_some_and(|file| file.section == section_id) {
+            current_file_index += 1;
+            file = self.suite.files.get(current_file_index.into());
+        }
+        let files = &self.suite.files[EmbeddedFileId::from_usize(self.current_file_index)
+            ..EmbeddedFileId::from_usize(current_file_index)];
+        self.current_file_index = current_file_index;
+        Some(MarkdownTest {
+            suite: self.suite,
+            section: &self.suite.sections[section_id],
+            files,
+        })
+    }
+}
+
+#[newtype_index]
+struct SectionId;
+
+/// A single header section of a [`MarkdownTestSuite`], or the implicit root "section".
+///
+/// A header section is the part of a Markdown file beginning with a `#`-prefixed header line, and
+/// extending until the next header line at the same or higher outline level (that is, with the
+/// same number or fewer `#` characters).
+///
+/// A header section may either contain one or more embedded Python files (making it a
+/// [`MarkdownTest`]), or it may contain nested sections (headers with more `#` characters), but
+/// not both.
+#[derive(Debug)]
+struct Section<'s> {
+    title: &'s str,
+    level: u8,
+    parent_id: Option<SectionId>,
+}
+
+#[newtype_index]
+struct EmbeddedFileId;
+
+/// A single file embedded in a [`Section`] as a fenced code block.
+///
+/// Currently must be a Python file (`py` language) or type stub (`pyi`). In the future we plan
+/// support other kinds of files as well (TOML configuration, typeshed VERSIONS, `pth` files...).
+///
+/// A Python embedded file makes its containing [`Section`] into a [`MarkdownTest`], and will be
+/// type-checked and searched for inline-comment assertions to match against the diagnostics from
+/// type checking.
+#[derive(Debug)]
+pub(crate) struct EmbeddedFile<'s> {
+    section: SectionId,
+    pub(crate) path: &'s str,
+    pub(crate) lang: &'s str,
+    pub(crate) code: &'s str,
+}
+
+/// Matches an arbitrary amount of whitespace (including newlines), followed by a sequence of `#`
+/// characters, followed by a title heading, followed by a newline.
+static HEADER_RE: Lazy<Regex> =
+    Lazy::new(|| Regex::new(r"^(\s*\n)*(?<level>#+)\s+(?<title>.+)\s*\n").unwrap());
+
+/// Matches a code block fenced by triple backticks, possibly with language and `key=val`
+/// configuration items following the opening backticks (in the "tag string" of the code block).
+static CODE_RE: Lazy<Regex> = Lazy::new(|| {
+    Regex::new(r"^```(?<lang>\w+)(?<config>( +\S+)*)\s*\n(?<code>(.|\n)*?)\n```\s*\n").unwrap()
+});
+
+#[derive(Debug)]
+struct SectionStack(Vec<SectionId>);
+
+impl SectionStack {
+    fn new(root_section_id: SectionId) -> Self {
+        Self(vec![root_section_id])
+    }
+
+    fn push(&mut self, section_id: SectionId) {
+        self.0.push(section_id);
+    }
+
+    fn pop(&mut self) -> Option<SectionId> {
+        let popped = self.0.pop();
+        debug_assert_ne!(popped, None, "Should never pop the implicit root section");
+        debug_assert!(
+            !self.0.is_empty(),
+            "Should never pop the implicit root section"
+        );
+        popped
+    }
+
+    fn parent(&mut self) -> SectionId {
+        *self
+            .0
+            .last()
+            .expect("Should never pop the implicit root section")
+    }
+}
+
+/// Parse the source of a Markdown file into a [`MarkdownTestSuite`].
+#[derive(Debug)]
+struct Parser<'s> {
+    /// [`Section`]s of the final [`MarkdownTestSuite`].
+    sections: IndexVec<SectionId, Section<'s>>,
+
+    /// [`EmbeddedFile`]s of the final [`MarkdownTestSuite`].
+    files: IndexVec<EmbeddedFileId, EmbeddedFile<'s>>,
+
+    /// The unparsed remainder of the Markdown source.
+    unparsed: &'s str,
+
+    /// Stack of ancestor sections.
+    stack: SectionStack,
+
+    /// Names of embedded files in current active section.
+    current_section_files: Option<FxHashSet<&'s str>>,
+}
+
+impl<'s> Parser<'s> {
+    fn new(title: &'s str, source: &'s str) -> Self {
+        let mut sections = IndexVec::default();
+        let root_section_id = sections.push(Section {
+            title,
+            level: 0,
+            parent_id: None,
+        });
+        Self {
+            sections,
+            files: IndexVec::default(),
+            unparsed: source,
+            stack: SectionStack::new(root_section_id),
+            current_section_files: None,
+        }
+    }
+
+    fn parse(mut self) -> anyhow::Result<MarkdownTestSuite<'s>> {
+        self.parse_impl()?;
+        Ok(self.finish())
+    }
+
+    fn finish(mut self) -> MarkdownTestSuite<'s> {
+        self.sections.shrink_to_fit();
+        self.files.shrink_to_fit();
+
+        MarkdownTestSuite {
+            sections: self.sections,
+            files: self.files,
+        }
+    }
+
+    fn parse_impl(&mut self) -> anyhow::Result<()> {
+        while !self.unparsed.is_empty() {
+            if let Some(captures) = self.scan(&HEADER_RE) {
+                self.parse_header(&captures)?;
+            } else if let Some(captures) = self.scan(&CODE_RE) {
+                self.parse_code_block(&captures)?;
+            } else {
+                // ignore other Markdown syntax (paragraphs, etc) used as comments in the test
+                if let Some(next_newline) = self.unparsed.find('\n') {
+                    (_, self.unparsed) = self.unparsed.split_at(next_newline + 1);
+                } else {
+                    break;
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    fn parse_header(&mut self, captures: &Captures<'s>) -> anyhow::Result<()> {
+        let header_level = captures["level"].len();
+        self.pop_sections_to_level(header_level);
+
+        let parent = self.stack.parent();
+
+        let section = Section {
+            // HEADER_RE can't match without a match for group 'title'.
+            title: captures.name("title").unwrap().into(),
+            level: header_level.try_into()?,
+            parent_id: Some(parent),
+        };
+
+        if self.current_section_files.is_some() {
+            return Err(anyhow::anyhow!(
+                "Header '{}' not valid inside a test case; parent '{}' has code files.",
+                section.title,
+                self.sections[parent].title,
+            ));
+        }
+
+        let section_id = self.sections.push(section);
+        self.stack.push(section_id);
+
+        self.current_section_files = None;
+
+        Ok(())
+    }
+
+    fn parse_code_block(&mut self, captures: &Captures<'s>) -> anyhow::Result<()> {
+        // We never pop the implicit root section.
+        let parent = self.stack.parent();
+
+        let mut config: FxHashMap<&'s str, &'s str> = FxHashMap::default();
+
+        if let Some(config_match) = captures.name("config") {
+            for item in config_match.as_str().split_whitespace() {
+                let mut parts = item.split('=');
+                let key = parts.next().unwrap();
+                let Some(val) = parts.next() else {
+                    return Err(anyhow::anyhow!("Invalid config item `{}`.", item));
+                };
+                if parts.next().is_some() {
+                    return Err(anyhow::anyhow!("Invalid config item `{}`.", item));
+                }
+                if config.insert(key, val).is_some() {
+                    return Err(anyhow::anyhow!("Duplicate config item `{}`.", item));
+                }
+            }
+        }
+
+        let path = config.get("path").copied().unwrap_or("test.py");
+
+        self.files.push(EmbeddedFile {
+            path,
+            section: parent,
+            // CODE_RE can't match without matches for 'lang' and 'code'.
+            lang: captures.name("lang").unwrap().into(),
+            code: captures.name("code").unwrap().into(),
+        });
+
+        if let Some(current_files) = &mut self.current_section_files {
+            if !current_files.insert(path) {
+                if path == "test.py" {
+                    return Err(anyhow::anyhow!(
+                        "Test `{}` has duplicate files named `{path}`. \
+                                (This is the default filename; \
+                                 consider giving some files an explicit name with `path=...`.)",
+                        self.sections[parent].title
+                    ));
+                }
+                return Err(anyhow::anyhow!(
+                    "Test `{}` has duplicate files named `{path}`.",
+                    self.sections[parent].title
+                ));
+            };
+        } else {
+            self.current_section_files = Some(FxHashSet::from_iter([path]));
+        }
+
+        Ok(())
+    }
+
+    fn pop_sections_to_level(&mut self, level: usize) {
+        while level <= self.sections[self.stack.parent()].level.into() {
+            self.stack.pop();
+            // We would have errored before pushing a child section if there were files, so we know
+            // no parent section can have files.
+            self.current_section_files = None;
+        }
+    }
+
+    /// Get capture groups and advance cursor past match if unparsed text matches `pattern`.
+    fn scan(&mut self, pattern: &Regex) -> Option<Captures<'s>> {
+        if let Some(captures) = pattern.captures(self.unparsed) {
+            let (_, unparsed) = self.unparsed.split_at(captures.get(0).unwrap().end());
+            self.unparsed = unparsed;
+            Some(captures)
+        } else {
+            None
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use ruff_python_trivia::textwrap::dedent;
+
+    #[test]
+    fn empty() {
+        let mf = super::parse("file.md", "").unwrap();
+
+        assert!(mf.tests().next().is_none());
+    }
+
+    #[test]
+    fn single_file_test() {
+        let source = dedent(
+            "
+            ```py
+            x = 1
+            ```
+            ",
+        );
+        let mf = super::parse("file.md", &source).unwrap();
+
+        let [test] = &mf.tests().collect::<Vec<_>>()[..] else {
+            panic!("expected one test");
+        };
+
+        assert_eq!(test.name(), "file.md");
+
+        let [file] = test.files().collect::<Vec<_>>()[..] else {
+            panic!("expected one file");
+        };
+
+        assert_eq!(file.path, "test.py");
+        assert_eq!(file.lang, "py");
+        assert_eq!(file.code, "x = 1");
+    }
+
+    #[test]
+    fn multiple_tests() {
+        let source = dedent(
+            "
+            # One
+
+            ```py
+            x = 1
+            ```
+
+            # Two
+
+            ```py
+            y = 2
+            ```
+            ",
+        );
+        let mf = super::parse("file.md", &source).unwrap();
+
+        let [test1, test2] = &mf.tests().collect::<Vec<_>>()[..] else {
+            panic!("expected two tests");
+        };
+
+        assert_eq!(test1.name(), "file.md - One");
+        assert_eq!(test2.name(), "file.md - Two");
+
+        let [file] = test1.files().collect::<Vec<_>>()[..] else {
+            panic!("expected one file");
+        };
+
+        assert_eq!(file.path, "test.py");
+        assert_eq!(file.lang, "py");
+        assert_eq!(file.code, "x = 1");
+
+        let [file] = test2.files().collect::<Vec<_>>()[..] else {
+            panic!("expected one file");
+        };
+
+        assert_eq!(file.path, "test.py");
+        assert_eq!(file.lang, "py");
+        assert_eq!(file.code, "y = 2");
+    }
+
+    #[test]
+    fn custom_file_path() {
+        let source = dedent(
+            "
+            ```py path=foo.py
+            x = 1
+            ```
+            ",
+        );
+        let mf = super::parse("file.md", &source).unwrap();
+
+        let [test] = &mf.tests().collect::<Vec<_>>()[..] else {
+            panic!("expected one test");
+        };
+        let [file] = test.files().collect::<Vec<_>>()[..] else {
+            panic!("expected one file");
+        };
+
+        assert_eq!(file.path, "foo.py");
+        assert_eq!(file.lang, "py");
+        assert_eq!(file.code, "x = 1");
+    }
+
+    #[test]
+    fn multi_line_file() {
+        let source = dedent(
+            "
+            ```py
+            x = 1
+            y = 2
+            ```
+            ",
+        );
+        let mf = super::parse("file.md", &source).unwrap();
+
+        let [test] = &mf.tests().collect::<Vec<_>>()[..] else {
+            panic!("expected one test");
+        };
+        let [file] = test.files().collect::<Vec<_>>()[..] else {
+            panic!("expected one file");
+        };
+
+        assert_eq!(file.code, "x = 1\ny = 2");
+    }
+
+    #[test]
+    fn no_header_inside_test() {
+        let source = dedent(
+            "
+            # One
+
+            ```py
+            x = 1
+            ```
+
+            ## Two
+            ",
+        );
+        let err = super::parse("file.md", &source).expect_err("Should fail to parse");
+        assert_eq!(
+            err.to_string(),
+            "Header 'Two' not valid inside a test case; parent 'One' has code files."
+        );
+    }
+
+    #[test]
+    fn invalid_config_item_no_equals() {
+        let source = dedent(
+            "
+            ```py foo
+            x = 1
+            ```
+            ",
+        );
+        let err = super::parse("file.md", &source).expect_err("Should fail to parse");
+        assert_eq!(err.to_string(), "Invalid config item `foo`.");
+    }
+
+    #[test]
+    fn invalid_config_item_too_many_equals() {
+        let source = dedent(
+            "
+            ```py foo=bar=baz
+            x = 1
+            ```
+            ",
+        );
+        let err = super::parse("file.md", &source).expect_err("Should fail to parse");
+        assert_eq!(err.to_string(), "Invalid config item `foo=bar=baz`.");
+    }
+
+    #[test]
+    fn invalid_config_item_duplicate() {
+        let source = dedent(
+            "
+            ```py foo=bar foo=baz
+            x = 1
+            ```
+            ",
+        );
+        let err = super::parse("file.md", &source).expect_err("Should fail to parse");
+        assert_eq!(err.to_string(), "Duplicate config item `foo=baz`.");
+    }
+
+    #[test]
+    fn no_duplicate_name_files_in_test() {
+        let source = dedent(
+            "
+            ```py
+            x = 1
+            ```
+
+            ```py
+            y = 2
+            ```
+            ",
+        );
+        let err = super::parse("file.md", &source).expect_err("Should fail to parse");
+        assert_eq!(
+            err.to_string(),
+            "Test `file.md` has duplicate files named `test.py`. \
+            (This is the default filename; consider giving some files an explicit name \
+            with `path=...`.)"
+        );
+    }
+
+    #[test]
+    fn no_duplicate_name_files_in_test_non_default() {
+        let source = dedent(
+            "
+            ```py path=foo.py
+            x = 1
+            ```
+
+            ```py path=foo.py
+            y = 2
+            ```
+            ",
+        );
+        let err = super::parse("file.md", &source).expect_err("Should fail to parse");
+        assert_eq!(
+            err.to_string(),
+            "Test `file.md` has duplicate files named `foo.py`."
+        );
+    }
+}

crates/red_knot_vendored/Cargo.toml

@@ -0,0 +1,32 @@
+[package]
+name = "red_knot_vendored"
+version = "0.0.0"
+publish = false
+authors = { workspace = true }
+edition = { workspace = true }
+rust-version = { workspace = true }
+homepage = { workspace = true }
+documentation = { workspace = true }
+repository = { workspace = true }
+license = { workspace = true }
+
+[dependencies]
+ruff_db = { workspace = true }
+once_cell = { workspace = true }
+zip = { workspace = true }
+
+[build-dependencies]
+path-slash = { workspace = true }
+walkdir = { workspace = true }
+zip = { workspace = true, features = ["zstd", "deflate"] }
+
+[dev-dependencies]
+walkdir = { workspace = true }
+
+[features]
+zstd = ["zip/zstd"]
+deflate = ["zip/deflate"]
+
+[lints]
+workspace = true
+

crates/red_knot_vendored/README.md

@@ -1,9 +1,5 @@
-# Red Knot
+# Vendored types for the stdlib
 
-Semantic analysis for the red-knot project.
-
-## Vendored types for the stdlib
-
-This crate vendors [typeshed](https://github.com/python/typeshed)'s stubs for the standard library. The vendored stubs can be found in `crates/red_knot_python_semantic/vendor/typeshed`. The file `crates/red_knot_python_semantic/vendor/typeshed/source_commit.txt` tells you the typeshed commit that our vendored stdlib stubs currently correspond to.
+This crate vendors [typeshed](https://github.com/python/typeshed)'s stubs for the standard library. The vendored stubs can be found in `crates/red_knot_vendored/vendor/typeshed`. The file `crates/red_knot_vendored/vendor/typeshed/source_commit.txt` tells you the typeshed commit that our vendored stdlib stubs currently correspond to.
 
 The typeshed stubs are updated every two weeks via an automated PR using the `sync_typeshed.yaml` workflow in the `.github/workflows` directory. This workflow can also be triggered at any time via [workflow dispatch](https://docs.github.com/en/actions/using-workflows/manually-running-a-workflow#running-a-workflow).

crates/red_knot_vendored/build.rs

@@ -0,0 +1,89 @@
+//! Build script to package our vendored typeshed files
+//! into a zip archive that can be included in the Ruff binary.
+//!
+//! This script should be automatically run at build time
+//! whenever the script itself changes, or whenever any files
+//! in `crates/red_knot_vendored/vendor/typeshed` change.
+
+use std::fs::File;
+use std::path::Path;
+
+use path_slash::PathExt;
+use zip::result::ZipResult;
+use zip::write::{FileOptions, ZipWriter};
+use zip::CompressionMethod;
+
+const TYPESHED_SOURCE_DIR: &str = "vendor/typeshed";
+const TYPESHED_ZIP_LOCATION: &str = "/zipped_typeshed.zip";
+
+/// Recursively zip the contents of an entire directory.
+///
+/// This routine is adapted from a recipe at
+/// <https://github.com/zip-rs/zip-old/blob/5d0f198124946b7be4e5969719a7f29f363118cd/examples/write_dir.rs>
+fn zip_dir(directory_path: &str, writer: File) -> ZipResult<File> {
+    let mut zip = ZipWriter::new(writer);
+
+    // Use deflated compression for WASM builds because compiling `zstd-sys` requires clang
+    // [source](https://github.com/gyscos/zstd-rs/wiki/Compile-for-WASM) which complicates the build
+    // by a lot. Deflated compression is slower but it shouldn't matter much for the WASM use case
+    // (WASM itself is already slower than a native build for a specific platform).
+    // We can't use `#[cfg(...)]` here because the target-arch in a build script is the
+    // architecture of the system running the build script and not the architecture of the build-target.
+    // That's why we use the `TARGET` environment variable here.
+    let method = if cfg!(feature = "zstd") {
+        CompressionMethod::Zstd
+    } else if cfg!(feature = "deflate") {
+        CompressionMethod::Deflated
+    } else {
+        CompressionMethod::Stored
+    };
+
+    let options = FileOptions::default()
+        .compression_method(method)
+        .unix_permissions(0o644);
+
+    for entry in walkdir::WalkDir::new(directory_path) {
+        let dir_entry = entry.unwrap();
+        let absolute_path = dir_entry.path();
+        let normalized_relative_path = absolute_path
+            .strip_prefix(Path::new(directory_path))
+            .unwrap()
+            .to_slash()
+            .expect("Unexpected non-utf8 typeshed path!");
+
+        // Write file or directory explicitly
+        // Some unzip tools unzip files with directory paths correctly, some do not!
+        if absolute_path.is_file() {
+            println!("adding file {absolute_path:?} as {normalized_relative_path:?} ...");
+            zip.start_file(normalized_relative_path, options)?;
+            let mut f = File::open(absolute_path)?;
+            std::io::copy(&mut f, &mut zip).unwrap();
+        } else if !normalized_relative_path.is_empty() {
+            // Only if not root! Avoids path spec / warning
+            // and mapname conversion failed error on unzip
+            println!("adding dir {absolute_path:?} as {normalized_relative_path:?} ...");
+            zip.add_directory(normalized_relative_path, options)?;
+        }
+    }
+    zip.finish()
+}
+
+fn main() {
+    println!("cargo:rerun-if-changed={TYPESHED_SOURCE_DIR}");
+    assert!(
+        Path::new(TYPESHED_SOURCE_DIR).is_dir(),
+        "Where is typeshed?"
+    );
+    let out_dir = std::env::var("OUT_DIR").unwrap();
+
+    // N.B. Deliberately using `format!()` instead of `Path::join()` here,
+    // so that we use `/` as a path separator on all platforms.
+    // That enables us to load the typeshed zip at compile time in `module.rs`
+    // (otherwise we'd have to dynamically determine the exact path to the typeshed zip
+    // based on the default path separator for the specific platform we're on,
+    // which can't be done at compile time.)
+    let zipped_typeshed_location = format!("{out_dir}{TYPESHED_ZIP_LOCATION}");
+
+    let zipped_typeshed = File::create(zipped_typeshed_location).unwrap();
+    zip_dir(TYPESHED_SOURCE_DIR, zipped_typeshed).unwrap();
+}

crates/red_knot_vendored/src/lib.rs

@@ -6,7 +6,7 @@ use ruff_db::vendored::VendoredFileSystem;
 // Luckily this crate will fail to build if this file isn't available at build time.
 static TYPESHED_ZIP_BYTES: &[u8] = include_bytes!(concat!(env!("OUT_DIR"), "/zipped_typeshed.zip"));
 
-pub fn vendored_typeshed_stubs() -> &'static VendoredFileSystem {
+pub fn file_system() -> &'static VendoredFileSystem {
     static VENDORED_TYPESHED_STUBS: Lazy<VendoredFileSystem> =
         Lazy::new(|| VendoredFileSystem::new_static(TYPESHED_ZIP_BYTES).unwrap());
     &VENDORED_TYPESHED_STUBS
@@ -42,7 +42,7 @@ mod tests {
     #[test]
     fn typeshed_vfs_consistent_with_vendored_stubs() {
         let vendored_typeshed_dir = Path::new("vendor/typeshed").canonicalize().unwrap();
-        let vendored_typeshed_stubs = vendored_typeshed_stubs();
+        let vendored_typeshed_stubs = file_system();
 
         let mut empty_iterator = true;
         for entry in walkdir::WalkDir::new(&vendored_typeshed_dir).min_depth(1) {

crates/red_knot_vendored/vendor/typeshed/source_commit.txt

@@ -0,0 +1 @@
+91a58b07cdd807b1d965e04ba85af2adab8bf924

crates/red_knot_vendored/vendor/typeshed/stdlib/VERSIONS

@@ -161,6 +161,8 @@ importlib.metadata._meta: 3.10-
 importlib.metadata.diagnose: 3.13-
 importlib.readers: 3.10-
 importlib.resources: 3.7-
+importlib.resources._common: 3.11-
+importlib.resources._functional: 3.13-
 importlib.resources.abc: 3.11-
 importlib.resources.readers: 3.11-
 importlib.resources.simple: 3.11-