[psteinroe]

Stacked on #928.

Adds a [workspace.lints] table and opts every member crate into it ([lints] workspace = true, also in the just new-crate template), and turns on the panic lints for pgls_lsp, the crate that handles client input first.

Workspace lints

clippy::todo, clippy::unimplemented and clippy::dbg_macro are now denied everywhere (dbg! stays allowed in tests, as before). The existing occurrences are fixed, not allowed:

• MemoryFileSystem::resolve_configuration was a todo!(), so a configuration that extends a package name (not a path) panicked on the memory file system, which the WASM build uses. It returns ResolveError::NotFound now, which is reported as "can't resolve" like an unresolvable package on disk.
• The TypeScript bindings generator had unimplemented!() for array and object literals and for boolean schemas. It now prints [a, b], { "k": v }, unknown and never. The generated bindings don't change.
• generate_tool_configuration returns an error for assists and the transformation category instead of panicking. unreachable!() got a message.

pgls_lsp: no unwrap, expect or indexing

pgls_lsp/src/lib.rs warns on clippy::unwrap_used, clippy::expect_used and clippy::indexing_slicing, and CI's -D warnings turns them into errors. Tests are exempt (allow-*-in-tests in clippy.toml). On top of #928, which already removed the params.arguments[0] indexing, there were 21 findings (19 unwrap, 2 expect):

• 12 were std::sync lock unwraps. pgls_lsp uses parking_lot locks now, which can't be poisoned. Before, a request handler that panicked while holding the documents lock (panic_to_lsp_error catches the panic) poisoned the lock, and every later request on that session panicked too.
• get_cursor_position and completions expected/unwrapped the client capabilities. They use Session::position_encoding(), which falls back to UTF-16 like the other handlers.
• A completion item whose range can't be converted is left out with a warning, instead of panicking the whole completion request.
• diagnostic.advices(..) errors are returned through ?. Serialising the executeStatement arguments drops the action on failure. The unused print_markup (two unwraps) is removed.
• Two can't fail and are #[expect(..., reason = "...")]: the u32 offset in LineIndex::new (documents over 4 GiB aren't supported anywhere) and the ConfigurationStatus atomic, which only ever stores a ConfigurationStatus as u8.

Counts for the next crate

pgls_workspace (non-test code) has 76 findings: 48 unwrap, 5 expect, 16 indexing and 7 slicing. About 34 of the unwraps are lock unwraps, which parking_lot would remove the same way. matcher/pattern.rs (20, the glob matcher's index arithmetic) and server.rs (16) have most of the rest. For reference, the same measurement over the other crates found 178 in pgls_configuration (mostly generated code), 89 in pgls_pretty_print, 70 in pgls_catalog and 39 in pgls_treesitter.

Risks

Open PRs that add std::sync locks with .unwrap() to pgls_lsp (#895, #908) will fail clippy once this is merged. The fix is to use parking_lot::{Mutex, RwLock} and drop the .unwrap(). client_capabilities() stays because #908, #911 and #919 use it.

Tests: cargo clippy --workspace --all-targets --all-features is clean, pgls_lsp and pgls_fs tests pass, and just gen-lint leaves the generated files unchanged.

Follow-ups: #895, #908 and #936 now use parking_lot locks. Only #881's own projects and workspace_folders locks remain to be switched when the two meet.