chore(build): upgrade toolchain to clang 22 and LLVM prebuilt r2 (#550)

## Summary

Toolchain upgrade:

- Compiler toolchain: clang 20.1.8 → 22.1.8, GCC 14.2.0 → 15.1.0
- LLVM prebuilt: 21.1.8+r1 → 21.1.8+r2 (rebuilt with the clang 22
toolchain)
- build-llvm workflow: DIA SDK off, patches applied from clice-llvm,
windows-2022 runner

macOS dynamic-linking fixes — libc++ ≥ 21 moved string hashing into the
dylib as `__hash_memory` (llvm/llvm-project#77653), which surfaced two
latent issues in our setup:

- Strip `_LIBCPP_DISABLE_AVAILABILITY` from the kotatsu targets. With
libc++ ≥ 21 headers that macro makes the headers emit references to
dylib-only symbols the macOS system libc++ does not ship, which broke
the x64 cross link and made binaries depend on conda's libc++ at
runtime. Temporary workaround until the flag is removed upstream in
kotatsu.
- Pass `--no-default-config` to clang on Apple, via the toolchain file
so the LLVM prebuilt builds inherit it. conda-forge clang 22's bundled
config files inject `-L`/`-rpath` pointing into the conda env at link
time, binding binaries to `@rpath/libc++.1.dylib` with an rpath valid
only on the build machine — the VSCode E2E job caught them failing to
load anywhere else. With config files disabled, binaries link the system
libc++, gated safely by availability annotations.
- Add a Debug-only rpath to the env's libc++: the prebuilt ASan dylibs
still reference conda's libc++, and Debug binaries are CI-internal. Goes
away with the next prebuilt respin.
- Cleanups: drop the redundant availability defines and the dead macOS
cross activation script/env.

## Test plan

- [x] Full CI matrix green: Linux/Windows/macOS native, all three cross
pairs, editor E2E on both platforms
- [x] macOS release artifacts link `/usr/lib/libc++.1.dylib` with no
conda env rpath (verified via `otool -L`)

feat(driver): one subcommand per translation unit (#539)

## Why

Upcoming features (`clice lint` with cross-TU caching, ahead-of-time
indexing, documentation data generation) are all CLI applications over
the same core, but the CLI surface lived as one monolithic `clice.cc`.
This PR gives every subcommand its own home before those features land.

## What

- `src/driver/` holds one translation unit per subcommand: `serve`,
`query`, `worker`, plus new stubs `index`, `doc`, `lint`, `format`. Each
file owns its options struct and registers a deco command with the root
`SubCommander` through an `add_*` builder.
- `clice.cc` shrinks to collection and registration (206 → ~60 lines).
Behavior parity everywhere: dispatch, `--version`, root/per-command
`--help`, missing/unknown subcommand handling, worker defaults, query
port validation.
- Stubs print a one-line self-description to stderr and exit non-zero,
so scripts can detect the command is unavailable while humans learn what
it will do.
- Commands are constructed inside the builders and moved into the
`SubCommander` (owning overload); help re-constructs a fresh command for
usage printing, so no closure captures a moved-from object.
- New `tests/integration/lifecycle/test_subcommands.py` pins the whole
surface: all seven names registered in root usage, stub messages and
exit codes, per-command help, unknown-subcommand failure.

## Tests

- Unit: 1059; Integration: 304 (incl. 4 new CLI surface tests); Smoke:
3/3.

feat(tests): generate folding-ranges doc from snapshot fixtures (#535)

## Why

Feature docs are hand-maintained support checklists with no mechanical
connection to tests: nothing verifies that a checked item actually
works, and nothing flags when behavior changes. This PR pilots the
inversion for folding ranges: snapshot fixtures become the single source
of truth and the doc checklist is generated from them.

## Design

Each fixture under `tests/data/folding_range/<section>/` carries a doc
header that is a plain markdown document behind `///`:

```cpp
/// # Preprocessor conditional folding (`#if` / `#ifdef` ...)
///
/// - status: partial
/// - issues: clangd#1661, clangd#2059
///
/// Branch regions delimited by `#else` fold today; a bare
/// `#if ... #endif` block without an `#else` does not fold yet.

<example code — doubles as the doc example>
```

- The h1 title marks the file as a doc item (files without it are
supplementary tests, excluded from docs); stripped of the `///` prefix
the whole header renders as-is in any markdown viewer.
- Grouping comes from one-level section subdirectories (`fold_kinds/`,
`refinements/`); generated doc regions are keyed by the directory name,
while human headings and prose stay hand-written.
- `status` semantics: `supported` → `[x]`, compiled, snapshot proves the
behavior; `partial` → `[ ]` with a _(partial)_ marker, compiled,
snapshot records the current partial behavior so improvements surface as
diffs; `unsupported` → skipped by the snapshot glob, snapshot pinned to
`UNSUPPORTED`, implementing the capability later surfaces as a diff
prompting the status flip.
- `tests/tools/feature_docs.py` renders headers into marker-delimited
regions of `docs/en/features/folding-ranges.md` (`update`) and fails on
drift (`check`, now part of `pixi run unit-test`). The C++ side reads
only the `status` key via a shared ten-line helper — there is no grammar
to drift between the two parsers.

## Notable corrections found by review

- Two preprocessor fixtures initially marked unsupported actually fold
`#else`-delimited branches today; they are now `partial` with snapshots
recording the true current behavior (bare `#if...#endif` still does not
fold — clangd#1661). The `partial` tier exists because of this finding.
- All supported items' snapshots were verified to demonstrate their
claimed capability.

## Migration

The 14 checklist items of the old hand-written doc are preserved
(titles, statuses, examples, issue links, client-support notes). Legacy
inline test cases are untouched and migrate in follow-up per-feature
PRs.

## Tests

- Unit: 1052 passed on both RelWithDebInfo and Debug, including the
doc-sync check
- Integration: 300 passed; Smoke: 3/3
- `format` → `check` round-trip verified stable

feat(tests): generate folding-ranges doc from snapshot fixtures (#535)

## Why

Feature docs are hand-maintained support checklists with no mechanical
connection to tests: nothing verifies that a checked item actually
works, and nothing flags when behavior changes. This PR pilots the
inversion for folding ranges: snapshot fixtures become the single source
of truth and the doc checklist is generated from them.

## Design

Each fixture under `tests/data/folding_range/<section>/` carries a doc
header that is a plain markdown document behind `///`:

```cpp
/// # Preprocessor conditional folding (`#if` / `#ifdef` ...)
///
/// - status: partial
/// - issues: clangd#1661, clangd#2059
///
/// Branch regions delimited by `#else` fold today; a bare
/// `#if ... #endif` block without an `#else` does not fold yet.

<example code — doubles as the doc example>
```

- The h1 title marks the file as a doc item (files without it are
supplementary tests, excluded from docs); stripped of the `///` prefix
the whole header renders as-is in any markdown viewer.
- Grouping comes from one-level section subdirectories (`fold_kinds/`,
`refinements/`); generated doc regions are keyed by the directory name,
while human headings and prose stay hand-written.
- `status` semantics: `supported` → `[x]`, compiled, snapshot proves the
behavior; `partial` → `[ ]` with a _(partial)_ marker, compiled,
snapshot records the current partial behavior so improvements surface as
diffs; `unsupported` → skipped by the snapshot glob, snapshot pinned to
`UNSUPPORTED`, implementing the capability later surfaces as a diff
prompting the status flip.
- `tests/tools/feature_docs.py` renders headers into marker-delimited
regions of `docs/en/features/folding-ranges.md` (`update`) and fails on
drift (`check`, now part of `pixi run unit-test`). The C++ side reads
only the `status` key via a shared ten-line helper — there is no grammar
to drift between the two parsers.

## Notable corrections found by review

- Two preprocessor fixtures initially marked unsupported actually fold
`#else`-delimited branches today; they are now `partial` with snapshots
recording the true current behavior (bare `#if...#endif` still does not
fold — clangd#1661). The `partial` tier exists because of this finding.
- All supported items' snapshots were verified to demonstrate their
claimed capability.

## Migration

The 14 checklist items of the old hand-written doc are preserved
(titles, statuses, examples, issue links, client-support notes). Legacy
inline test cases are untouched and migrate in follow-up per-feature
PRs.

## Tests

- Unit: 1052 passed on both RelWithDebInfo and Debug, including the
doc-sync check
- Integration: 300 passed; Smoke: 3/3
- `format` → `check` round-trip verified stable

refactor(tests): replace shared corpus with per-feature snapshot dirs (…

…#534)

## Why

The shared snapshot corpus (`tests/corpus/`) ran every corpus file
through five unrelated feature suites (folding_range, semantic_tokens,
document_symbol, inlay_hint, tu_index). Language constructs do not map
1:1 across features, so each corpus file generated noise snapshots for
features that do not care about it, making snapshot review scale badly.

## What

- Remove the shared corpus mechanism: `tests/corpus/`, the
`--corpus-dir` CLI option, the `corpus_dir` global, and its references
in `pixi.toml` and the docs.
- Re-point each of the five snapshot glob tests at its own curated
directory `tests/data/<feature>/`, following the existing
`tests/data/hover/` pattern. The dump/transform logic is kept unchanged
for reuse in the upcoming per-feature snapshot migration.
- Seed each directory with a feature-tailored `basic.cpp` plus a
`DisableFormat: true` `.clang-format` (fixture layout is part of the
tested input; formatting would invalidate line/column-sensitive
snapshots).
- Regenerate snapshots; each seed produces richer output for its feature
than the old shared corpus file did.

## Tests

- Unit: 1026 passed (12 pre-existing skips)
- Integration: 300 passed
- Smoke: 3/3 passed

fix(vscode): interim display name pending marketplace (#529)

Microsoft's Marketplace policy permanently reserves deleted extension
display names, which currently blocks publishing `clice-io.clice` with
the display name `clice` (held by the removed predecessor extension).
Use `clice vscode` as an interim display name so the extension can ship;
a release request for the original name is pending with Marketplace
support. Display-name changes are plain updates — the extension ID stays
`clice-io.clice`, so this is fully reversible once the name is released.