perf: per-root re-analysis (remove IncludeOnly pass)

[pfitzseb]

Motivation

Profiling showed that a single keystroke forced a full re-lint of the entire
workspace, once per root. Three compounding causes:

1. The include layer defeated Salsa's early-exit on every edit.
   derived_all_includes returned a tuple of the include graph and a global
   Dict{objectid(EXPR) → URI} used by the semantic pass to resolve
   include() calls. Objectids change on every reparse, so this value never
   compared equal after any edit — backdating never fired, and everything
   downstream (derived_roots, derived_roots_for_uri,
   derived_is_indirect_file, every lint meta) invalidated even when the
   include structure was untouched. derived_roots_for_uri additionally
   re-ran a per-root BFS on every navigation request.
2. Lint state was monolithic per root but global in its dependencies.
   derived_static_lint_meta_for_root seeded meta and ran check_all +
   resolve_remaining_getfields! over every Julia file in the workspace for
   every root. That made the work O(files × roots) — and the results for
   files outside a root's include closure were pure waste, since the
   diagnostics pass only ever reads the closure. It also meant any edit
   anywhere invalidated every root's meta.
3. Diagnostics were one global node. derived_static_lint_all_diagnostics
   looped all roots, so any edit re-ran collect_hints for every root's
   closure.

Measured baseline (LanguageServer package, 57 files, 12 roots, warm caches):
~35 ms per root per keystroke, i.e. every root paid the full price on every
edit regardless of where the edit was.

Implementation

Commit 1 — include graph from per-file records. The graph
(derived_uri2included) is now built by a worklist over the existing
per-file, value-stable derived_file_include_records — its value only changes
when include structure changes, so early-exit protects all graph consumers
on ordinary edits. The objectid dict is gone from the Salsa layer entirely:
semantic_pass/followinclude build an objectid→target map per entered file
at traversal time from a shared walker (which also replaces the duplicated
IncludeOnly traversal — one include-detection implementation instead of
two). One behavioral edge preserved and pinned by test: files without a
filesystem path (unsaved buffers) still resolve absolute include paths.

Commit 2 — lint meta scoped to the include closure. New
derived_include_closure(rt, uri) (transitive closure over the stable graph).
Meta seeding, check_all, and resolve_remaining_getfields! — in both
derived_static_lint_meta_for_root and derived_deved_package_meta — now
cover the closure instead of all files. Each root's lint state depends only on
its own closure's CSTs, so edits outside it re-verify from cache.

Commit 3 — per-root diagnostics. derived_static_lint_all_diagnostics
replaced by per-root derived_static_lint_diagnostics_for_root; the per-file
query unions results from derived_roots_for_uri, preserving the cross-root
Set dedup. Iterating the closure instead of the old ad-hoc worklist also
fixed a latent non-termination: the old walk had no visited-set, so an include
cycle inside a project would loop forever (pinned by a new test).

Results

Per-keystroke lint cost went from O(workspace files × roots) to O(edited
root's closure). Measured: editing a file recomputes its own root in ~33 ms
(unchanged, as expected — the closure spans the whole src tree), while an
unrelated root drops from ~35 ms to ~0.1 ms; derived_roots re-verifies
in ~1 ms and derived_roots_for_uri serves from cache in ~4 µs. Behavior is
otherwise identical: full suite green (3658 passed), including all 585
staticlint tests, plus 22 new tests in test/test_includes.jl covering graph
shape, value stability under include-preserving edits, closure computation,
cross-file resolution after edits, the unsaved-buffer edge, and cycle
termination.

---

@davidanthoff I know you added the separate IncludeOnly pass for a reason, but I can't quite remember the reasoning. derived_includes is still around, so maybe that's good enough?