Repository navigation
fix(kernel): finish the pinned resolution port and the three regressions it exposed - #582
Merged
Merged
Conversation
A lone same-file definition chosen by findBestMatch scored by directory proximity (0.7/0.4), losing to a framework's naming guess in another file. is_decisive_same_file_match scores it 0.9 when it is the only same-file candidate, is not the caller itself, and, as a member, belongs to an owner the caller is inside. Ports upstream cde4cd7 (colbymchenry#1938). same-file-match-confidence: 7/7.
The resolver is not dropped before process exit on the CLI index path, so CODEGRAPH_KERNEL_PROF=1 accumulated timings that never printed. close() is the deterministic teardown point.
Upstream's kernel defers a file whose parse tree has errors, so the test expected tryKernelExtract to return null. The fork's tsjs walker extracts the recovered tree natively (cba9256) and keeps the re-export signature as written — assert that contract instead.
An alias, workspace or direct-path join kept a specifier's lexical spelling, so `@/widgets/Chart/` probed `Chart//index.tsx` and `src/widgets/./Chart` probed `./` segments no files row is keyed by. probe_extensions now spells each candidate as the index keys it for the ES-module family; other languages keep the path as found. Ports upstream 8651990 (colbymchenry#2792). import-path-spelling: 5/5.
A Dart `_name` is private to its library — the library file and every file its `part` directives name — not to its file. private_is_file_local drops `dart`; is_visible_across_files gets the dart arm (a private candidate is seen only from a file of its own library); and the exact-name pre-pick filter declines a private candidate another library declares so the caller's own library's declaration wins. Ports upstream f7820b5 (colbymchenry#2832). dart-part-privacy: 10/12 — the two remaining failures resolve a typed receiver's private method through resolve_method_on_type's member path, whose visibility gate still needs tracing (upstream's dart arm at name-matcher.ts ~13147 / preferOwnDartLibrary is the next suspect).
finish() applied name_post_guard but not the file/library-locality veto resolveOne's nameResult gets at index.ts:1867, so typed-receiver picks (instance-method) linked private members of another library. With the dart arm in place this completes f7820b5: dart-part-privacy 12/12.
…stored The file lock was released as soon as indexing started — the promise returned unawaited, the try ended, the finally let the lock go. Awaited, the lock now holds until the files are stored, as indexAll and sync hold it (upstream 865ddf4 / colbymchenry#2845).
settleKernelOutcome merged kernel candidates beside the frameworks', so a project namesake could answer a call Nuxt's navigateTo owns (upstream 58eb57a / colbymchenry#2826). With ownsReference set, the frameworks' best answer wins or the ref stays unresolved.
…ompiled in Upstream's test asserted wasm-preload sets (9ee6718 / colbymchenry#2824); the fork has no wasm loader, so a tsx <script> can never lack its grammar. The block-mapping half was already fixed in sfcScriptLanguage.
… only
Two corrections to the finish() is_visible_across_files port:
- Run it on the node after supertype/overload/alias retargets, the way
upstream vets the target matchReference reports — a pre-forward object
property or import alias is not the node upstream inspects.
- Vet only name-matcher picks: upstream's guard sits inside the
matchReference arm, so import, file-path and framework results never
reach it. An import pick's alias forward can legitimately land on an
unexported module member (export const api = { run: right }).
Also ports the Rust arm upstream added for typed receivers: a method a
trait declares carries no pub of its own — reached through a typed
receiver it is the trait's (rust_trait_of + rust_typed_refs, marking
rust_written_call exits). Fixes the rust-* regressions the guard exposed
(rust-typed-receivers, rust-receiver-chains, rust-generic-bound-paths,
rust-method-value-types, rust-self-field-types, rust-wrapped-self-calls)
while keeping dart-part-privacy 12/12.
Remaining exposed failures are import-binding port debt, not guard bugs:
items$.getState().inc (import-type-modifier) and styled(Disclosure)
(function-ref-project-imports) resolve upstream through the import path
the kernel does not fully carry.
…alue call arms Port goReceiverChainType/goElementOf/goPackageValueType semantics tree-natively into the kernel: package-level var/const typing, indexed and chained receiver calls, and declared-type terminal vetoes for literals/outside types. Bound-receiver misses are terminal unless the receiver is an asserted local the index cannot type (upstream strategy-3 passthrough). TS gateTargetKind gains the missing upstream arms (import self-edge, Go extends qualification, dot-imported supertypes) and goBareTypeDir for type-position targets.
Upstream's matchGoPackageValueCall resolves at 0.9; the arm inherited go_method_in_package's 0.85.
goElementOf (named slice/map types resolve through their declarations), goFieldTypeOf (struct field written types), selector-expression and call-expression result arms in go_expr_written. Param/decl locals the declaration reader misses fall back to infer_local_receiver_type.
…ersion refs (colbymchenry#2581) Port upstream 800e582's three Go semantics into the kernel: - A `pkg.Type(…)` conversion resolves its type through the import member arm — the kind filter now accepts Go exported member kinds (struct/interface/type_alias/variable/constant), which the TS edge-kind rewrite then turns into a `references` edge. - `go_name_visible` is clause-aware: files sharing a directory but declaring different package clauses (`a` vs `a_test`) see each other's bare top-level names only through an import — qualified, or bare after a dot import and exported. - New `go_build.rs` ports `goFileConstraint`/`preferGoBuildTwin`: `//go:build` and `// +build` expressions plus `_GOOS`/`_GOARCH` filename suffixes evaluate against the ordered GO_BUILD_TARGETS; a resolved target with constrained twins retargets to the twin the reference's own file builds with, and `go_pick_twin` collapses multi-candidate package members the import arm couldn't pick. Also fixes the fork's `go-type-visibility.ts`: module-root packages spell their dir `''` (matching `getGoPackageDir` and kernel `pos_dirname`), so `goTypePositionTarget` retargets a namesake function to the module-root type.
…colbymchenry#2762) Port upstream dbb6aa9: VbReceiver gains a namespace arm (vb_known_namespaces/vb_namespace_around/vb_namespace_named/ vb_types_in_namespace/vb_module_member_in) so reads like Billing.Invoice.MaxLines and Acme.Billing.Invoice.MaxLines link the type or Module member the namespace declares; vb_takes_no_arguments + vb_runnable_overload pick the overload VB.NET runs when a method is named without parentheses (Me.Load, Optional/ParamArray, a ctor's Me.New never resolves to itself); vb_head_member + vb_with_head link a bare member before the dot (Button1.Text) when no local or parameter shadows it, gated by foreign-Module and outside-namesake checks. vb_receiver_type's member search is extracted as vb_member_around.
colbymchenry#2502, colbymchenry#2456) Port e8ddc4e's declared-entry import resolution into the kernel and its sync-side invalidation into codegraph.ts. Kernel (napi config grows entriesByDir + scopedByName): - resolve_workspace_import takes from_file; workspace_import_targets tries declared entries first — exports (exact, wildcard, fallback arrays, nested conditions), then source/module/main/types/typings in order — then an indexed direct probe, then emitted→source mapping, then the direct probe as last resort. Path deps (file:/link:/portal:) resolve scoped by the importer's nearest manifest, before global names; longest package name wins. - imports.rs gains longest_package_name/match_workspace_package/ declared_workspace_entries/indexed_target/built_source_of plus WorkspaceTargets/exported_targets/built_sources_of; prefilter passes the importing file through. TS sync side (upstream index.ts port): - resolver: refreshWorkspacePackages/getWorkspaceFingerprint/ workspacePackageDirOf; a moved map closes the kernel since the config is construction-time. - codegraph.ts: js_workspace_fingerprint metadata stamped at index and re-checked on full reconcile; a moved fingerprint re-opens every ESM-language ref for the orphan sweep. reopenReferencesThroughAddedFiles gets packageDirOf so a late-appearing declared entry re-links importers. - loader.ts KernelWorkspaceIn + serialization gain the two new maps. workspace-package-entries: 26/26. Regression sweep: 146/147, the one failure a pre-existing tracked bucket (7180fd0 colbymchenry#2788).
…s type Port 449eefa's resolver half: `this.items[0].save()` arrives as `this.items[].save` from the extractors (already ported) and now binds the element type the field's declaration writes instead of bare-name matching any project method. - method_call_shape: elementOf arm `^this\.(#?[\w$]+)\[\]\.([\w$]+)$` → match_ts_this_field_call(field, method, r, element=true) — exclusive, like the plain `this.field.m` arm. - TsFieldDecl gains element; ts_field_element_type reads it off the written type (`Item[]`, `Array<Item>`, `ReadonlyArray<Item>`, `Record<string, Item>`) and rejects unions, arrays of arrays and non-plain elements; ts_field_element adds tsFieldRead's multi-line continuation guard (`| Other[]` on the lines below). - ts_field_held filters element reads through ts_class_type_parameter — `class Store<Item> { items: Item[] }` holds no project `Item`. - ts_class_type_parameter ported upstream: `class|interface|type` keyword anchor, 24-line multi-line head, returns true when the list's `>` is not found (drop the type rather than take a parameter for a class). ts-this-field-element 11/11; expression-receiver-calls + ts-array-annotated-receivers green; receiver sweep shows only tracked debt (08478ee colbymchenry#2811, 013b91c colbymchenry#2825); golden dumps unchanged.
…amed method (colbymchenry#2821) A python call `reg.get()` on a builtin-named method was dropped by the builtin filter before any strategy could read the receiver. Port matchPythonTypedLocalCall (colbymchenry#2583) into python_attribute_verdict: a plain local or parameter receiver settles only through scope evidence — `python_local_class` agrees a binding def's parameter annotation (quoted, Optional[T], Union[..., None], T | None, read off the raw def header) or its assignments (ctor/factory/None; any foreign binding shape, rebinding or contested evidence refuses). Members come from `python_strict_members` — the class's own or a declared base's, resolved through the file's imports, so a base from outside the project holds none of its methods and the extends edges never substitute a same-name class. Sync side: FAILED_PY_ATTR_REF widened to the one-dot `x.y` shape. The fork's delta-narrowed retry keys (perf/sync-delta-retry) never list an unchanged name like `get`, so a signature-only edit to the factory a failed typed-local call reads its type from gave no retry key; the attr-retry lane kernel-probes candidates and retries only resolvable ones, which also lands the edge's typeFrom stamp for future edits.
…turn type (colbymchenry#2822) `repository().pending()` — and `val repo = repository()` locals — type the receiver by the inner call's declared return, so `pending` is the method the returned project type declares, never a same-named method elsewhere. A type from outside the project leaves the name strategies untouched, and an extension the bound-member walk can't see keeps the old name match. The extractor's return_type keeps only a user_type's first identifier (`Request`), so `kotlin_scoped_hit` retypes the callee through its signature (`(): Request.Builder`) and writes a lone nested name through its declaring outer (`Builder` → `Shop.Builder`). Two pre-existing filters were misaligned with the kotlinCall/kotlinBare split this relies on: member-scope reachability vetted receiver-written names (`text().reverse`) as if receiver-less, and a Chain-shape call with a non-project receiver type (`StringBuilder`) vetoed every name match — upstream only constrains standard-library names there. Both now match upstream's isReceiverLessKotlinCall / kotlinStdChain boundary. Also: python_local_class returns Option<Answer> — a name no enclosing def binds (module-level, global) is upstream's `undefined`, not a refused local, which restores module-global `store = Store()` receivers the strict arm claimed and dropped.
…nt (colbymchenry#2790) Port upstream e69546e's mockless context: a ref from non-test code resolves on a graph without the project's __mocks__/ files, while test code keeps the whole graph. resolve_ref sets a per-ref mockless flag; nodes_by_name, nodes_by_lower_name and nodes_by_qualified_name then drop __mocks__/ nodes (gated on the cached mock-name sets so unaffected names aren't re-collected). The TS callback-synthesis guards were already in place. Also fix ts_scoped_reference's type-ref tail: upstream importedTypeMember falls through (undefined) on a found-miss unless the specifier is outside the repository — a resolved package directory whose exports can't be read is not a refusal. The old gate refused references/extends refs bound to a no-entry workspace package, which upstream answers by name-match.
…nputs (colbymchenry#2563) collectSynthesisInputs only kept files whose own content matched mayFeedSynthesis, so a barrel or index that merely forwarded another module's component was never recorded as an input. When a later sync edited only that index, the JSX pass never reran and the tag kept its stale endpoint. Port upstream's importedTagModules: for each JSX-bearing file, map its capitalized tags to project imports and collect every file exportedComponent reads through, memoized per collection run.
…ymchenry#2831) Port upstream 8fdc8a1: a field chain or struct literal receiver's declared type decides the call — `vis.map.get` is HashMap's, `Xyz{}.to_val` is Xyz's — and a path call naming a type std or a dependency owns reaches only the project's impls of that type or the named trait, never a project namesake. - rust_expr_type types tuple-struct/enum-variant constructor calls - undeclared-type guard limited to plain named `field_expression` heads, exempting generics, Self, outside types, use-aliases and wrapper-own methods; derived-trait provided methods resolve through `#[derive]` - rust_path_call_reachable gates name-match on the path's outside type Rust sweep 28 files/206 tests green; kernel golden dumps clean; the 16 remaining full-suite failures are all tracked port debt.
…onJS (colbymchenry#2825) A named class expression is a class bound inside its expression: index its members and supertypes, but resolve references written outside the expression on a graph where that class and its members are not — a same-named declaration or binding keeps its links, and `new Widget()` elsewhere links nothing. References written inside see the whole graph unless the name is another expression's, bound inside that one alone. `module.exports = class Router` / `exports.Store = class StoreImpl` are the module's value: the class stays reachable to every importer, and collect_cjs_export records the declared name as the export's local binding. Inheritance refs resolved through an import binding emit upstream's import confidence 0.9; the scoped arm's 0.95 stays for type-position references.
colbymchenry#2801) Nothing calls an interface or a type alias, or makes an instance of one. A calls/instantiates result that found a TypeScript type by name or through an import meant the value its file declares beside it — the function under `interface Ok` in a result module, the class merged with an interface — and a global the project only types, like `new Response()` or `new Worker()`, links nothing.
tree-sitter-cpp parses `d.operator+=(c)` as a call on the receiver alone, with the operator's name stranded in an ERROR sibling (`function: (identifier) (ERROR ".operator+=") arguments: (c)`), so the callee never spells `operator+=` and cpp_call_argument_types found no call. Every overload then tied at 0 and the first — `operator+=(int)` — took a call that `operator+=(const Counter&)` fits. Accept a call whose error spells the operator name (the member-access dot is part of the error text), so the argument types are read and the class overload wins. `size_type` and `difference_type` count as numbers now, the container convention upstream uses (6bd217a): nlohmann/json's `p->operator[](0)` reaches the index overload, not the `T*` one.
The colbymchenry#2821 port (0b564c78) settled a builtin-named method call through pythonLocalClass alone, which refuses when a name is bound more than once. `store = A(); store = B(); store.get()` then resolved to nothing, though the fork's receiver-scope rule has always read the binding in effect at the call (B). Read that binding: among plain assignments at or before the call, the last one decides, and a bare annotation (`store: B`) does not rebind. python_receiver_uncertain still holds back a same-line replacement and a later replacement a closure's captured cell sees, and a later binding no longer poisons an earlier one when the call reads before it. A def's parameter keeps its annotation when nothing rebinds it. Three shapes the strict read owns stay with it: a binding that is not a plain assignment (a branch, loop, `with`, comprehension), a base from outside the project, and a future binding read from a lambda body. A receiver naming a class declared here (`File.write(f, "x")`) and a lambda parameter of the receiver's name keep the ordinary paths.
2ca95adb applied isVisibleAcrossFiles to every strategy's pick. Upstream guards just the name match (index.ts:1867), and the port's ESM leg vetoed picks that were never name guesses: a typed store action (`items$.getState().inc()`) and an imported component function value were both dropped as unexported module bindings. Keep every language's privacy rule for a structured pick, and skip the ESM export check for a pick a receiver's type or a function value made.
Every case held on each corpus (json 11 absent + 7 present, flask 2 present, vite 6 absent + 2 present, svelte 1 absent).
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
 |
Owner
Author
|
Review round (GPT-6.1 Sol, xhigh) returned fix-then-ship with four findings; all four are fixed in
Re-ran after the fixes: the seven focused suites 180 tests with only the pre-existing Pascal failure, and the full suite 11,068 passed with the same 23 pre-existing failures as origin. |
- C++ (P3): a comment inside the access (`b./* note */operator+=(a)`) left the comment in the error's text, so the operator name never matched. Compare the `operator_name` node the grammar still reads, and keep the trimmed-text form only when there is none. - Python (P2): a local read before its only assignment in the same def (`result = store.get(); store = A()`) is an UnboundLocalError, not the later value; refuse it, keeping the future-initialization exception for a nested def that captures the outer name. - Python (P2): two conditional bindings in different blocks leave no value in effect (`if flag: store = A() else: store = B()`, then `store.get()`); refuse. One conditional beside a `None` placeholder or a later plain assignment still reads, and a conflicting set whose binding in effect may not run stays unknown. - ESM (P2): a function value is import-justified only when the file's own import resolves to the target's file, so a bare name fallback can no longer take an unexported namesake from another module. A default-exported wrapper still reaches its inner function through the default import, and `export default function Name` counts as an exported name. Regression tests: commented operator access, a future local read, two class-binding branches, and an imported name the module does not export.
bompus
force-pushed
the
fix/kernel-port-debt
branch
from
October 29, 2026 16:47
c17ae18 to
3289308
Compare
Owner
Author
|
Landing with a documented waiver, on the owner's instruction. macOS legs (waived). Windows flake (resolved). One Review. GPT-6.1 Sol (xhigh) returned fix-then-ship; all four findings are fixed in Scope of what the waived leg covered. |
Merged
5 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Finishes the pinned upstream resolution port in the Rust kernel: 29 commits, 26 of them ports of upstream behaviour, plus three fixes this branch found by running the fork's own tests against them.
What the port covers
this-field element calls, framework ownership past the name strategies, Dart library privacy, indexed import-path spelling, lone same-file match confidence,finish()visibility.The three fixes on top
48b54075). tree-sitter-cpp parsesd.operator+=(c)as a call on the receiver with the operator name stranded in anERRORsibling, socpp_call_argument_typesfound no call, every overload tied at 0, andoperator+=(int)took a call thatoperator+=(const Counter&)fits. The error text is accepted as the callee spelling.size_type/difference_typecount as numbers, as upstream does, sop->operator[](0)reaches the index overload.operator-method-names.test.ts3/5 → 5/5.f9fb8f12). The fix(python): calls like reg.get() on a typed local or parameter reach the class's method colbymchenry/codegraph#2821 port settled builtin-named method calls throughpythonLocalClassalone, which refuses any name bound twice;store = A(); store = B(); store.get()then resolved to nothing. The binding in effect at the call now decides, with the fork's existing uncertainty rules (same-line replacement, captured cell, deferred lambda) still refusing, and the fix(python): calls like reg.get() on a typed local or parameter reach the class's method colbymchenry/codegraph#2821 prohibitions (branch/loop/with/comprehension bindings, outside-project bases) still holding.receiver-scope-lookup61/69 → 69/69,method-call-owner-class12/13 → 13/13,python-typed-builtin-method-calls24/24.0ae0b138).2ca95adbappliedisVisibleAcrossFilesto every strategy, but upstream guards only the name match (index.ts:1867). Its ESM leg dropped a typed store action and an imported component function value as unexported module bindings. Structured picks keep every language's privacy rule and skip the ESM export check.import-type-modifier9/10 → 10/10,function-ref-project-imports2/3 → 3/3, Dart part privacy still 12/12.Checks
vitest run: 11,064 passed, 23 failed, 43 skipped (928 files). All 23 failures are pre-existing onorigin/fork/consolidatedat421ea886; I ran the same files on a clean worktree of that commit and they fail there identically: explore substring ordering 3, Pascal function-ref 1,index-recovery-off-thread8 (the test expects anidx_edges_synthesis_siteindex no code creates), Java records 2, JS receiver binding 1,ui-package8 (only_child is not a functionin the Svelte runtime)..dumpchange, so no re-baseline.eval:precisionon the corpora this touches: json 11 absent + 7 present held, flask 2/2, vite 6+2, svelte 1/1. Results committed in3c708552.npm testcannot run on this host: itspretestrunscheck:kernel-rules, which reports 10 diagnostics incodegraph-kernel/src/resolve/pascal_scope.rs. That file is untouched by this branch andorigin/fork/consolidatedhas the same five sites, so the default branch fails that gate too. CI does not run the script. The suite above isvitest rundirectly.README rows
Checked the "About this fork" tables and the benchmark rows: none move. This is resolution correctness with no user-visible surface, command, language, framework or config change, and no measured number in the README comes from these paths.