parser-matrix.md

Parser Construct Support Matrix

Status: Step 07 snapshot of the hardening blueprint

Scope: What each KnowCode parser extracts today, and the explicit limitations a user or downstream step can rely on. This matrix is enforced by the cross-language integrity gates in tests/unit/indexing/ and tests/unit/parsers/, and the exact fixture contracts in tests/fixtures/parser_contracts/. It is the single source of truth referenced by the release gate (blueprint Step 22).

The parser and graph invariants themselves live in ADR 1. Anything not listed as supported here is either an explicit limitation below or outside the committed fixtures.

How to read this matrix

• Supported means a committed fixture or gate proves exact extraction, location, and graph-identity behavior through GraphBuilder, not only direct parser output.
• Limitation means the construct is either not extracted or extracted incompletely, and the gap is visible (reported through ParseResult.errors or documented here) rather than silently lost. Silent loss is a defect, not a limitation.

Endpoint and identity contract (all languages)

Every relationship endpoint is exactly one of:

Classification	Form
Internal	<canonical-file>::<qualified-name>
External	external::<namespace>::<symbol>
Unresolved	unresolved::<language>::<file>::<scope>::<symbol>

Legacy pseudo-namespaces (type::, trait::, ::method::, ::data::) are invalid and rejected by the gates. A parser may emit a transient ref::<name> placeholder for a reference it cannot fully qualify at parse time; GraphBuilder._resolve_references links it to a local entity when one exists. Committed fixtures and the mixed-language merge contain no lingering ref:: or invalid endpoints.

Scoping note: this endpoint contract is enforced for the hardened parsers above (JS/TS, Python, Vue, Rust) and the mixed-language gate fixture. The Java parser predates it and still emits non-canonical forms — see its limitations below; the mixed-language fixture contains no .java file, so the gates do not cover it.

Duplicate declarations in one file never produce duplicate entity IDs: the parser keeps the first and reports the dropped collision (Step 07). The synthetic per-file module entity (named after the file stem) is intentionally excluded from that dedupe so a top-level declaration whose name matches the file stem — notably a Java public class in ClassName.java — remains a redundant wrapper that the graph merge resolves in favor of the declaration.

JavaScript and TypeScript (tree-sitter)

Supported:

• Classes, functions, arrow functions, and function-valued variable declarations.
• TypeScript interfaces, type aliases, and enums (extracted as class entities until the shared entity-kind schema expands).
• Named and default export declarations, unwrapped through one shared dispatch path.
• extends for simple identifiers, member expressions, and the explicitly supported complex grammar forms; nonlocal bases become unresolved references.
• Module entity, containment, and call edges.

Limitations:

• .tsx JSX bodies are reported as tree-sitter syntax errors, because TypeScriptParser uses the plain typescript grammar (no TSX); tree-sitter error recovery still yields partial extraction, but the entity set is not guaranteed beyond the module entity. .jsx files route to JavaScriptParser, whose javascript grammar accepts JSX and extracts normally — JSX tags are not extracted as entities in either case. JSX-free TypeScript in a .tsx file parses and extracts normally.
• TypeScript interfaces, aliases, and enums share the class entity kind.
• Duplicate top-level declarations are deduped and reported (Step 07); the synthetic module entity is exempt so a Java-style filename match does not false-trigger.

Python (ast)

Supported:

• Nested classes, nested functions, async nested functions, with lexical qualified names (Outer.Inner, Outer.method.local).
• Decorators on classes, functions, and methods; entity location and source begin at the first decorator.
• Module Assign and AnnAssign produce one variable entity per simple bound name, including each name in chained or tuple assignment.
• Scoped call resolution: a call resolves to the lexical scope that owns it and never leaks across nested definition boundaries.
• Imports (external) and inheritance.

Limitations:

• A syntax error fails the whole file (ast.parse raises); there is no partial extraction. The failure is visible in ParseResult.errors and deterministic.
• Module-local (non-module-scope) assignments are intentionally not entities.

Vue SFC (custom parser)

Supported:

• Attribute-order-insensitive SFC section scanning with exact byte/line offsets; malformed or unclosed sections are reported.
• Composition API (<script setup>) and Options API; script content parsed with the JS or TS parser and rebased into the .vue file.
• Template bindings (v-model, @event handlers), emitted events, CSS v-bind(), imports, and composables, resolved through a per-component symbol table. Template v-bind: / : directives are not extracted — only the CSS v-bind() form in <style> is.
• Exact declaration lines and source snippets.

Limitations:

• v-model.trim and other .modifier forms drop the binding edge.
• Quoted data() keys ("count": 0) and array-form defineProps(['title']) produce no entity.
• Generic defineEmits<{ (e: 'save'); (e: 'cancel') }>() captures only the first event.
• An import statement inside a block comment still registers.
• _get_component_name lowercases interior capitals, so MyButton.vue yields the entity Mybutton while importers yield external::vue_component::MyButton — this blocks future cross-file resolution until normalized.
• A component pairing a plain <script> with <script setup> indexes only the setup block; the plain block's Options API declarations are missing and reported.
• Duplicate Options API method keys or data() keys silently keep the first declaration (not reported); template-binding name collisions are reported.
• Relationship identity is (source, target, kind, binding_type); the shared fixture helper keys on (source, target, kind), so two legitimate edges that differ only by binding_type are both retained but cannot yet be expressed in one fixture.

Rust (tree-sitter)

Supported:

• Structs, enums, traits, inherent and trait impl blocks, methods, generics, and qualified trait paths.
• Lexical module-scope reference resolution: a bare name declared once resolves to its entity; qualified or foreign paths become scoped unresolved references.
• Calls for local functions and Type::method naming a same-scope method; imports from use trees (including self, grouped, and as forms).

Limitations:

• const, static, and type aliases are extracted with no containment edge, so they are unreachable by graph traversal.
• Trait bodies contribute no entities; default methods and required signatures are missing.
• Associated const / type items inside impl blocks are skipped.
• Import-aware external trait classification is not performed, so a trait reached through an imported path (use std::fmt; then impl fmt::Display) is unresolved rather than external.
• Structs, enums, traits, and type aliases share the class entity kind.
• A field and a method on the same type can collide on the Type.name qualified-name scheme; the collision is dropped and reported.

Java (tree-sitter)

Supported:

• Classes and interfaces (both class entities), methods and constructors (both method entities), with containment edges from the file module.
• Imports as IMPORTS edges; extends/implements superclasses as INHERITS edges via ref:: placeholders resolved against local entities by GraphBuilder.
• Calls: method invocations (foo.bar(...), bare bar(...)) and new Foo() constructor calls, emitted as CALLS edges via ref:: placeholders.

Limitations (pre-hardening parser; not covered by the endpoint gates):

• Endpoint forms predate ADR 1: imports emit external::<dotted.name> (e.g. external::java.util.List) instead of external::<namespace>::<symbol>, and internal IDs are raw file::name strings without the shared normalization helpers, so classify_endpoint_id marks them invalid.
• ref:: placeholders without a matching local entity linger in the graph instead of becoming unresolved:: endpoints.
• Nested method invocations can emit duplicate CALLS edges (the cursor walk re-visits a parent invocation after its children).
• Class fields, enum declarations, and annotation declarations are not extracted; package names are not part of qualified names.

Markdown (custom parser)

Supported:

• One DOCUMENT entity per file (first paragraph as docstring) and one SECTION entity per ATX heading (#–######), with CONTAINS relationships following heading nesting, heading levels in metadata, and exact section line ranges.
• Headings inside fenced code blocks are deliberately skipped.

Limitations:

• Setext headings (underlined with === / ---) are not detected — only ATX # headings are.
• Section IDs are slugified titles; duplicate heading titles collide on one ID (the first wins after graph dedupe).

reStructuredText (custom parser)

Supported:

• Mirrors the Markdown parser: one DOCUMENT entity plus one SECTION entity per heading with a CONTAINS hierarchy. Heading detection uses the shared RST adornment scanner (prose_chunker._scan_rst_headings), so section levels stay consistent between the graph and the retrieval index.

Limitations:

• Inherits the Markdown shape: no cross-references, directives, or role markup; only section structure is extracted.

YAML (custom parser)

Supported:

• One DOCUMENT entity per file plus one CONFIG_KEY entity per mapping key at any nesting depth (dotted qualified names), with CONTAINS relationships following key nesting, value type and a bounded value preview in metadata. yaml.safe_load with parse errors visible and deterministic.

Limitations:

• Keys only: scalar/list values contribute a preview, never entities; a non-mapping root (list or scalar) yields only the document entity.
• Key line numbers use a textual search heuristic and can misattribute on duplicate keys, keys inside multiline strings, or commented-out keys that look real.

Gate coverage

Invariant	Gate
Exact entities/relationships/locations through GraphBuilder	test_graph_builder_references.py::test_graph_builder_matches_fixture_contract (parametrized over every fixture)
Unique entity IDs; collisions reported	test_graph_integrity_gates.py::test_duplicate_declarations_never_produce_duplicate_entity_ids
Mixed-language merge has no invalid/dangling endpoints	test_graph_integrity_gates.py::test_mixed_language_merge_has_no_invalid_or_dangling_endpoints
Output independent of scan order	test_graph_integrity_gates.py::test_mixed_language_graph_is_independent_of_scan_order
Malformed input is visible and deterministic	tests/unit/parsers/test_parser_negative_fixtures.py
Extension dispatch (.tsx, .jsx, Vue TS)	tests/unit/parsers/test_parser_extension_dispatch.py