fix(parsers): recover Python annotated parameters (40% loss) and Go grouped imports - #1561

Fixes #1520 and #1521 — two parser defects that silently dropped large fractions of the graph. Both were found by an audit and are reproduced below.

Full suite: 1065 passed, 7 skipped, no failures.

Python — annotated parameters without a default (#1520)

typed_parameter has no name field in tree-sitter-python; it exposes only type. Its children are (<identifier|splat>, ':', type):

typed_parameter -> children: [('identifier', 'x'), (':', ':'), ('type', 'int')]

So child_by_field_name('name') returned None and the parameter was discarded. x: int = 5 survived only because it is a different node type — typed_default_parameter, which does have the field — which is exactly why this went unnoticed for so long.

The fix reads the first child instead, and covers the splat forms (*args: int, **kw: str) that also appear in that position.

Measured against the stdlib ast module over src/codegraphcontext/:

Since this repo is heavily annotated, self-indexing was losing over 40% of Parameter nodes and their HAS_PARAMETER edges.

Go — grouped import (...) blocks (#1521)

The imports query required import_spec to be a direct child of import_declaration. That shape only occurs for a single-line import "fmt"; a grouped block nests them one level deeper:

import_declaration
└── import_spec_list          <- not matched
    ├── import_spec  "context"
    └── import_spec  "fmt"

gofmt rewrites any file with more than one import into a grouped block, so essentially every real Go file lost every import. With no Module nodes and no IMPORTS edges, cross-file call resolution for Go had nothing to work from.

Added the nested alternatives for both the plain and aliased forms. Verified:

grouped:        [('context', None), ('fmt', None), ('sql', 'sql')]
single:         [('fmt', None)]
aliased single: [('fmt', 'f')]

The repo's own sample_project_go/main.go fixture went from 0 imports to 2.

Goldens

sample_project and sample_project_go regenerated. Both diffs are purely additive — verified by counting node labels before and after:

sample_project:     Parameter  139 -> 143   (+4)
sample_project_go:  Module       2 ->  16   (+14)

No nodes lost in either. I deliberately did not touch the goldens for other languages.

Note for reviewers

Fixing parsers changes what the graph contains, so node and edge counts will move. That is the point of the change, but it will shift the numbers in #1393 and is worth keeping in mind for the Database Parity Check. The remaining parser bugs (#1522 PHP, #1523 Ruby, #1524 C/C++, #1526 JS/TS) will each do the same, so they are probably best merged deliberately rather than all at once.

🤖 Generated with Claude Code