Meta: Go / cgo / Go+C extraction — tracking issue #1932

Tracking issue for the Go / cgo / Go+C extraction work. Method: count a construct in source, count the same thing in the graph, treat any mismatch as a defect. Every item was minimised to a dummy fixture before filing, measured on one real Go + C/C++ repository (~1150 Go files, eBPF C probes, a Windows C++ driver — counts only), and carries a SQL detection recipe that works on any indexed repo. Each fix PR ships a reproduce-first test (RED proven before the change) and a fresh field census on the same tree.

Ledger — bugs

Combined stack (#1937 + #1940 + #1944) vs main: USAGE 29198 → 21035, WRITES 2681 → 774, +4533 field nodes — the graph gets nodes and loses junk edges.

Ledger — design-shaped (awaiting maintainer ack, no PR)

Ledger — CI

Merge order

• Stack A: fix(pipeline): suppress weak short-name matches for Go selector calls #1907 → fix(extract): receiver-qualify Go method QNs #1913 → fix(extract): give each Go init() its own QN #1915 → fix(pipeline): bind field_type_hint by owning QN segment, not substring #1936 (each rides the previous; fix(pipeline): bind field_type_hint by owning QN segment, not substring #1936's probe relies on fix(pipeline): suppress weak short-name matches for Go selector calls #1907's guard + fix(extract): receiver-qualify Go method QNs #1913's receiver QNs). Direction accepted 2026-09-03; fix(pipeline): suppress weak short-name matches for Go selector calls #1907/fix(extract): receiver-qualify Go method QNs #1913 rebased onto post-memory-core main, fix(extract): give each Go init() its own QN #1915/fix(pipeline): bind field_type_hint by owning QN segment, not substring #1936 held per maintainer until those two land.
• Stack B: fix(pipeline): guard USAGE/WRITES/READS against cross-language binds #1937 → fix(extract): descend into Go struct field_declaration_list #1940 → fix(pipeline): a bare Go reference never binds a struct Field #1944 (fields must not land before the reference guards; fix(extract): descend into Go struct field_declaration_list #1940 + fix(pipeline): a bare Go reference never binds a struct Field #1944 ideally merge together).
• Independent: fix(extract): drop cgo's import "C" pseudo-package #1931, fix(pipeline): never bind a Go import to a symbol #1938.
• Related external: fix(go): resolve cross-package field-chain calls via type dispatch #1939 (@xkchok) — cross-package field-chain dispatch via lsp_type_dispatch; complementary (consumes fix(extract): descend into Go struct field_declaration_list #1940's field nodes, replaces the weak path fix(pipeline): bind field_type_hint by owning QN segment, not substring #1936 tightens). Overlaps extract_defs.c with fix(extract): descend into Go struct field_declaration_list #1940 — whichever merges second rebases.

Cross-cutting theme

One failure mode at every layer: a heuristic binds by name shape and nothing checks the language or what the name can denote. #1906 (CALLS, guarded) → #1927 (field_type_hint, exempted on a false premise) → #1928 (references, never guarded) → #1934 (imports, label-blind) → #1942 (references, label-blind). Guarding one relocates mass to the next unguarded one — which is why every fix PR above now carries negative/exact-count probes (ei_edge_count_is, zero-project-edge assertions, parallel-twin fixtures) rather than floor-only asserts, closing #1932's original test-infra gap in place.

Verified working (audit surface closed, no issues needed)

Generics (receiver-qualified QNs unwrap generic_type; no [ in any QN) · embedding + interface dispatch (lsp_embed_dispatch 42, lsp_interface_dispatch 79, fan-out via 434 IMPLEMENTS) · //go:linkname (body-less decls get nodes; callers bind lsp_direct) · defer/closures (attributed to the enclosing named function) · implicit interface satisfaction + method values (lsp_callable_alias) · TESTS edges and Variable coverage · internal import resolution (2206/~2242 same-module imports → correct package Folder; residual was #1934). Assembly (.s): untested — no fixtures in the measured repo.

Detection recipes (run on any indexed repo)

-- call-graph shape
SELECT json_extract(properties,'$.strategy') AS strategy, count(*) AS n,
       round(avg(json_extract(properties,'$.confidence')), 3) AS conf
FROM edges WHERE type='CALLS' GROUP BY 1 ORDER BY n DESC;

-- node recovery vs source: grep -rc '^func init()' --include='*.go' .
SELECT count(*) FROM nodes WHERE name='init';

-- swallowed twins: a file whose only node is its own __file__
SELECT file_path, count(*) c FROM nodes GROUP BY 1 HAVING c = 1;

-- cross-language reference edges (any Go↔C/C++ row is suspect)
SELECT e.type, count(*) FROM edges e
JOIN nodes s ON s.id=e.source_id JOIN nodes t ON t.id=e.target_id
WHERE s.file_path LIKE '%.go'
  AND (t.file_path LIKE '%.c' OR t.file_path LIKE '%.h'
    OR t.file_path LIKE '%.cpp' OR t.file_path LIKE '%.hpp')
GROUP BY 1;

-- references onto Go fields (must be zero: field access is selector-only)
SELECT e.type, count(*) FROM edges e JOIN nodes t ON t.id=e.target_id
WHERE t.label='Field' AND t.file_path LIKE '%.go'
  AND e.type IN ('READS','WRITES','USAGE') GROUP BY 1;

-- Go import targets (anything not Folder/Module/File is false)
SELECT t.label, count(*) FROM edges e
JOIN nodes s ON s.id=e.source_id JOIN nodes t ON t.id=e.target_id
WHERE e.type='IMPORTS' AND s.file_path LIKE '%.go' GROUP BY 1;

Every census lives inline in its issue's body (and in the fix PR's description); issues in this family are self-contained.

Related work in other ecosystems: #1893 (Swift, method-QN class), #1276 (Python, closed), #495 (Rust cfg twins — the pattern #1911 wants), #725 (cross-language CALLS guard), #1114 (dispatch fan-out), #787 (file attribution).