feat: Monkey C (Garmin Connect IQ) extractor + member-call resolver

[stepan-orlov]

Summary

Adds Monkey C (Garmin Connect IQ, .mc) as a supported code language.

Today .mc is in no extension set at all, so a Connect IQ app is silently dropped: classify_file() returns None, the files land in unclassified (which the skill never reports), and graphify extract . --code-only on such a repo ends in Graph is empty. There is no tree-sitter grammar for Monkey C, so this follows the Pascal precedent (_extract_pascal_regex) with a scanner-based extractor in its own module, graphify/extractors/monkeyc.py, wired per ARCHITECTURE.md → "Adding a new language extractor".

What is extracted

Comments, string and char literals are blanked with offsets preserved (line numbers stay exact — '{' and "a { b // c" don't desync anything), then a single brace-depth scan attributes every declaration to its enclosing scope and every call to its function.

Construct	Emitted
module M { } / class C { } (nested too)	node + contains from file/owner
function f(...) in a class / module / file	.f() + method edge / f() + contains edge; _callable marked
class C extends B	inherits; same-file base → real node; other base → sourceless stub labelled with the alias-expanded name (extends Ui.View under using Toybox.WatchUi as Ui → Toybox.WatchUi.View), so _rewire_unique_stub_nodes collapses it onto the real class when the base is in the corpus and SDK bases stay one shared external node
using X.Y as Z; / import X.Y;	imports_from (context import); a bare import Utils; (a module of the same app) targets the local module if declared in the file, else a sourceless stub for the same rewire; Toybox.* stays dangling and is dropped at build like a Python stdlib import
f(), me.f(), self.f(), $.f(), LocalType.f(), new LocalClass()	same-file calls (own class → enclosing (innermost) module → file scope), EXTRACTED, context call; new Toybox.*() is never a raw call
method(:sym), me.method(:sym), new Lang.Method(self, :sym)	indirect_call (INFERRED, context callback) — the Connect IQ timer/callback idiom
everything else	raw_calls tagged lang="monkeyc": bare calls for the shared cross-file pass; member calls with receiver, receiver_type and receiver_kind (static for an explicit Type.fn() qualifier, typed when the receiver is typed via var x as T / a typed parameter / x = new T()); unresolved bare calls made from a class body carry self_scope

New language resolver: monkeyc_member_calls

Monkey C code is dominated by member calls the shared pass skips by design (#543/#1219): an app is one namespace of modules and classes, so Utils.getSecret(), WristlaBleSession.enroll() and _transport.write() are the architecture. Registered right after java_member_calls, additive, single-definition guarded:

• Module.fn() / Type.fn() → calls EXTRACTED; type known but function not (a field / SDK method on a subclass) → type-level references;
• receiver typed by local inference → calls INFERRED;
• an unresolved bare call from a class body → looked up along the caller's inherits chain (initialize() / onShow() provided by a base class in another file) → EXTRACTED; if the shared pass already bound the same pair by corpus-unique name nothing is added (strictly additive, no mutation of other passes' edges);
• Toybox.* receivers are the SDK and never bind to app code; _LANGUAGE_BUILTIN_GLOBALS respected; a module's functions are found through contains, a class's through method.

Wiring

_DISPATCH (collect_files() derives from it), CODE_EXTENSIONS, _LANG_FAMILY_BY_EXT (+ mirrored build._EDGE_LANG_FAMILY) as family monkeyc, cli._HOOK_SOURCE_EXTS, LANGUAGE_EXTRACTORS, facade re-export in extract.py, README extension table, CHANGELOG (0.9.45 unreleased).

Tests

• tests/fixtures/sample.mc + tests/fixtures/monkeyc_cross_file/{Base,App}.mc + tests/fixtures/monkeyc_ambiguous/{A,B,Caller}.mc
• 15 tests in tests/test_languages.py: modules/classes/functions, contains/method, local + stub inherits, imports, same-file calls (incl. $. qualifier), callbacks in all three spellings (method(:sym), me.method(:sym), new Lang.Method(self, :sym)), raw-call typing (static/typed/self_scope, alias expansion, SDK constructors never raw), no dangling edges, line numbers, dispatch/detect registration, a full extract() two-file run (inherits rewired onto the real class with no leftover stub, Module.fn(), Base.initialize(), new Base(), inherited bare call, type-level references, import Store; rewired onto the real module node), and the exactly-one-definition guard (two same-named classes → no edge).
• Full suite: 4464 passed, 48 skipped (the 4 test_ollama_retry_cap tests fail identically on pristine v8 in this env — openai extra not installed). tools.skillgen --check/--audit-coverage/--schema-singleton/--monolith-roundtrip/--always-on-roundtrip all OK; ruff check clean on the new module.

On a real corpus

A production Connect IQ app (18 files, ~19k lines): 879 nodes / 2660 edges in 0.6 s, 428 cross-file call edges (the UI layer → BLE session module → protocol/crypto/transport modules chain comes out as the top pairs), 15 inherits (all onto alias-expanded SDK stubs), a bare import <Module>; present in 8 files rewired onto the one real module node.

Notes / caveats

• .mc is also the extension of sendmail m4 configs and Windows Message Compiler sources; those contain none of the Monkey C keywords the scanner keys on, so such a file degrades to a lone file node rather than misparsing.
• Scanner-based, so it will miss what a real grammar would not: chained calls a.b().c() yield only b, and a function whose parameter list nests parentheses more than one level deep is skipped (braces still balance, so nothing downstream desyncs).

🤖 Generated with Claude Code

[graphify-labs]

Graphify reviewed this change.

Worth a look — the grounded gate found no coupling regressions or blocking issues, but 2 advisory finding(s) below merit a look before merge.

---

Graphify review — findings

Adds a scanner-based Monkey C (.mc) extractor in graphify/extractors/monkeyc.py covering modules, classes, functions/methods, extends, using/import, calls, and method(:sym) callbacks, plus a monkeyc_member_calls corpus-wide resolver for qualified/typed-receiver/inherited calls. Registers .mc across the extractor dispatch, edge-family map, CODE_EXTENSIONS, hook source exts, and the language-family tables, and documents it in the README and CHANGELOG.

Worth a look

• Qualified Monkey C type labels are skipped before exact lookup — graphify/extractors/monkeyc.py · Escalate · medium
  • agreed by 2 of 2 members but NOT verified (no proof, no reproducing execution) — consensus is not a verdict; needs human review Execution auto-disposal is off for this run; enable it (with sandbox isolation) to have Graphify try to confirm or refute this automatically.
• Scoped duplicate type names resolve to the first declaration — graphify/extractors/monkeyc.py:236 · Escalate · medium
  • agreed by 2 of 2 members but NOT verified (no proof, no reproducing execution) — consensus is not a verdict; needs human review Execution auto-disposal is off for this run; enable it (with sandbox isolation) to have Graphify try to confirm or refute this automatically.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 3135 functions depend on the 1136 functions this change touches.

Health — this change adds coupling hotspots:

• new: extract() — 464 callers, 41 callees
• new: _rebuild_code() — 98 callers, 51 callees
• new: build_from_json() — 153 callers, 18 callees
• new: detect() — 99 callers, 15 callees
• new: build_merge() — 43 callers, 14 callees
• new: save_manifest() — 34 callers, 11 callees
• new: to_obsidian() — 29 callers, 12 callees
• new: extract_files_direct() — 17 callers, 20 callees
• …and 74 more — each is listed as a finding

Verification — 3135 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 2955 function(s) in the blast radius were not formally verified this run

· 1 grounded finding(s) anchored inline below; 81 more finding(s) on lines outside this diff (see the check run).

[graphify-labs]

⚠️ Health regression — extract_monkeyc()

fans out to 16 callees (efferent coupling).

Grounded coupling-delta finding (deterministic), not an LLM guess.