feat(languages): add MQL4/MQL5 extractor + mql-lsp-server registration

[davalillo]

Summary

Adds MQL4/MQL5 (.mq4/.mq5/.mqh) as a first-class language: bespoke tree-sitter extraction (following the #805 shape: extractor + tests + registration + docs) plus the LSP registration that upgrades MQL edges and unlocks the on-demand action tools.

• One grammar serves all three extensions: modern MQL4 (build 600+) shares MQL5 syntax; pre-600 legacy sources are a C subset. The grammar, davalillo/tree-sitter-mql5, is independently maintained since 2026 — it began as a fork of mskelton/tree-sitter-mql5 (tree-sitter-cpp + the MQL5 input storage class), but the upstream has been dormant since 2023, so that repo is now the canonical home: full git history preserved, per-contribution attribution in ACKNOWLEDGMENTS.md. It has moved well past the original fork — rebased onto tree-sitter-cpp v0.23.4 (tree-sitter CLI 0.25, ABI 15) with a Go binding and corpus-driven extensions: sinput, interface_specifier, MQL-native C'…'/D'…' literals, input group, MQL primitive types, declarations as un-braced if/else consequences, and empty member declarations after inline constructor bodies. The external scanner is the C scanner of the cpp base (renamed tree_sitter_mql5_*; raw strings are not MQL5), keeping the build free of a C++ toolchain.
• Extraction reuses the shared machinery (combined query, deferred calls, funcRanges attribution) and adds the MQL surface: input/sinput/extern variables, Meta["methods"] interfaces, methods with EdgeMemberOf, #include imports, Meta["dialect"] on file nodes (content-sniffed for .mqh).
• LSP: registers mql-lsp-server (my project, davalillo/mql-language-server: an OmniSharp-based stdio LSP) so definition/implementation enrichment produces lsp_resolved edges and get_diagnostics/get_code_actions/fix_all_in_file work on MQL. .mqh deliberately maps to a languageId the server does not short-circuit on, so its LanguageDetection routes headers by content sniffing — a header's dialect isn't knowable from its extension. Plain stdio, no args (the --stdio flag its docs show is accepted and ignored). Install: self-contained binary from GitHub Releases (no .NET runtime) or dotnet tool install -g mql-language-server (.NET 10).
• Grammar module location (discussed with @zzetorg on Discord): it stays under my account, github.com/davalillo/tree-sitter-mql5 — I'm its maintainer and will keep evolving it alongside mql-language-server. Happy to transfer to gortexhq in the future if that ever becomes the better home.

Corpus audit

The grammar extensions are data-driven, not speculative. Parsing a 53-file MQL corpus with the stock grammar leaves 46 files clean with 37 ERROR nodes; after the extensions, 47 files clean and 10 ERROR nodes — every remaining one is a macro-expansion-only declaration (Controls event map), the same boundary mql-language-server enforces with its macro layer.

On a real 21.4k-line / 720 KB production EA, the stock grammar produced 40 contained ERROR nodes; the two rules added for it (declaration-as-statement, stray semicolon after inline constructor bodies) bring it to zero ERROR nodes. The v0.2.0 grammar also ships a 111-file real-corpus regression harness with CI (fails if the per-file error count rises). 100/100 tree-sitter tests pass, including the full upstream C++ corpus (no C++ regressions).

Extraction performance on that EA: ~145 ms per file (parse + graph), 400 nodes (198 functions, 175 variables, 18 types) and 4,668 edges (4,263 calls).

Changes

File	Change
go.mod, go.sum	pin github.com/davalillo/tree-sitter-mql5 v0.2.0
internal/parser/tsitter/mql/mql.go	grammar shim (dart.go pattern)
internal/parser/languages/mql.go	combined-query extractor
internal/parser/languages/mql_test.go	happy path, empty input, dialect stamping
internal/parser/languages/register.go	register before forest languages
docs/languages.md	count 257, matrix row, MQL specifics + not-covered boundaries
internal/parser/languages/mql.go (+test)	input variables now mint for array and multi-declarator forms (input double Rates[][6];, extern int A, B;) — previously silently dropped
internal/semantic/lsp/registry.go	mql-lsp-server spec (priority 5)
docs/lsp.md	server matrix row + install instructions, pinned to the v2.4.2 tag with SHA256 verification (CHECKSUMS.txt) and the full platform asset matrix
docs/languages.md	grammar described as independently maintained (no longer 'a fork of mskelton', no 'scanner stub')

Testing

• New tests added: go test -race ./internal/parser/languages/ -run TestMQL → 7/7
• go test -race ./internal/semantic/lsp/ → 376/376
• Meta-tests (extension collisions, coverage clusters) pass; gofmt clean; go build ./... passes
• go test -race ./... in full — blocked on my machine by a pre-existing glibc heap corruption inside tree-sitter cgo (SIGABRT, moving crash site), reproduced identically at base 2b5480bf with no MQL code present. Happy to file a separate issue with the reproduction details.
  • Addendum (independent environment, Linux/glibc): the same corruption reproduces without -race on ./internal/parser/languages (corrupted size vs. prev_size, moving crash site mid-Go-extractor tests) and at the PR base with zero MQL changes (~5/6 runs), so it is not -race-specific. All MQL tests, the LSP suite (376/376) and meta-tests pass; only the flaky abort interrupts the full-package run.
  • Same environment shows a second, unrelated pre-existing breakage: internal/parser/forest does not compile (go-sitter-forest perl scanner.c vs system gcc/glibc) — no MQL code involved.

Checklist

• Code follows existing patterns in the codebase (shim over binding, one combined query, shared helpers)
• No unnecessary abstractions added
• Language extractor includes Meta["methods"] for interfaces
• Methods have EdgeMemberOf edges to their containing type

[davalillo]

Context: why MQL belongs in a code-intelligence engine

MQL4/MQL5 is the proprietary language of MetaTrader 4/5, the retail trading platform family by MetaQuotes. Far from niche:

• Per MetaQuotes, "hundreds of brokers and banks worldwide" run MetaTrader 5, serving millions of traders; Finance Magnates Intelligence measured that 83.8% of retail brokers offered MT4/MT5 (Q2 2022).
• MQL5.com — the language's community hub — reports 7M+ monthly visitors, and its public CodeBase holds 12,000+ programs. On GitHub alone there are ~2,700 MQL5 repos and ~29,500 .mq5 files.
• The code is financial logic in production: Expert Advisors trading real accounts, often as decade-old, multi-thousand-line files that urgently need code navigation and refactoring tooling. Gortex's graph (callers, usages, blast radius) is precisely the tool that codebase lacks.

The grammar and extractor in this PR were validated against a real 21.4k-line production EA from that ecosystem (zero parse errors after the corpus-driven grammar extensions).

[davalillo]

Hi @zzet, the macOS test job failed on the latest run, but the failure looks pre-existing and unrelated to this diff. Could you re-run the failed jobs when you have a minute?

Details: test (macos-latest, 1.27) failed on TestCheckoutStaleDiscoveryProofNeverReturnsBaseData in internal/mcp/checkout_control_test.go, an area this PR does not touch (the diff is parser, LSP registry, docs, and go.mod/go.sum pins). The failure mode is a freshness fallback to the base view with fallback_reason: "view_building" while the checkout view was still building, which is timing-sensitive. Two data points that point at a flake rather than a regression:

• The same checkout_control family already failed on main on Sept 23, on Windows (run 35849558723: TestObserveCheckoutPathRejectsInventoryWithReboundGitMarker, TestCheckoutMutationConcurrentAdmissionIsBoundedByCaller), with no MQL code involved.
• The earlier CI run on this branch (38bc3ea) was fully green, including macOS.

If it reproduces after the re-run I'll dig deeper, but right now it looks like a timing flake in the checkout control tests rather than something introduced here.

[zzet]

Hey @davalillo,

I found one issue I'd like to highlight: Template member methods are silently omitted

walkMqlTypeBody handles direct function_definition nodes but does not handle template_declaration. At the same time, emitMqlTemplateFunction explicitly skips templates inside a type.

As a result, valid templated MQL methods produce no method or member_of node, and calls inside them cannot be attributed to their enclosing method.

Suggested fix:

• Handle template_declaration in walkMqlTypeBody.
• Extract its wrapped function_definition and pass it through the method-emission path.
• Add a regression test asserting:
  • the method node exists;
  • the member_of edge exists;
  • calls inside the template method are attributed correctly.

Relevant code:

• walkMqlTypeBody
• emitMqlTemplateFunction
• MQL5 template-method documentation

[davalillo]

Thanks @zzet, spot on. I verified the finding empirically before fixing, and it was actually two shapes: a plain-declarator template member leaked out as a free function, while a pointer-return one (T *Create(...)) was dropped entirely, calls and all.

Fixed in 93a256d:

• walkMqlTypeBody now handles template_declaration, unwraps its function_definition, and passes it through addMqlMethodFromNode. The _method_L marker that path seeds keeps the func.def dispatch from double-emitting it, and a template wrapping a nested class keeps the class path.
• The regression test asserts exactly what you listed (method node, member_of edge, call attribution) plus the pointer-return form.

Fixing it also surfaced a sibling gap worth flagging: qMqlAll only had the plain declarator pattern, so any free function with a pointer return (CObject *New(), the idiomatic MQL5 factory shape) was silently dropped. The C/C++ queries carry pointer/double-pointer/reference variants; I copied those three patterns over and added tests for those too.

TestMQL passes with -race, LSP suite still 376/376. The full-package glibc abort mentioned in the PR body still reproduces with TestMQL skipped, so it stays pre-existing.