feat(luau): add Luau/Roblox language support (issue #52)

[CyanoTex]

Summary

Draft skeleton for issue #52. Adds a luau language tier so Rojo-synced Roblox projects produce a real dependency graph instead of the reported 21 disconnected nodes.

This PR wires the full language path end-to-end with the exception of Rojo default.project.json-aware resolution for game.<Service>... absolute instance paths, which is deliberately scoped to a follow-up.

What's in scope

• LanguageSpec(tag="luau", ...) in the central registry, replacing the existing git-blame-only lua passthrough spec. One source of truth for both .lua and .luau (Luau's grammar is a strict superset of Lua 5.1).
• "luau" added to LanguageTag Literal in models.py so the extension flows through FileTraverser/ASTParser instead of being filtered out at models.py:61.
• LANGUAGE_CONFIGS["luau"] entry in parser.py mapping function_declaration → function and type_definition → type_alias.
• queries/luau.scm with captures for symbols, require(...) arguments, and call sites.
• resolvers/luau.py handling:
  • String literals: require("rel/path") — relative then stem fallback.
  • script / script.Parent chains — Roblox semantics where the first Parent is identity (the importer's container directory) and each subsequent Parent walks up.
  • game.<Service>.Path... — registers as external node; full Rojo-tree resolution is the next PR.
• tree-sitter-luau>=1.2,<2 added to pyproject.toml dependencies.
• docs/LANGUAGE_SUPPORT.md updated to document the new tier.
• Unit tests: TestLuauParser in test_parser.py, test_luau_resolver.py for the resolver. Absolute instance-path resolution is pinned with an xfail(strict=True) test documenting the expected end state.

What's deliberately NOT in scope

• Reading default.project.json to map service subtrees to filesystem directories. The shape is documented in the resolver's module docstring and the failing xfail test. Planned home: core/ingestion/dynamic_hints/rojo.py.
• Bindings / heritage extractors. Luau modules typically don't use class inheritance; revisit if needed.
• .lua files in pure-vanilla-Lua projects (not Rojo) will now be AST-parsed where before they were git-blame-only. This is aligned with issue [Feature] Luau/Roblox support: dependency graph parsing and language detection #52's stated goal, but flagging explicitly since it's a behavior change for non-Roblox Lua repos.
• The pre-existing broken repowise.core.ingestion.parsers reference in pyproject.toml's packages list (unrelated to this PR; surfaced in PR feat: add language support for wiki generation #75 review). Left untouched per the maintainer's preference to keep that fix in its own PR.

Verification

uv run pytest tests/unit/ingestion/ -q
# 227 passed, 1 xfailed
uv run ruff check packages/core/.../luau.py .../registry.py .../parser.py .../models.py .../resolvers/__init__.py tests/.../test_luau*.py tests/.../test_parser.py
# All checks passed

Open questions / risks

1. tree-sitter-luau ABI pin. The published wheel's [core] extra pins tree-sitter~=0.22, while repowise requires >=0.23,<1. Installing tree-sitter-luau without the [core] extra (which is what the standard dep resolution does) worked locally against tree-sitter 0.23.x — verified by test_parses_require_imports exercising the loader. Worth a CI run against each matrix slot to confirm.
2. .lua ownership. Taking .lua into the luau spec is a judgment call. Pure-Lua projects will now get AST-parsed; vanilla Lua parses cleanly as a Luau subset. Open to splitting the extensions if preferred.
3. Resolver for game.<Service>.... Currently returns an external node rather than silently matching the nearest-stem file. This matches the maintainer's "no silent fallback" norm from PRs feat: add OpenRouter as LLM and embedding provider #56 / feat(zai): adopt tier framework for plan-aware rate limiting #83, but means these edges won't land in the graph until the Rojo reader ships.

Test plan

• uv run pytest tests/unit/ingestion/ -q — green (230 passed, 2 xfailed)
• uv run ruff check on changed files — clean
• uv run ruff format --check on changed files — clean
• Run against a real Rojo project locally via repowise init — done against a private Rojo project + upstream luau-lang/luau (see comment for numbers); Rojo-tree resolution remains pending the follow-up PR
• Confirm CI ABI compatibility across the matrix

[CyanoTex]

Three follow-up commits pushed on top of 6e5c199 — all in the spirit of test-plan box 4 (running init against real Luau projects). Numbers and rationale below.

Commits

• fix(luau): string-literal requires never entered the resolver — parser.py:705 strips quotes from the captured @import.module text before calling the resolver, but the resolver's string-literal branch keyed off arg.startswith('"'). Result: every require("./path") fell through to the external-node fallback in production. The unit tests passed because they fed explicitly-quoted strings, which the production parser never emits. Fix detects literal paths by elimination (not script.*, not game.*, not @alias) and rewrites the tests to mirror what the parser actually produces.
• feat(luau): resolve :WaitForChild / :FindFirstChild require chains — On real Rojo code, script.Parent:WaitForChild("Foo") is vastly more common than the bare dot-chain form the original _SCRIPT_RELATIVE regex matched (it's the race-safe idiom against Rojo's filesystem-sync ordering). Normalizes both method-call forms to the dot shape before regex matching; the path-walking logic in _resolve_script_relative is reused unchanged. Variable-aliased chains (SharedFolder:WaitForChild(...)) still go external — that needs lightweight flow tracking and is deliberately out of scope.
• test(luau): pin .luaurc @alias resolution as a follow-up (xfail) — Second known follow-up alongside the existing Rojo xfail. .luaurc is Luau's official require-by-string alias mechanism (parallel to tsconfig paths). Needs a .luaurc reader; shape is documented in the xfail.

Runtime-verified numbers

Ran repowise init --index-only against two Luau codebases before and after:

Private Rojo project (161 .luau files across src/client, src/server, src/shared):

	Before	After
Internal (luau→luau) import edges	13 (2.8%)	333 (72.2%)
External edges	448	128

~26× more resolutions. The 128 remaining externals are all known-unsupported shapes: variable-aliased chains (out of scope here), and service-based lookups (ReplicatedStorage:WaitForChild(...), game:GetService(...)) which are in the same family as the existing Rojo follow-up.

Upstream luau-lang/luau (237 Luau files, mostly tests/ and bench/):

	Before	After
Internal (luau→luau) import edges	1 (0.4%)	113 (45.0%)
External edges	250	138

~113× more resolutions. The 138 remaining break down as:

• 107× external:script.Parent.bench_support — correct behavior; these are the intentionally-non-matching arm of the require(script.Parent.bench_support) or require("../bench_support") or-chain idiom in bench/*. The ../bench_support branch now resolves (was previously 0 internal, now all 107), so the pattern works end-to-end.
• 24× @alias references — the .luaurc follow-up (pinned as xfail by the third commit).
• ~7× scattered test-fixture edge cases.

Suite

uv run pytest tests/unit/ingestion/ -q
# 230 passed, 2 xfailed

Both xfails are the strict-mode documentation-only tests pinning the two deferred follow-ups (Rojo project-tree resolution, .luaurc alias resolution).

One side-finding worth flagging separately

While running init to generate the numbers above, I hit what looked like a serious data-loss bug in packages/cli/src/repowise/cli/mcp_config.py unrelated to this PR — JSON parse errors on an existing ~/.claude/settings.json silently discard the user's config. Filed as #93 with repro and recommended fix, so it's out of this PR's scope and you can triage independently.

-Opus 4.7