Parse remaining GRIT_POWER grammar extensions (INTABLE, TABINIT, SAFENAV, CCOMMENT)

[Entytaiment25]

Summary

Fills in the FiveM/GRIT-Lua syntax extensions this fork's parser/lexer didn't yet recognize, checked directly against citizenfx/lua at luaglm-548 (CMakeLists.txt GRIT_POWER_* options + lparser.c/llex.c), not just from memory of the feature names.

Each commit is one feature, self-contained:

• destructuring locals (local a, b in t) and destructuring reassignment (a, b in t) — the two grammar positions GRIT_POWER_INTABLE covers (localstat and restassign in the reference parser). Both desugar at parse time into the equivalent field-access assignment (a, b = t.a, t.b), so the rest of the checker treats them like ordinary assignments with zero changes elsewhere. For the reassignment form, each lhs target's own name/field selects the key read back out of t (x.y reads t.y, x[1] reads t[1]); targets without a literal name or index (x[y]) are rejected, matching the same restriction getassignkey enforces in the reference implementation.
• table constructor set sugar (GRIT_POWER_TABINIT) — .name = val as an alternative to name =/[expr] =, and either dot or bracket form may omit = expr, defaulting the value to true.
• safe navigation on calls/method calls (GRIT_POWER_SAFENAV) — ?. and ?[ already worked; this adds ?:, ?(, ?{, ?"...". The lexer already falls back to a bare "?" token for anything other than ?./?[ (confirmed by reading lex_qm, not assumed), so this is one generic suffix_handlers["?"] that delegates to whichever plain handler follows, restricted to the funcargs/method token set so a stray t??.b still correctly errors (caught by a second review pass, see commit history).
• C-style block comments (GRIT_POWER_CCOMMENT) — /* ... */, reusing the existing "long_comment" token so the parser needs no changes; treated exactly like --[[ ]] for line tracking, and like --[[ ]] it doesn't feed inline luacheck: ignore directives (only -- short comments do that today).

Deliberately not included: GRIT_POWER_EACH/BLOB/WOW/CHRONO/READLINE_HISTORY/DEFER_OLD are runtime/stdlib features (a __iter metamethod, C API additions, timers, REPL history, a library function), not grammar — nothing for a static parser to recognize. GRIT_POWER_NAMESPACE_SEL ships OFF in this repo's own CMakeLists.txt/makefile, so it isn't part of the actual reference build. GRIT_POWER_COMPOUND (+= etc.) turned out to already work end-to-end (compound_operators + the OpSet node), so there's no commit for it.

Known limitation (disclosed, not fixed)

All three in/desugar-style features share one artifact: if the right-hand base expression is itself undefined, you get one warning per destructured name/target, at the same line:column, since the desugar references one physical source occurrence multiple times. E.g. local a, b in undefinedTbl reports accessing undefined variable 'undefinedTbl' twice at the identical position, rather than the hand-written form's two distinct columns. Happy to add dedup if you'd rather have it, but didn't want to build a hidden-temp-local AST shape to chase a cosmetic duplicate without checking first.

Test plan

• Every construct verified against actual parser/checker output (not just "doesn't throw" — luacheck.check_strings reports syntax errors as report items rather than exceptions, so this required checking for code 011 explicitly, not pcall alone)
• Spec tests added per feature in spec/parser_spec.lua / spec/lexer_spec.lua, matching existing style and asserting real captured offsets, not guessed ones
• Regression sweep after each commit against all prior forms (old table-constructor syntax, plain division/idiv, existing ?./?[, existing comments, etc.)
• Full busted suite — not run locally (no lua/luarocks/busted in this environment, only luajit), relying on this repo's own CI to run it; flagging explicitly rather than silently omitting

[Entytaiment25]

*Claude Code Assistance