compiler: stamp node ids once, at compile time

[davydog187]

Follow-up to #400. That PR keyed scope resolution and codegen by meta.id, stamping ids in Lua.Parser.parse_chunk/1 and then again in Lua.Compiler.compile/2 — the second stamp is required for correctness (a Lua.AST.Builder chunk, or a parsed chunk with hand-built nodes spliced in, arrives unstamped or partially stamped, and an "already stamped?" early-out would cost the very walk it saves). So every parsed chunk paid for two full AST walks.

Only Lua.Compiler.Scope and Lua.Compiler.Codegen read meta.id — nothing between parse and compile touches it (verified across lib/ and website/). That makes the parser-side pass pure duplication, so this drops it.

Result

One walk per compile instead of two. Interleaved A/B (8 rounds × 20 iterations, min-of-rounds, comparing parse against parse + the removed Ids.assign):

chunk	parse now	parse + old stamp	saved
api.lua (32 KB)	1102 µs	1561 µs	458 µs (29.4%)
strings.lua (13 KB)	466 µs	688 µs	222 µs (32.2%)

This is strictly better than both current main and pre-#400: every path that compiles still gets numbered exactly once, and the hand-built path that #400 fixed stays fixed.

Behavior change

Lua.Parser.parse/1 (and parse_raw/1, parse_structured/1, parse_chunk/1) now return chunks with meta.id unset — ids are a compiler concern, established at the compiler's boundary. Node ids were introduced in #400 and have never shipped in a release (latest tag is v1.0.1, which predates it), so no published contract changes. Docstrings on Lua.Parser.parse_chunk/1 and Lua.AST.Ids updated to say where stamping happens and that it reassigns from scratch.

test/lua/ast/ids_test.exs now applies Ids.assign/1 before asserting — including the idempotence test, which becomes the real property (assign(assign(c)) == assign(c)) rather than a statement about parser output.

Validation

• mix compile --warnings-as-errors, mix format --check-formatted: clean
• mix test: 2651 passed
• mix test --include lua53: 2671 passed, 16 skipped
• Confirmed by hand that an unstamped chunk straight from the parser compiles correctly, and compiler: key scope resolution by node id and dedupe upvalue descriptors #400's hand-built-AST regression test still passes.

https://claude.ai/code/session_01BYHGFLoHBJAsUrTzjVp5nC