fix: don't treat Object.prototype members as reserved words

[mindlace]

Problem

readWord classifies a command-start word as a keyword with text in RESERVED_WORDS. in walks the prototype chain, so any Object.prototype member name matches, and RESERVED_WORDS[text] then hands setToken a function instead of a Token.

Two names get past the existing fast-path guard (fc >= CH_a && fc <= CH_z && text.length <= 8): toString and valueOf. Both are plausible names for a shell command or function.

The failure is silent — the bogus token makes the parser abandon the rest of the script, with nothing added to errors:

parse("toString foo")                     // commands: []
parse("valueOf x")                        // commands: []
parse("toString() { echo hi; }")          // commands: []
parse("echo hi; toString foo; echo bye")  // commands: [echo hi]   <- rest dropped

Inside a compound command it instead produces spurious errors:

parse("if toString; then echo a; fi")
// errors: ["expected 'then'", "expected 'fi' to close 'if'"]

The blast radius happens to be limited to those two names only because of the length <= 8 fast-path check, which is a performance heuristic — not something that should be load-bearing for correctness. Widen or drop that guard and constructor, hasOwnProperty, toLocaleString, etc. become reachable too.

Fix

Replace the in test plus second lookup with a single lookup guarded by typeof reserved === "number". Own entries are always numeric Token values; inherited members never are.

This also removes one hash lookup from a hot path — bench/self.ts is unchanged to slightly faster on every group (e.g. short 11.67 → 10.76 µs/iter, large 6.75 → 6.39 ms/iter, within noise but never worse).

I considered Object.create(null) / __proto__: null for RESERVED_WORDS, but null-prototype objects drop to dictionary mode in V8, which would cost more than it saves here.

Scope check

I swept the other string-keyed lookup tables for the same defect class:

Table	Lookup	Reachable?
RESERVED_WORDS	text in RESERVED_WORDS	yes — fixed here
UNARY_TEST_OPS	UNARY_TEST_OPS[val] === 1	no — value-guarded
BINARY_TEST_OPS	BINARY_TEST_OPS[nt.value] === 1	no — value-guarded
REDIRECT_OPS	REDIRECT_OPS[t.value] ?? ">"	no — keys are only lexer-produced operator strings, never user words

REDIRECT_OPS would return the inherited function rather than falling back to ">" if a prototype key ever reached it, but no code path can supply one, so I left it alone rather than adding an unreachable check to the hot path. Happy to harden it if you'd prefer belt-and-braces.

Tests

Two tests added to test/parser.test.ts, both failing before the change:

• Object.prototype member names parse as ordinary command names (toString, valueOf, constructor, hasOwnProperty, __proto__, isPrototypeOf)
• an Object.prototype member name doesn't truncate the rest of the script

Full suite: 1443/1443 passing. tsc --noEmit, oxlint, and oxfmt --check clean on the touched files.

[webpro]

Oof, nasty. Thanks for the find + PR!