fix(bundle): a strict TOML reader in loaders/toml.js

[exo-nikita]

Two commits, split out of #175 at review's request.

1. The move (c406cdb). The TOML reading that Cargo.toml, Cargo.lock and foundry.toml share moves out of cargo.js into stasis/src/loaders/toml.js, with the two bracket helpers (matchClose, splitTopLevel) the Rust scanner uses for cfg predicates and macro bodies. A pure move; the new file is in the package's files.

2. Strictness (review: "the reader silently misreads invalid TOML"). The reader took whatever it was handed and guessed. It now reads TOML 1.0 whole and throws a TomlError naming the file and line on anything else. The review's table:

input	before	now
"\uD800"	a lone surrogate	"\uD800" is not a Unicode scalar value
"a\qb", "a\eb"	aqb, aeb	invalid escape "\q" / "\e"
"abc	"abc as text	unterminated string
"a" junk	a	unexpected text after the value
["a" "b"]	["a"]	expected a comma or "]" after the array item
["a", b]	["a", "b"]	invalid value "b"
{ a = 1, a = 2 }	last wins	duplicate key "t.a"
{ a = "x"	{ a: '"x' }	unterminated inline table
[a] twice	merged	duplicate table [a] (at the second)
"abc at the end of the file	accepted	unterminated string
1_000, 0x1f, +inf, [1, "a"], version = 3	their text	1000, 31, Infinity, [1, 'a'], 3

Also: leading zeros, 1__0, 1., .5, -0x1, True, impossible dates, control characters in strings and comments, a lone CR, a trailing comma in an inline table, a table header after dotted keys or an inline table defined it, [[x]] over a table and [x] over an array of tables, text after a header.

Values are JavaScript ones: numbers (an integer past 2^53 is a BigInt, so a foundry gas_limit = 9223372036854775807 stays exact), booleans, strings (all four kinds), arrays, inline tables (null-prototype objects, so a key is only a key), dates and times checked and kept as their text. A leading byte-order mark is skipped, as cargo and forge do.

The API is one function: tomlEntries(text, { file }) → the document's [header]s and key = value pairs in order, each with the table path it lands at and its physical line span. The line-based helpers (logicalLines, TABLE_HEADER_RE, KEY_VALUE_RE, parseTomlValue, splitTomlKey) are gone. parseCargoManifest and parseCargoLock read by table path, so [dependencies.foo] features = […], [dependencies] foo.features = […] and foo = { features = […] } are one thing, and a [workspace.*] table alone marks a workspace root (as cargo sees it). Readers pass the file name: the manifest's project-relative path (crates/app/Cargo.toml:3: unterminated string comes out of createCargoContext), Cargo.lock, a foundry.toml's path; a --manifests config that can't be read is refused rather than carried unredacted.

Verification:

• toml-test 1.0.0 file list: 703 of 709 right; the six left are the invalid-UTF-8 byte cases, which a decoded string never holds (the bundle checks UTF-8 before reading).
• Differential against Python 3.11's tomllib: the 132 Cargo.toml / Cargo.lock / *.toml files of a vendored corpus (tokio, mio, serde, regex, libc, …) all agree; of 158 hand-written edge cases all agree but the BOM (accepted here) and CRLF inside a multi-line string (kept as written; tomllib normalizes).
• 781 KiB of corpus manifests and locks read in 40 ms; a 4.5 MiB synthetic Cargo.lock in 0.3 s.
• oxlint clean; full suite 2030 pass, 0 fail, 3 skipped (the cargo/rustc-dependent ones). New tests/toml.test.js covers every row above plus numbers, strings, dates, spans and the redefinition rules; the cargo and foundry tests cover the file:line messages end to end. One fixture changed: the redactFoundryToml test defined rpc_endpoints both inline and as a header, which is not TOML.

#175 is independent (the same move taken back out); whichever merges second needs a mechanical rebase: the import lines, and #175's .cargo/config readers onto tomlEntries.

🤖 Generated with Claude Code

https://claude.ai/code/session_015nbp2vC4jLPbGN8cmorxAL

---

Generated by Claude Code