docs: two claims that were asserted rather than checked

[EricAndrechek]

Both are the same defect the 0.1.2 CHANGELOG is itself about, which is why they are worth fixing rather than leaving.

1. The CHANGELOG enumerated three files. Twenty-one qualify.

The 0.1.2 entry, in all four CHANGELOGs, ended:

"What remains naming it is the handful of places where the two-repository split is the actual subject — the README, SECURITY.md, and docs/reference/README.md."

Measured on main, excluding generated fixtures:

21 tracked files name "core repository" / "core repo"

All of them legitimately — describing the split is exactly what lint-public.sh deliberately permits. The problem is not the content, it is that the count was never checked and is wrong by eighteen.

The paragraph immediately above it in the same entry is the one explaining that a sweep reported itself complete while 148 of its targets survived, and that the fix which outlives that is a rule a linter can enforce. And then it enumerated a file list.

Replaced with the rule — describing the split is legal, naming something a reader cannot open is not, scripts/lint-public.sh decides — plus an explicit note of what the old count got wrong. Silently correcting it would repeat the original mistake: this entry has now been wrong once, and a reader who acted on the list deserves to know. All four CHANGELOGs edited identically; they are released together and tell one story.

⚠️ This already shipped inside 0.1.2 on all four registries. A fixed repository is not a fixed registry — the correction reaches npm, PyPI, crates.io and the Go proxy at the next release, not before.

2. A normative MUST that half the bindings ignore

docs/reference/artifact.md said:

"A loader SHOULD verify the hash before dlopen … and MUST when it came from a network."

with no indication that half the reference bindings don't. Measured:

binding	load-time hash verification
Python	✅ verify_hashes
TypeScript	✅ verifyChecksums
Go	❌ no hash path in multiversion.go at all
Rust	❌ registry.rs — "This crate does not hash (it takes no crypto dependency)"

A MUST that two of four implementations silently ignore is worse than no MUST, because a reader assumes the guarantee is there. The requirement stands — it now names who meets it, what the other two rely on instead (the release pipeline's own verification), and what a caller pointing either at network-sourced bytes must therefore do themselves.

Same treatment as the mandatory-symbol count in this file in #23: describe reality, do not quietly pick a winner. The divergence is #13 A2 and was already filed.

Verification

lint-prose, lint-public, markdownlint, dprint check clean; parity 9/9 (it reads these docs). Prose lint caught a British "licence" in my own new sentence before it left the branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_018ckJDLjWhAAttQVrYnQEJD