[jdrouet]

Breaking — for 7.0. Removes LocalIncludeLoader::default().

The type documents that it "doesn't allow to go fetch a template that is in a parent directory of the root directory". That guarantee fails in both directions.

It reads anything. The struct derived Default, giving root == PathBuf::default() == "", and Path::starts_with("") is unconditionally true. So every canonicalized path passes:

LocalIncludeLoader::default().resolve("file:///../../../../../../../../etc/hostname")
-> Ok("<contents of /etc/hostname>")

file:////etc/hostname gets there too — the fourth slash survives trim_start_matches("file:///"), and PathBuf::join with an absolute path discards the root.

It reads nothing. build_path canonicalizes the target but compares it against the root as supplied. Any root that is not already canonical matches nothing, so the loader rejects every input — including legitimate ones. The crate's own documented example is in this state: its root is relative, so it could never load a file. Same for mrml --local-loader ../partials, and for any root reached through a symlinked ancestor.

The containment check itself is right — with an absolute, canonical root I confirmed both traversal payloads are blocked. So this fixes the two ways its precondition gets violated, not the check:

• The Default derive is gone, so the empty-root state is no longer constructible.
• The root is canonicalized before comparison, so relative roots, .. and symlinked ancestors all behave. If the root does not exist, that is now an error rather than a silent fallback to the raw path — a fallback would just reinstate the second bug.
• The doc example is fixed, and the security note now states the guarantee precisely: the resolved file must sit inside the root, with .. and symlinks resolved first.

mrml-cli and mrml-python both build their root absolutely and call new, so neither needed a change.

Three regression tests, each checked failing before and passing after: a relative root loading a legitimate file, and both traversal payloads still refused under a relative root. should_handle_dots_with_existing_file used to pass for the wrong reason — its relative root rejected everything — and now passes because the path genuinely resolves to a sibling of the root.

Tests 621 -> 624.

[github-actions]

📈 This is how your metrics evolved 📉