Add scripts/build_browsesafe_set.py: a seeded, label-stratified sample of
the BrowseSafe-Bench test split at a pinned revision (row ids committed in
evals/labelled/browsesafe-test-sample.json, dataset not vendored), its
conversion to the guard's labelled-set JSONL (html or reader text), and the
share of page text inspected at the 6,000-character clip.

Labelled-set evaluation now applies the spec's runtime clip
(RailSpec.labelled_state) and records failed decisions as failure rows
instead of aborting, with per-label band counts in the summary.

…#75)

A --parquet input went straight to read_rows, so bytes from another revision
were stamped with the pinned revision and digest. Every command now verifies
the source's sha256 first; an offline regression checks a mismatching file is
refused before decoding or writing output.

The .part file is hashed and only promoted to the cache path when it matches
PARQUET_SHA256; a mismatching download is removed, and a mismatching cached
file left by an older run is discarded and fetched again, so a bad response no
longer leaves sample, build and coverage stuck on poisoned bytes.

Resolve CHANGELOG.md by keeping BrowseSafe-Bench sample entries alongside
mainline MCP tool annotations and ticket-router JSONL fixes.

Co-authored-by: Danyang Chen <birfychen@gmail.com>

Keep both Unreleased Added entries: BrowseSafe-Bench sample / clip coverage
and the offline public231 quality consumer from #69.

Keep the BrowseSafe-Bench sample, labelled-state clip and failure rows
together with main's coverage policy, public injection set, and the
changelog, NOTICE and license entries that landed on main.

Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>