v0.2.2

v0.2.2

A live fetch now has to return something shaped like the dataset before it is
believed. A 200 carrying valid JSON with no feature records indexed to zero
features and was returned as source: "live" with no warning, so an endpoint that
had moved produced a confident empty matrix: search_features reported every
feature as nonexistent, and check_feature_support and lint_email failed with "No
client matches", which reads as the caller's typo. A body that fails the check
now takes the same route as a 500, down to the cache and then the bundle, with
the warning that says so.

The live-fetch test runs nightly instead of never. It has always been gated
behind CANIEMAIL_TEST_NETWORK=1 so a caniemail.com outage could not read as a
broken commit, but nothing ever set that variable, and the one path that talks
to upstream had no automated coverage at all.

Nothing changes for a caller getting real data.

v0.2.1

v0.2.1

Markup using the command and commandfor attributes is now linted.

caniemail added command and commandfor attributes upstream, and the title
tables only derived attribute names from singular "X attribute" titles — the
one existing plural, srcset and sizes attributes, was reachable only because
it had been written into a hardcoded table. Both now come from one rule, so
the next multi-attribute feature is picked up without a release here.

The offline dataset snapshot both surfaces ship moves from 2026-07-20 to
2026-08-10: one feature added, none removed, and thirteen verdict cells
changed across css-inset and html-popover. Detection output is unchanged
on every fixture.

No API changes. Anything reading 0.2.0's output keeps working.

v0.2.0

v0.2.0

lint_email returns a different shape. A finding carries only what its verdict
decides, with everything constant across a feature's findings hoisted into a
features legend keyed by slug; positions are "line:col-line:col" strings;
and version_requested has moved off the entries of check_feature_support's
support array onto the result. Anything reading those fields off a finding
follows the legend instead.

Nothing was removed from what a result tells you, and a lint of a realistic
newsletter against all 48 clients costs about half what it did — 75KB to 41KB,
roughly 19k tokens to 10k. Findings now report every occurrence of a feature
rather than the first.

Also: the MCP server revalidates its dataset every 15 minutes instead of
loading once at startup, a version pin resolves per client instead of failing
the whole call, the dataset cache is covered by tests against a loopback
server, and a weekly workflow refetches the snapshot so it cannot decay
unnoticed.

See CHANGELOG.md for the full entry.