fix: bracketed link labels drop the edge; add released_in (#48, #49)

[RichardHightower]

Two independent fixes, one commit each, plus the release bump.

Commit	Issue	Change
fix: match link labels containing brackets	#48	LINK_RE accepts balanced brackets. One regex, one call site.
feat: add released_in to KNOWN_RELS	#49	One frozenset entry plus the three prose copies.
chore: release 0.4.0	both	Four manifests, README, CHANGELOG.

Why #48 is worth reading first

LINK_RE's label class was [^\]]+, which stops at the first ]. So a
bracketed label matched nothing:

>>> LINK_RE.findall('- [[AREA NAME]](/requirements/area-name.md)')
[]

The important part is what that does downstream. It produces a missing edge,
not a broken one — and validate reports broken link only for edges that
exist. The orphan check needs a concept to have neither inbound nor
outbound links, so any concept carrying one outbound link loses its catalog
backlink with zero diagnostics. I built that bundle to check: validate
returned 0 issues and backlinks returned [].

The concept itself is never lost — load_bundle globs every file, so it stays
queryable by path. Only the inbound edge goes.

This matters because the capture plugins interpolate concept titles into catalog
links unescaped, and bracketed titles are ordinary in exported wiki content
([AREA] prefixes, [DEPRECATED] suffixes). Companion issues cover that
producer side; this is the reader half, and it is the half that also repairs
bundles already on disk.

The regex, and two ways I got it wrong first

LINK_RE = re.compile(r"\[((?:\[[^\[\]]*\]|[^\]])+)\]\(([^)]+)\)")

The label alternation keeps [^\]] — the entirety of the old language — and
adds balanced pairs, tried first. Keeping the old branch is what makes it a
superset. Two earlier drafts regressed:

• An escape-aware branch (\\.) swallowed the closing bracket on a label ending
  in a backslash, so [a\](/c.md) stopped matching.
• * instead of + started matching [](/x.md), an empty label the old pattern
  rejects — which would have added edges to existing bundles.

Measured, rather than asserted:

Check	Result
20k generated links (trailing \, embedded ], multi-bracket labels)	0 lost
sample-okf edge sets, old vs new	identical — 0 lost, 0 gained
[](/x.md)	still unmatched, as before
"[" * 200	< 1 ms, no backtracking blowup

test_link_re_is_a_superset_of_the_plain_form pins both regressions so they
cannot come back.

#49

released_in always worked — the guard passes unknown non-empty rels through
unchanged. Only the vocabulary was missing, so a bundle modelling releases got
one info line per shipped work item.

Worth noting this repo already has a release axis on the worklog side:
milestone is a core work-item field and bin/ia_graph.py turns it into
edge(key, "targets", "release/" + milestone). Whether the two should be
unified under one name is a design question I have not assumed an answer to.

The rel list is duplicated in three prose files, so those move together and the
new test pins them to the frozenset.

Verification

python3 tests/test_okf_graph.py          # 29/29
bash tests/test_okf_curate.sh            # ok
python3 scripts/okf-graph.py validate sample-okf --strict   # 0 errors, 0 issues

Both new tests fail on the old regex (27/28) and pass on the new one.

Closes #48
Closes #49