issue_102: Separate artifact validity from review

[dieterbaier]

Closes #102.

Documentation only. #101 implements the axis and stays open.

An accepted mitigates relation must not become rejected when its target risk
stops existing. rejected is defined as a verdict about the claim — "considered
and intentionally rejected" — so using it to express the passage of time
rewrites a statement that was true and reviewed when it was made.

The artifact status enum had the same gap from the other side. It offered
deprecated without ever defining it, mixed review maturity with temporal
validity, and for Risk artifacts collided with risk-management vocabulary,
where an accepted risk is one the team deliberately carries.

issue_102: Separate artifact validity from review

ADR-009 keeps relationship status a review state and moves "does this still
hold" onto a separate validity axis: validity (active | retired) with
retired_on, retired_reason and retired_note. A retired risk keeps
status: accepted, because both statements are true at once — the description
is accepted knowledge, and the risk is gone.

Three options were weighed with a Pugh matrix: a closed value in the shared
status enum (-1), the separate axis (4), and a prose-only closure note
(0). The cheap option loses because a closed artifact has no review state
left, and because "closed" means nothing for an ADR or a component.

The metamodel document gains the artifact status table it never had. Every
status value now has a definition, including deprecated, which means "still in
force, discouraged for new work" — a component being phased out, not an artifact
whose subject is gone.

issue_102: Close the gaps the convergence check found

From the convergence check on #101. The first version delegated four semantic
questions to the implementation, where they would have become accidental
decisions in generator code.

Either endpoint, not just the target. A relation is inactive when its source
is retired too. "Active relation" has to mean the same thing from both sides.

Allowed and warned, not rejected. Rejecting a new relation to a retired
artifact would block legitimate historical links such as documents or
supersedes; a silent accept would hide stale references.

A residual risk stays active. Retirement means the subject is gone, not that
it is knowingly carried. Marking an accepted residual risk as retired would make
a present risk disappear from the active-risk view, so accepted-as-residual is
not a retirement reason.

Reason and explanation stop competing. retired_reason is machine-readable
for filtering, retired_note is for human readers. The vocabulary is
no-longer-applicable and removed for every type, plus mitigated and
materialized for Risk.

The metamodel also states the invariants the axis needs, including that an
absent validity is read as active by validator and generator — a JSON Schema
default annotates without materializing anything.

issue_102: Name the two values the split left behind

From the review. The section claimed two independent axes while superseded and
deprecated stayed on status carrying temporal meaning — a clean separation
that ADR-009 deliberately has not completed, because that migration overlaps
with deriving superseded_by from an outgoing supersedes relation (#67).

The lead paragraph now states the exception instead of implying it away, and
both values are marked "Transitional, temporal" where they are defined.

Convergence Check — authoritative, against 13510ca

Run immediately before integration, against the commit being integrated. It
replaces the provisional results; the one recorded against 9fe97ca was void.
No commit landed after 13510ca, local and remote heads are identical, and CI
validate is green.

Integration hazard, handled. #105 is stacked on this branch, and this
repository deletes head branches on merge. Deleting issue_102 would close #105
irrecoverably (#73), so #105 was retargeted to main before this merge.

Result: Converged.

#	Question	Answer
1	Intent and scope	Passed (assisted). All four acceptance criteria of #102 are met. The check found one scope change recorded only in the diff — the four semantics closed by the second commit — and it is now recorded in #102 as well. Disposition: fixed.
2	Behaviour	Not applicable (deterministic). No observable behaviour changes: the diff is two .adoc files under src/docs, and no script, schema, template or skill is touched. This question would have failed had the schema or validator moved with the documentation, which is exactly what the NOTE in the metamodel prevents.
3	Architecture impact	Passed (assisted). Two artifacts changed, both genuinely affected: the new ADR and the metamodel document it refines. No collateral edits to other chapters.
4	Decisions	Passed (human tier, applied). The decision is the ADR. Proportionality was checked against the four criteria in #102 before drafting. Three questions are explicitly left open rather than answered by accident, and the overlap with #67 on superseded/deprecated is recorded as coordination, not resolved silently.
5	Traceability	Passed (deterministic for the relation path). #102 → ADR-009 → refines DOC-04001-metamodel, addresses QS-004, introduces_risk R-004 → regenerated impact and traceability fragments. Validation reports no dangling targets.
6	Implementation and verification	Passed (deterministic). See below.
7	Documentation and delivery state	Passed (assisted). One review finding, fixed in 13510ca: the section described the axis separation as complete while two temporal values stayed on status. Stale references are gone — no accepted-as-residual remains anywhere. The not-yet-implemented NOTE is deliberately kept and names the schema constraint that makes it necessary. Deferred work has follow-ups: #101 for the axis, #67 for the superseded_by overlap.

Verification

Run	Result
./build.sh all on 13510ca	Ruby 29 runs, 144 assertions, 0 failures; CLI 4 runs, 13 assertions, 0 failures; JS 41 pass, 0 fail; adapters current
./build.sh build	Errors: 0. Renders build/architecture/index.html, no Asciidoctor warnings, both new xrefs resolve.
./build.sh validate	38 warnings, all pre-existing bidirectional-relation false positives tracked as #93. None names an artifact from this change.
node --test, all JS suites	41 pass, 0 fail
ruby -Itest test/validate_metamodel_test.rb	29 runs, 144 assertions, 0 failures
ruby -Itest test/validate_metamodel_cli_test.rb	4 runs, 13 assertions, 0 failures

Residual risk. The validity axis is documented before it is implementable.
Until #101 lands, metamodel/artifact.schema.yaml rejects the fields the
metamodel now describes. The NOTE in the document is the mitigation, and it is
removed by #101's acceptance criteria — a review-enforced guard, not a
deterministic one, because no check can read prose intent.

[dieterbaier]

The central decision is sound, the four findings from the convergence check are now explicitly resolved, and the documentation-only scope is respected. CI is green, and I found no implementation changes hidden in this PR.

One finding to resolve before merge: the current wording overstates the separation that ADR-009 actually delivers. status is called a pure review axis while superseded and deprecated deliberately remain temporal states in that same enum pending a follow-up decision. The ADR itself correctly acknowledges this transitional mixture, so the metamodel should say so as well.

Non-blocking delivery note: #101 still contains the earlier reason vocabulary (including accepted-as-residual) and still asks questions now decided by ADR-009. Since #101 is the implementation follow-up, its body should be synchronized with #102/ADR-009 before implementation starts.

GitHub does not allow a formal request-changes review on one's own PR, so this is submitted as a comment review.

[dieterbaier]

Re-check against 13510ca: the finding is resolved. The metamodel now states the transitional exception explicitly, marks superseded and deprecated as temporal values still carried in status, and links the deferral to the overlap with #67. #101 has also been synchronized with ADR-009/#102, including the endpoint semantics, residual-risk treatment, reason/note split, metadata invariants, and acceptance criteria. The current Validate run is green. No further findings from this review.