issue_69: Separate what a risk affects from what mitigates it

[dieterbaier]

Closes #69. Two commits, one per problem.

1. The register labelled harm as mitigation

fa0c854. The Risk index headed a column Mitigation/action and filled it
from the risk's outgoing affects relations — which point at what the risk
endangers. This repository's own register was the clearest case: every risk
names a quality scenario there, so it read as if the scenario mitigated the risk
it is threatened by.

mitigates was in the relation enum and no generator had ever read it.
grep -n mitigates scripts/validate-metamodel.rb returned nothing, while four
ADRs here declare it. Relations are only ever authored outgoing, so the question
"what mitigates this risk?" can only be answered from the incoming direction —
which ArtifactRenderHelper already had the artifacts for.

The column is split rather than replaced, per the discussion on #69: what a risk
affects is worth seeing next to what is being done about it.

Before, for this repository's own R-001:

ID	Risk	…	Mitigation/action
R-001	AI Suggestions Are Mistaken For Reviewed Architecture Truth	…	QS-004-ai-suggestion-reviewability

After:

ID	Risk	…	Affects	Mitigation/action
R-001	AI Suggestions Are Mistaken For Reviewed Architecture Truth	…	QS-004-ai-suggestion-reviewability	ADR-007-artifact-provenance-metadata

That ADR-007 link is data this repository has carried all along.

2. Prose assessments collapsed the table

1bbb773. Likelihood, Impact and Priority were lifted verbatim into
columns one tenth of the table wide. Nothing said those fields had to be terse,
so writing the explanatory sentence an assessment actually wants wrecked the
generated register — discoverable only by rendering it.

summary_cell takes the leading sentence, so the verdict reaches the register
and the reasoning stays in the artifact. Only a real sentence boundary counts, so
Medium (rises when hosted) survives whole.

Truncating silently would swap one invisible rule for another, so
templates/risk.adoc and skills/risk/SKILL.md now state it, and the template
gained a place to put the reasoning.

Verification

• ./build.sh test, validate, check-adapters, generate — all exit 0, and
  the worktree is clean after generate.
• Two new tests in test/validate_metamodel_test.rb. Both were confirmed to fail
  against main before the fix — the register test on the column header, the
  truncation test with NoMethodError.
• The 38 pre-existing bidirectional-relation warnings are unchanged; they are on
  main too and are not touched here.

Provenance

Found while writing a second risk artifact in a project that consumes the
toolkit. The register only became visibly wrong once there was a second row to
compare against.

🤖 Generated with Claude Code

[dieterbaier]

Review-Ergebnis: sieht gut aus. ✅

Die beiden in #69 beschriebenen Probleme sind sauber und an der richtigen Stelle behoben:

• Der Generator trennt jetzt korrekt zwischen dem, was ein Risiko gefährdet (affects), und dem, was es mindert (eingehende mitigates-Relationen). Damit wird die Semantik aus dem Metamodell nicht länger im wichtigsten Überblicksartefakt umgedreht.
• incoming_relation_sources folgt konsequent der bestehenden Regel, Relationen nur ausgehend zu pflegen und die Gegenrichtung bei der Generierung abzuleiten. Sortierung und Fallback sind deterministisch.
• Die Bewertungsspalten übernehmen mit summary_cell nur noch das führende Urteil. Gleichzeitig wird diese Projektionsregel nicht heimlich eingeführt, sondern sowohl im Risk-Skill als auch im Template erklärt.
• Die Tests decken die falsche Spaltenbelegung sowie Satzgrenze, Klammerzusatz und leeren Wert ab.
• Der GitHub-Workflow Validate ist für den aktuellen Head erfolgreich durchgelaufen.

Besonders gut: Affects wurde nicht einfach entfernt. Beide Richtungen sind für ein Risk Register relevant und stehen nun mit ehrlichen Überschriften nebeneinander.

Von meiner Seite keine blockierenden Findings. Eine formale Freigabe lässt GitHub nicht zu, weil der PR vom verbundenen Benutzer selbst stammt.