skills: add triage-issue, for judging an issue against the architecture

[Yaraslaut]

Adds .claude/skills/triage-issue/SKILL.md. Repo-specific by design — the value is in morph's own invariants and morph's own worked examples, not in generic triage advice.

Why

Roughly a third of the issues acted on recently needed something other than what they asked for. Not carelessness: a symptom seen from inside one layer reads as a framework gap when it is really the architecture working.

• No framework seam for a model's own execute() to post background work and later update its own state #129 asked for a seam that must not exist (background work mutating model state, off-strand).
• codecov/patch under-reports coverage in templated code: llvm-cov's per-instantiation line records are never aggregated #92's premise was false — llvm-cov already aggregates the records it said nothing aggregates.
• QML bridge boilerplate: the duplication is O(QML surface), not O(models) — extract the two model-independent helpers #86's growth law was wrong — the boilerplate is O(QML surface), not O(models).
• core: HasLifetime and lifetime-bound Completion overloads (stage 1, draft) #150 solved half its problem with a mechanism (inheritance) too strong for its consumers.

Each was caught by the same few questions asked in the same order. So they are written down.

What's in it

Seven architecture invariants, each having already caught a real mis-framed issue, each cited to its authoritative source rather than to the skill file — a model does not know the framework exists and is single-threaded per ModelId; a requirement on consumers must compose rather than inherit; the value type reports the fact and the layer decides the policy; templates parameterised on application types are the design; spec-vs-code disagreement is itself the finding; a guarantee reimplemented per call site belongs in one place; and "would this still pass if the feature did nothing?"

The step it puts hardest is checking the premise empirically before evaluating the fix. That is the step that has changed the most verdicts and the one most often skipped. An issue reasoned out from reading source is a hypothesis; ten lines usually settles it.

Five worked examples with the actual evidence — including what the experiment turned out to be — so a reader sees the shape of a triage, not just its rules.

A failure-modes section recording two mistakes genuinely made here: undoing a deliberate fix without reading its history (the per-commit cache key existed to stop branches deleting each other's caches), and verdict-by-assertion.

Labels

Six triage: verdicts and ten area: labels, with the gh label create commands. Already created on the repo, and the verdicts established this session are already applied to 20 open and closed issues — so the taxonomy arrives populated rather than theoretical.

Verification

Every quoted source in the skill was checked against the file it cites; one quote didn't match and was corrected to the spec's exact wording rather than left as a paraphrase. Every issue number cited was checked to exist and match the described verdict.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!