feat(test-placement): add an authoring step and close the L1/L2 gap

[gustavolira]

The problem

The skill could route a developer correctly to "L2, in rhdh" — and then have nothing to say.

Steps 2–3 conclude with a layer, but Step 4 was organised by repo, and its rhdh
section covered only L3, L4a and L4b. L1 and L2 existed nowhere except as two cells in the
Step 3 ladder table. Across the whole skill, mockServices, backend-test-utils and
frontend-test-utils appeared zero times.

So the decision output and the creation guidance did not line up. I suspect that mismatch
is why L1/L2 fell through — it follows from the organising axis, not from an oversight.

What changed

Step 4 — the rhdh section is reordered cheapest-first to match the Step 3 ladder
(L1 → L2 → L3 → L4a → L4b) and gains the missing L1 and L2 entries. Each names a template
that exists on main today:

Layer	Template
L1	plugins/scalprum-backend/src/service/router.test.ts
L2	plugins/scalprum-backend/src/service/router.integration.test.ts

Step 5 (new) — turns the skill from an advisor into an author. It is a procedure,
not a knowledge dump: open the template Step 4 named, consult rhdh: docs/testing.md,
mirror the imports and setup, run the layer's command, apply a per-layer checklist.

Why the API details are not in the skill

Deliberate. The Backstage test surface moves — @backstage/test-utils was renamed to
@backstage/frontend-test-utils, and the package went 0.5.1 → 0.6.1 inside 123 commits of
main. Nobody bumps a skill when a package is renamed, so an API name pinned here goes
stale silently and starts telling people to import something that no longer exists.

A path to a file on main cannot go stale the same way, and Step 5 carries a glob fallback
for the case where the template itself is moved. Same rule the skill already follows for
layer-migration-matrix.md: the decision lives in the skill, the detailed reference lives
in the repo.

This depends on rhdh#5140
(RHIDP-13234), merged, which added docs/testing.md — the L1–L3 authoring reference Step 5
points at. Added to References.

One opinionated line

Step 5 states that a test must be justified by the failure it catches, not by a coverage
delta — Codecov is informational: true on both the project and patch statuses in rhdh
and rhdh-plugins, so no coverage number can block a PR. Worth encoding, since a PR
justified that way has already been rejected in review.

Also

The frontmatter description gains "how do I write a test for this" / "write a test for X"
triggers. The previous list was all where-questions, which would not fire for someone
asking the skill to actually write the test.

Verification

• All eight rhdh paths cited in the diff verified against main via the GitHub API.
• Frontmatter is valid YAML; name matches the directory and the description is 638 chars
  (tests/unit/test_skill_structure.py requires 20 < len <= 1024).
• Diff is +39 / −7 on one file; the skill is 152 lines.

Per AGENTS.md this is a behaviour change, so it wants a patch tag after merge.

[schultzp2020]

/lgtm