fix: refuse a build that emitted no landing

[toastygm]

Since package-build 15 (package-build#182) a package's landing is published at /<package>/homepage-<shortcode>/, and Hugo separately generates /<package>/index.html — chrome around an empty <main>, authored by nothing. The guard's root check was written when those were one file, so losing the landing used to mean losing that file. It no longer does: a homepage-mode build that emitted no landing at all leaves one index.html at the package root, tallies one page, passes every check, and replaces the live site with a blank page.

Require the addressed landing. The difficulty is that a pre-#182 tree and a collapsed post-#182 tree are identical on disk — one index.html at the package root and nothing else — so the tree cannot say which it is.

The signal is the installed @heroiclands/package-build. npm ci has already run by the time the guard executes, so node_modules/@heroiclands/package-build/package.json states the version that actually built this tree. From 15 on, no landing is a collapsed build and is refused; below 15 the package root is the landing and the tree is counted exactly as before.

Not the declared range. A range is a claim about what may be resolved, and permission is not resolution: >=9.0.0 permits 15 while the lockfile pins 9, so reading that range as new-shape would refuse a legitimate old-shape build — the one outcome this must not produce. It would also put semver-range comparison (9 || 15, *, a dist-tag) into shell to answer a question the resolved version answers with a string compare. Two other signals were considered and rejected: the _redirects entry that sends /<package>/ to the landing is authored by the package, so its absence proves nothing; and testing <main> for emptiness in <dir>/index.html would couple the guard to the Hugo theme's markup, which is precisely the coupling this workflow exists to avoid.

An unreadable version is refused, like every other guard in this file. It is only consulted once the landing is already missing, so the tree is under suspicion by then; "which shape is this?" going unanswered is not a reason to deploy it.

The landing is matched with -s rather than -f, so an empty landing is a collapsed landing rather than a discountable one.

What is not touched. The root check stays the separate assertion it is — the deployment root exists and is non-empty. The tally is unchanged, so the homepage licensing bound still refuses a genuine second page in both shapes.

Verification

The guard step's run: body was extracted verbatim with yq and driven against real trees, each built from a pristine git archive origin/main extraction plus npm ci && npm run build:site. The three pre-#182 trees emit no _headers, so the workflow's own default-_headers step was extracted and run against them first, as it is in the job.

tree	installed package-build	expected	result
kethira, as built	15.0.0	pass	pass — 2 index.html file(s); 1 page(s)
kethira, landing deleted, Hugo's root kept	15.0.0	reject	rejected — "the build emitted no landing"
harnadventures, as built	9.0.0	pass	pass — 1 page(s), landing is the package root
harnensemble, as built	10.0.0	pass	pass — 1 page(s), landing is the package root
hm3, as built	9.0.0	pass	pass — 1 page(s), landing is the package root
harnadventures, root index.html removed	9.0.0	reject	rejected by the root check
kethira + a genuine second content page	15.0.0	reject	rejected — 2 page(s), licensing bound
kethira, no landing, no node_modules	—	reject	rejected — version unreadable
kethira, landing present but empty	15.0.0	reject	rejected

shellcheck -s bash is clean on the extracted body, actionlint is clean on the file, and it parses as YAML.

What this does not do

#21's first two acceptance criteria are not met and the issue stays open for them: harn-adventures (9.0.0), harn-ensemble (10.0.0) and HarnMaster-3-FoundryVTT (9.0.0) are still on pre-#182 majors, and #20's transitional branch is still here because they need it. Bumping those three crosses five or six majors each and is three content migrations, not a version bump — and the hole would stay open for sohl-kethira-basic, the only caller on 15, for as long as they took. This closes it now for any caller on the new shape while leaving the old shape working; the branch and the version read both go when the last caller reaches 15. Reasoning recorded on the issue.

Refs #21