Ship the example project skeleton with an initialized checklist (#648)

[PavloSEO]

Advances #648.

The gap

Epic #648's definition of done says project status on the examples/ skeleton prints a coverage figure matching a hand count. It did not:

$ seohead project-status --directory examples/project-skeleton
  "checklist": {
    "state": "not_initialized",
    "reason": "coverage checklist is initialized by project checklist setup, not project creation"
  },

The one project in this repository a reader can run the documented command against answered with a state instead of a number. The epic also asked that the skeleton be executable and walkable by the doc gates, and grep -rn "project-skeleton" tests/ docs/ scripts/ returned nothing — no test and no document named that directory.

Separate initialization is deliberate and documented (docs/PROJECTS.md), so project new is unchanged. The example is fixed the other way: it now ships already initialized.

What changed

• examples/project-skeleton/coverage.json is committed: 250 definitions, one per catalogue item, every one not_run with an empty records list. Nothing has been run against the synthetic site, so nothing claims a result.
• Initialization stamps each definition with the wall clock, which was the only machine-specific value the generated file carried. Those 250 observed_at values are normalized to the project's own created_at (2026-09-09T00:00:00Z) — the only observation time this example can honestly state — so the committed file is byte-stable and carries no generating machine's clock. No absolute paths, no fabricated results.
• scripts/generate_project_skeleton_coverage.py regenerates that file in one command, following the repository's other scripts/generate_*.py artifacts. It initializes a throwaway copy, so --check never writes into the shipped project and a stale committed file is compared against a first initialization rather than a reconciliation of itself.
• tests/test_project_skeleton_coverage.py is the gate: it fails if the file disappears, if item IDs drift from the live catalogue, if any shipped item claims a result, if a definition time is not the project's creation time, if the bytes differ from a fresh generation, or if project-status stops reporting counts. The drift and staleness messages name python scripts/generate_project_skeleton_coverage.py, matching the other generated-artifact gates.
• docs/PROJECTS.md describes the skeleton as already initialized and names it by path, in a runnable fence — so tests/test_docs_commands_execute.py now executes project status against the shipped project. That gate no longer creates a synthetic project when the documented directory already exists under examples/.
• examples/README.md gains a section for the skeleton; changelog.d/648.md records the change.

Validation

Hand count of the live catalogue:

checks (seohead.sf.core.registry.CHECKS): 161
skills (.claude/skills/*/SKILL.md + seohead/skills/*/SKILL.md): 23 + 7 = 30
scenarios (docs/scenarios/*.md, excluding README): 59
161 + 30 + 59 = 250

seohead project-status --directory examples/project-skeleton after the change:

"checklist": {
  "state": "initialized",
  "revision": 1,
  "counts": {"run": 0, "not_applicable": 0, "not_run": 250, "total": 250,
             "disabled": 0, "complete": 0, "stale": 0, "remaining": 250},
  "by_kind": {"check": {"total": 161, "complete": 0},
              "skill": {"total": 30, "complete": 0},
              "scenario": {"total": 59, "complete": 0},
              "custom": {"total": 0, "complete": 0}}
}

The printed figure matches the hand count.

The gate was proven to fail without the fix, three ways, each restored afterwards:

• committed file removed: all 6 tests fail, including state == "initialized" asserting against not_initialized — the exact pre-change behaviour.
• one item ID dropped: AssertionError: examples/project-skeleton/coverage.json has drifted from the packaged catalogue: regenerate it with 'python scripts/generate_project_skeleton_coverage.py'
• a succeeded record injected into one item: test_no_shipped_item_claims_a_result fails with Left contains one more item: {'reason': 'fabricated', 'status': 'succeeded'}.

Restored, the file matches a fresh generation byte for byte and all 6 pass.

Gate output:

$ ruff check .
All checks passed!
$ ruff format --check .
696 files already formatted

$ pytest tests/test_project_skeleton_coverage.py -q
6 passed in 0.05s

$ pytest tests/test_docs_drift.py tests/test_doc_counts.py tests/test_project_catalogue.py tests/test_docs_scenario_claims.py -q
227 passed, 1 warning in 10.85s

$ pytest -q
4528 passed, 2 skipped, 1 warning in 410.85s (0:06:50)

Note on the full run: on the uncommitted working tree, ten documented crawl-site lines failed with native scan provenance requires a clean source checkout. That is the build-provenance gate reacting to a dirty tree, not to this change; the run above, after committing, is clean.