Restore 100% test coverage #201

Context

pyproject.toml historically set fail_under = 100 for the coverage
gate. The recent rewrite temporarily dropped coverage from 100% to 87%,
so the gate was relaxed to 85% to land that work.
The relaxation is documented but it's debt — the project's quality
bar said 100% and the goal is to get back there.

What this issue tracks

Bring test coverage back up to 100% so fail_under = 100 can be
restored.

Why it matters

The 100% rule is a discipline, not just a number. It catches dead
code, forces edge cases to be tested, and makes deletions safe (if
something's only covered by the test that imports it, deleting the
test catches the unused code). Letting it slide is a one-way ratchet.

Where the gaps are

Run uv run coverage report --skip-covered for the current view.
Headline gaps from the last run (~297 missed statements + 135
partial branches across 24 files):

• src/openrange/core/episode.py (~70 missed) — advance() until /
  timeout paths, checkpoint/restore/fork, _terminal_state
  branches, NPC error swallowing, close(). Needs new tests after
  the dashboard tests were dropped.
• src/openrange/dashboard/ (~53 missed) — same root cause: old
  dashboard tests were deleted during the rewrite; replacements not
  yet written.
• src/openrange/packs/cyber_webapp_offense_v1/ modules — error
  branches in llm_generation (LLM call fail / bad JSON / bad
  source), sampling (validation paths in sample_int /
  weighted_pool), codegen/handlers (each kind branch in
  _default_handler_body, AST splice failure paths), NPC factory
  bad-config paths.
• Smaller gaps in core/builder.py, core/snapshot.py,
  core/runtime_backing.py (defensive paths).

Where to start

• pyproject.toml — the fail_under setting + the coverage section.
• tests/ — the existing tests as model for what to mirror.
• Run uv run coverage report --show-missing after a test run to
  see exact missed lines.

Acceptance

• fail_under = 100 restored in pyproject.toml.
• CI passes.

Notes

This is mechanical but real test-writing work — probably ~200 lines
of test code spread across several files. Good help wanted task
for someone who wants to learn the codebase by exercising every
branch.

Some of the older branches that lost coverage are honestly defensive
code that's never going to fire in practice. For those, a
# pragma: no cover is fine — but don't reach for it as the first
move; write the test if you can.