test: cover the edge cases the suite was missing

[VijitSingh97]

What this is

A coverage-gap pass over the suite: an audit found behaviours, branches, and degenerate inputs in rigforge.sh that no test exercised, and this fills them. ~58 new assertions, tests/run.sh only — no new files, SHELL_FILES unchanged.

Every addition lands at the layer tests/README.md prescribes: logic / config-gen / command behaviour, driven by PATH stubs, hardware-independent, bash-3.2-safe. Each is the lowest layer that proves the behaviour honestly — nothing here needed a new /etc effect, macOS-native tool, or real hardware beyond what the existing layers already cover. Each test fails if the logic breaks (e.g. the commit-mismatch test asserts the clone directory is gone, not just that the error prints; the reboot-gate test asserts start/restart are absent from the systemctl call log).

Behaviours now covered

Area	Gap closed
check_prerequisites	The entire jq-bootstrap was untested: apt / dnf / pacman / no-manager / brew / no-brew / already-present
install_dependencies	dnf + pacman branches (right command, distro-appropriate package); no-manager warn-not-abort; the apt-cache-present linux-tools-<rel> branch
install_service (#audit A2)	Reboot pending → enable but NOT start (would run degraded under Restart=always); restart on rebuild; start on steady state; CPUPOWER_PATH resolved in ExecStartPre
compile_xmrig (#18)	Commit mismatch drops the clone (rm -rf xmrig) so the next run starts clean
config-gen	Bracketed-IPv6 url round-trips unmangled; single-pool tls:true reaches config.json; empty ACCESS_TOKEN → JSON null
restore	Corrupt / config-less archive fails and leaves the existing config.json untouched (no silent clobber)
uninstall	Without --yes, answering n aborts cleanly and reverts nothing
macOS	mac_start no-worker guard ("run setup first"); mac_stop launchd-owned delegation
tune numerics	_median/_stddev/_mean on empty & single sample; _watts_from_energy backwards counter with no wrap-max → empty (not negative); no-power fallback in _accept_better and _autotune_score
kernel/limits	compute_build_jobs sub-2GB max<1→1 floor; 1G fstab line on a fresh write; the rest of the xmrig.service hardening block

Residual gaps (deliberately NOT in this PR — the real completeness note)

These are gaps a stub can't honestly close; recording them rather than pretending they're covered:

• Post-GRUB-reboot autostart — CI now proves the decision (reboot pending → enable, withhold start). That systemd actually starts the miner after the reboot can't be proven by any runner; it's a manual release-gate observation (e2e-real confirms the enabled-but-stopped state; the across-reboot start is a human check).
• dnf / pacman on real hardware — the new tests prove dispatch (right command, distro-appropriate package). RigForge's real-hardware gate and CI are both Ubuntu, so a package-name drift on Fedora/Arch would be caught by no layer. Acceptable given the documented Ubuntu target, but named, not assumed-covered.
• Runtime vm.nr_hugepages value wiring — the reserved count is still proven only by proposed-grub.sh's isolated math; the sysctl stub swallows -w. The actual reservation is proven by e2e-real on real hardware. A CI logging-sysctl assert would only re-derive the math without proving anything new, so it's intentionally left to the release gate.
• tune_kernel /etc/modules MSR fallback content — the primary modules-load.d path is covered in-container; the fallback (systems without modules-load.d) is asserted only via the uninstall pre-seed. Low-value residual gap.

Bugs

None. Every branch exercised behaves correctly — the suite passing against previously-untested paths is itself the evidence. This PR adds coverage, it does not change rigforge.sh.

Verification

• make lint — clean (shellcheck + shfmt)
• make test — 666/666 local
• make coverage — 91.60% total (rigforge.sh 91.45%, util 96.15%) vs floor 80; kcov ran the full suite (689 checks). Patch-coverage gate is N/A — the change touches only tests/run.sh, so there are zero changed lines in the instrumented rigforge.sh/util.
• Real hardware: suite passes 689/689 on a live Ubuntu 24.04 / bash 5.2 rig (miner-0) — validation was fully stubbed, so the running miner was never disturbed and the rig was left as found.

🤖 Generated with Claude Code