v0.6.0

Added

• CpuInfo.cpu_stepping_int (defaults to 0). resolve() takes it as an optional arg and detect() fills it in on all three platforms
• zenmaster.init() re-exported at top level.
• smu.read_pm_table_full(family). Every backend was running the full mailbox sequence twice per read_pm_sensors() call: once for read_pm_table(), once again for read_pm_table_version(), because they were written as two separate functions that each redid the same version/address/transfer round trip. Now there's one fetch, and both older functions call into it.
• Linux ModuleStatus.reason can be not_installed or unsigned now, not just not_loaded. Previously "never installed," "installed but unsigned," and "installed but idle" all collapsed into the same reason, so anyone consuming this had to shell out to modinfo themselves to tell them apart.
• smu.close() (also at top level as zenmaster.close()) releases whatever the active backend is holding: the Windows PawnIO handle, macOS's DirectHW/IOKit connection, or the Linux SMU file descriptor. Nothing closed these before; a long-running process (a daemon that wants to drop and reacquire access) had no way to let go of them.
• More of smu's surface is re-exported at the top level: active_backend, pm_table_supported, send_mp1, send_rsmu, query_mp1, query_rsmu, read_pm_table, read_pm_table_version, read_pm_table_full. The CLI already needed all of these; now import zenmaster gets you the same surface without dropping down to zenmaster.smu.
• BackendUnavailable now appends a link to the wiki's Installation guide to whatever message it's raised with. Every "PawnIO isn't installed" / "ryzen_smu is too old" / "DirectHW isn't loaded" error now points somewhere with the actual setup steps, instead of leaving you to go find them.

Removed

• AMDFamily0F.bin, AMDFamily10.bin, AMDFamily17.bin, AMDReset.bin.

Changed

• send_arg() raises UnsupportedCPU for a family with no SMU support.
• send_arg() no longer swallows every exception as a hardware rejection. Only OSError, struct.error, and ctypes.ArgumentError count as SMU_FAILED now. A real bug in a backend used to look identical to the SMU just saying no.
• CLI shows help instead of doing nothing when you run it with no action flag and no tuning args. zenmaster --json by itself used to just exit 0 silently.
• hardware.detect() caches its result for the life of the process. It used to re-parse /proc/cpuinfo, PROCESSOR_IDENTIFIER, or sysctl and recompute the codename on every single call, which is why callers (U4L included) kept building their own memoization around it. resolve() is untouched and still takes whatever you pass it every time, since that's the one meant for feeding in values you didn't read off the real CPU.
• The mailbox send/query/poll/retry logic that macOS and Windows each had their own near-identical copy of now lives once in zenmaster/mailbox.py. Linux already shared its two backends this way; the other two platforms didn't.
• All three backends now check "has init() run yet" the same way: a bare _require_init() guard called at the top of send_mp1/send_rsmu/query_mp1/query_rsmu, raising SMUNotInitialized before any hardware I/O happens.

Fixed

• Windows: the PawnIO registry lookup (_pawnio_info) was a separate registry read for module_version(), module_version_ok(), module_status(), and is_available() each. It's read once and cached now.
• Windows: read_pm_table_full didn't check for a zero table version, unlike Linux and macOS, so a rejected version query could fall through and read from a garbage address instead of returning None.
• apply() now counts an unparseable value (--stapm-limit=notanumber) as a rejection. It already returned an error result for this case, but had_rejection stayed False, so a caller only checking that flag would think the preset applied cleanly.