Add mqttstats: headless Commodity for Home Assistant telemetry

[sidick]

Summary

• New mqttstats tool (src/amiga/mqttstats_main.c): a headless AmigaOS Commodity publishing Amiga telemetry (uptime, chip/fast RAM free, CPU model) to Home Assistant via MQTT Discovery, entirely configured through Workbench ToolTypes (amiga.lib's ArgArrayInit) and safe to drop into WBStartup: - no window ever, Commodities Exchange (Enable/Disable/Kill) is its only control surface.
• New portable src/tools/ha_discovery.[ch] module: builds HA MQTT Discovery topics/payloads, host-unit-tested (tests/test_ha_discovery.c).
• This is the telemetry piece of I2C/hardware sensor support (scope addition beyond the MQTT proposal) #6, scoped ahead of I2C sensor support per the issue discussion; I2C remains a follow-up.
• userdocs/mqttstats.md added (ToolTypes reference, Home Assistant setup, WBStartup instructions), wired into mkdocs.yml nav and tools/docs2guide.py for midge.guide.
• docs/ARCHITECTURE.md updated with the new file, Makefile's dist target now ships mqttstats in the release archive.

Test plan

• make test - 276 passed, 0 failed (26 new checks for ha_discovery)
• make lint - clean
• make m68k-docker - all 6 binaries (including mqttstats) build warning-free
• make guide + mkdocs build --strict - docs build clean
• On-target (Copperline, real AmigaOS 3.2.2 boot): verified full happy path (HOST configured) publishes discovery/availability/state topics correctly via mosquitto_sub, and verified the missing-HOST error path returns cleanly with no window and no Shell/Startup-Sequence hang

🤖 Generated with Claude Code

https://claude.ai/code/session_01QWq1CRLCQy9AaD9UtLRSM9

[sidick]

Added a Workbench icon (`icons/mqttstats.info`, via a new `tools/mkicons.py` ported from AmiAuth's generator) - without one, `mqttstats` couldn't actually be double-clicked or dropped into `WBStartup:`, which was a direct requirement. Ships alongside the binary in `make dist`; `userdocs/mqttstats.md` and `Installation.md` updated accordingly.

[sidick]

Follow-up round addressing feedback:

• Icon: now lists every ToolType from `read_config()` (not just four), all disabled/commented by default.
• Auth/config failures no longer silent: initial `MQTT_Connect()` failures now retry with capped exponential backoff (1s→32s) rather than quitting once - `mco_AutoReconnect` only covers a drop after a successful connect, not a failed first attempt, which matters for a WBStartup daemon racing the TCP/IP stack's own startup. A CONNACK refusal specifically (bad credentials, rejected client id, ...) is different - retrying won't fix it, so `mqttstats` pops a one-off `EasyRequest` (via `intuition.library`, opened only for this) before quitting. `mqtt.library` itself stays completely UI-free - it just returns a new public constant, `MQTT_CONNECT_REFUSED` (-102), added to `<libraries/mqtt.h>` (`mqtt.doc`/api-reference updated too); deciding what to do with it is entirely up to the caller.
• Two new sensors: Kickstart version and Workbench version (they can genuinely differ).

Also found and worked around a real m68k-amigaos-gcc `-O2` miscompile hit while adding those two sensors (garbage upper bits when a `UWORD` struct field reaches `sprintf`'s `%u` varargs) - tracked separately in #23 since it could affect other code built at `-O2`, not just this file.

All verified on-target under Copperline: retry/backoff timing, discovery+state publish for all 8 sensors now including the two new ones, and the corrected version values (confirmed against the test image's actual Kickstart ROM).

[sidick]

Fixed a real bug found by @sidick after checking a live sensor: uptime was reporting ~1,535,701,114s (~48.7 years).

Root cause: `TR_GETSYSTIME` doesn't measure time since reset - it returns the same absolute, calendar-rooted system clock `DateStamp()` is built on (seconds since 1978-01-01, set from the battery-backed RTC at boot), just with sub-second precision. It was picked specifically because it's immune to issue #8's clock-jump hazard, but that reasoning was based on a wrong assumption about what it actually measures.

Switched to `ReadEClock()`'s free-running hardware tick counter instead - genuinely resets to 0 at power-on, no notion of calendar date at all, and is exactly as immune to the clock-jump hazard the original code was trying to avoid. Verified on-target: uptime now reports a sane small value matching actual elapsed time.

[sidick]

Fixed the Exchange description truncation: reordered "Publishes Amiga telemetry to Home Assistant via MQTT" -> "Amiga telemetry to Home Assistant via MQTT" so "Home Assistant" lands safely inside the ~37-character cutoff Exchange applies on 3.1, instead of running right off the end into "...to Home Ass".