V2.4.33

Publish the internal/DMA-RAM split to MQTT so Home Assistant can graph the long-uptime net-stack drain. Follow-on to V2.4.32: the per-cycle diag_log_heap() lands in /log → FTP, which means watching the multi-day trend meant pulling and diffing log files. The MQTT rich state already published heap_free / heap_min / heap_max_alloc — but those are the PSRAM-dominated totals, i.e. the same misleading numbers that hid the problem. This adds the gauge that matters.

Changes

• Three new rich-state fields (mqtt.c): heap_int_free, heap_int_largest, heap_dma_largest (heap_caps_* for MALLOC_CAP_INTERNAL / MALLOC_CAP_DMA), emitted next to the existing heap_* totals.
• Three HA discovery entities (mqtt_discovery.c): "Heap internal free / internal largest / DMA largest", data_size/B/measurement (so HA keeps long-term statistics), diagnostic category — same shape as the existing heap entities.
• Gated MQTT_RICH_STATE (PSRAM boards only). On the Heltec (no PSRAM) the total heap is the internal heap, so the existing heap_free/heap_max_alloc already serve as the internal gauge there — the split would only duplicate them. The split is informative exactly where total ≠ internal, which is the rich-state boards.

Why

heap_int_largest is the high-value one — the contiguous-block ceiling on internal/DMA RAM is what a sustained inbound TLS/OTA receive needs, and a shrinking largest over days is the fragmentation signature behind the OTA-upload stalls (see V2.4.32 / the long-uptime OTA-stall investigation). Graphing it in HA — with an optional alert when it drops below ~40 KB — catches a node before its OTA starts failing, instead of after. Diagnostic/observability only; no data-path change. Adds ~72 B to the rich-state JSON (well within the 1280 B buffer; observed publish ~676 → ~748 B).