[gabrielmancini]

nif_esp_nvs_fetch_binary() (esp:nvs_get_binary/1,2) opens an NVS handle and closes it only after reading the blob. Three returns after a successful nvs_open skip nvs_close:

• the key does not exist ({error, not_found}, which esp:nvs_get_binary turns into undefined)
• nvs_get_blob fails while reading the size
• memory_ensure_free fails (the not_found tuple or the binary)

Each of them loses the handle ESP-IDF allocated in nvs_open, about 72 bytes of heap. Reading a missing key now and then is common, for example an optional setting, and it went unnoticed. Reading one on every loop does not: we read an optional setting on every 2 s poll, and an ESP32 without PSRAM lost ~1.2 KB/min and aborted with Cannot handle out of memory every ~20 minutes.

The fix calls nvs_close before each of those returns. put_binary and both erase NIFs already close before returning.

Test. test_nvs (ESP32) writes a key so the namespace exists, then reads a missing key 500 times as a warm-up and 500 times more, and checks the free heap (esp32_free_heap_size) did not shrink by 4 KB or more across the second batch.

• QEMU esp32, without the fix: {leaked,36032}, 36 KB lost in 500 reads; with it, test_nvs passes (test_socket fails here as it does on plain release-0.7, no network inside QEMU).
• On hardware: the same board ran 74 minutes with a flat free heap (47–49 KB) once our firmware stopped reading the key on every poll. The VM fix itself is proven in QEMU only so far.

These changes are made under both the "Apache 2.0" and the "GNU Lesser General Public License 2.1 or later" license terms (dual license).

SPDX-License-Identifier: Apache-2.0 OR LGPL-2.1-or-later

[gabrielmancini]

Rebased on release-0.7 after #2436 was merged: the conflict was the CHANGELOG and the test list only, the code is the approved one.