v3.3.3 — Memory-safe trap/inform sends, flood-calibrated pool, alarm-free on ESP8266

[syntax1269]

Fork of 0neblock/Arduino_SNMP (last upstream release 2.1.0, ~3 years ago), carried forward as a maintained, hardware-validated line. This upload is v3.3.3.

v3.3.3 — memory-safety release

The long-standing "slot-12 DOUBLE RELEASE" pool alarm (fired once per campaign since v3.1.23) is root-caused and eliminated, together with a latent cross-tick use-after-reuse in the same path.

• Root cause: SNMPTrap/SNMPInform objects are persistent sketch globals whose pool-backed members (packet tree, varbind containers) outlived the send. The agent's per-tick pool reset flag-freed those slots; once response traffic recycled them, the next trap's teardown struck re-armed slots — mostly a benign stale delete (the alarm), in the worst case a destructor on a live object.
• Fix: traps and informs are now fully stateless per send — a fresh tree is built on every call and all pool state is released on every exit path. Inform retries rebuild and re-apply the stored request ID (previously a retry could serialise an already-freed tree — possible garbage on the wire).
• Pool right-sizing: the transient-burst term is reshaped with three anchors — 768 B packet → pool 56 (flood-measured peak 51 + margin 5), 1024 B → pool 72 (campaign-proven, unchanged), 1400 B → pool 78 (generic, unchanged).
• Verified: host suite 175 assertions / 17 cases (default) + 176/17 (TINY), both ALL GREEN with a new colour-coded per-case summary. ESP8266 hardware: cold-boot startup trap alarm-free for the first time in the campaign; 1-minute 200 ms flood soak (GET/GETNEXT/GETBULK/SET + concurrent traps): 176 ops, 0 failures, 0 UDP losses, 0 pool alarms, 0 reboots, heap flat, pool peak 51/56.

What this fork adds over upstream 2.1.0

• Memory-safety hardening: pool double-release detection (alarms on true double-destroy), stale-after-bulk-reset deletes stay silent, stateless trap/inform sends — no pool-backed state survives a transmit.
• Derived, not magic: varbind cap and pool size are computed from the packet budget ((packet − 40) / 160 per-varbind worst case) and the sketch's handler count; a compile-time static_assert rejects mismatched overrides. Boot-time arena lock-in pre-allocates the pool before setup()/WiFi — zero heap-fragmentation exposure on long-running nodes (ESP8266 aborts on a failed new; this removes the time bomb).
• Loud failure modes (no silent drops): GetBulk above the varbind cap answers tooBig instead of silently truncating; a response that cannot fit the packet budget returns an RFC 3416 tooBig error PDU instead of burning the manager's timeout.
• Proven on an ESP8266 (1 MB, ~80 KB RAM): 30+ minute soaks and 200 ms flood tests with zero crashes, zero pool alarms, flat heap, ~1% fragmentation.
• CI-validated: arduino-lint, host Catch2 suite, and an ESP8266 + ESP32 example compile matrix (.github/workflows/test.yml).
• Docs/examples refreshed: README current, all example and demo sketches compile out-of-the-box on both targets.