P1: cap the path table at 150 records, and pick up the stack fixes

[A13xB0]

The solar P1 ran out of memory after 30 hours: ContainerAllocator failed to allocate memory (64 bytes), with 30 KB nominally free — fragmentation, not exhaustion. Three things were behind it, found from the node's own Heap Pool Stats over a 30-hour soak plus an audit of the stack's containers.

The path table was sized for flash, not RAM

The P1's QSPI branch set path_table_maxsize(500), copied from the RAK external-flash modules. But the path store's index lives in RAM: every record costs a FileStore index entry plus a _path_states entry, about 140 bytes together. Measured on the node, the table sat at 446–479 records — roughly 65 KB of a 140 KB pool, which accounted for the entire 104 KB baseline it booted with.

At 150 records the same node boots at 61% pool used with 46 KB free, against 73% and 22 KB before. The store is LRU, so the paths in use are the ones it keeps.

The neighbour scan ran every jobs() tick (ScotMesh/microReticulum a58ef77)

_scan_neighbor_stats() walked the whole stats map on every tick — four times a second — allocating a candidate vector and a hex string per neighbour per pass. In one hour of the soak: 7,082 scans and 13,166 skip traces. The per-neighbour rate limits bound the probe traffic, not the scan. Nothing it inspects can change faster than the 300-second suspicion and probe windows, so it is now gated at 15 seconds like every other periodic job in that loop.

cull_announce_table stopped on the wrong limit (same commit)

It triggers on _announce_table_maxsize but broke on _path_table_maxsize, so what it trimmed to depended on an unrelated table's size.

Measured

	before	after
Path table	446–479	150
Pool used	104,444 B (73%)	79,516–87,580 B (55–61%)
Free	~22 KB (15%)	~46 KB (32%)
Live objects	2,104 → 2,395 over 30 h	1,527

Still open

There is no age-based path expiry in this port: the PATHFINDER_E sweep is a comment, DestinationEntry._expires is computed and stored but never read, and the only TTL is a week checked lazily on access. The cap is a ceiling, not a cure — the underlying drift was +307 bytes/hour. _held_announces and _announce_rate_table also have no removal path, though both are dormant on this node.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SJZ75CWob1eMXeTa8Q2nLc