Remember .btf peers that answered, and dial them first on start

[Bitflash-sh]

Attacks the seven-minute wait before a node finds its first peer, which we measured repeatedly today.

Why it takes so long

Discovery is entirely dependent on the Nostr relays, and the P2P protocol carries no peer information at all — version, block, tx, inv, getblocks, getdata and nothing else. Two nodes can be connected and syncing without ever telling each other that anyone else exists.

So every start repeats the same sequence: reach the relays, pull down ~270 descriptors, dial them blindly, and wait while roughly 96% fail because the peers behind them are long gone.

A node that has been running for hours knows exactly which addresses answer. It throws all of it away on exit.

What this does

Every peer that completes a tunnel is written to btfpeers.json along with its meeting node and encryption key — everything ConnectNodeBtfResolved needs to redial it without touching a relay. On start, a thread tries those first, stopping once eight are connected.

Entries expire after a week and the file holds fifty at most, so a node that has been off for a long time still falls back to discovery rather than hammering addresses that have certainly moved.

Scope

Cache only. A first-ever run is unchanged, and a missing or unparseable file just falls through to the relays. Nothing here is consensus.

What it does not fix

The dependency itself. If the relays are unreachable — and today's logs show relay.damus.io refusing handshakes and relay.nostr.band failing outright — a node with an empty cache still cannot find anyone.

The real answer is peer exchange over the existing P2P link. A .btf address reveals no IP, being a hash of a public key, so gossiping them is safer than Bitcoin's addr messages ever were. After first contact the network would sustain itself, with the relays as bootstrap rather than lifeline. That is a bigger change and wants its own PR.

Not built yet — needs make windows.

[Bitflash-sh]

Runtime test done — this was the piece missing before merge.

Two rounds on a clean isolated datadir, same binary, same machine, network in its
current (healthy) state. Timing is wall-clock from process start to the first
Added time data line, which only appears once a peer completes the version
handshake — i.e. the first moment a real peer exists.

Round	Cache	First peer	First block
1	empty datadir, no btfpeers.json	36.3s	38.3s
2	same datadir, cache present	2.1s	4.1s

~17x faster to first peer. Round 1 wrote a cache with 2 entries; round 2 dialed
them directly and skipped the relays entirely.

Cache file after round 1, well-formed:

[
  {
    "btf": "varrjzrzlzlc63h2l7hmssdinrjr75t3jz2gx4vurpjlcmcl4564gra.btf",
    "enc": "5cd605d288e654fd15983f00e11b53e0c3e41761ab758856b5ee0b87d03e983b",
    "meeting": "92.246.128.180:8434",
    "seen": 1785272843
  },
  ...
]

One thing worth knowing, not a blocker: both entries pin the same meeting relay.
That relay was down for part of today, and a cached entry whose meeting node has died
is undialable until the node falls back to Nostr re-resolution. The cache is a
fast path, not a replacement for discovery — which is how it is written, so this is
an observation rather than a defect.

Unrelated finding while setting the test up: net.cpp says nListenPort is
"tunable via /port", but nothing in main_gui.cpp parses /port or -port. The
option does not exist, so two nodes cannot share a machine. Either the comment or
the option should go; separate change.

[Bitflash-sh]

For the record, since this PR's description does not match what it merged.

It was reviewed as one file (net.cpp, +169/-0, the peer cache). I later pushed peer exchange and a /port option onto the same branch while the PR was open, which widened it to six files without updating the description. The merge carried all of it.

The peer cache itself is what the numbers above measured and is unaffected. The peer exchange had not been validated — its two-node test never completed. #27 removes it from main; it will come back as its own PR once it passes.

Release v1.2.2 was built from 3caee27, before this merge, so nothing published is affected.