fix: use small buffers to make tests more efficient

[dannykopping]

Warning: vibe-coded!

Dependency of coder/coder#21251

I asked Opus 4.5 to justify its changes:

1. IdealBatchSize: 128 → 4

What it controls: Number of packets handled per read/write batch operation.

Memory impact: This affects container pool allocations - each container holds BatchSize() element pointers. With 128, you get slices of 128 pointers; with 4, just 4 pointers.

Safety: ✅ Safe. If more than 4 packets need to be processed, wireguard simply does multiple batch operations. This only affects throughput, not correctness. Tests don't need high throughput.

---

2. QueueOutboundSize/InboundSize/HandshakeSize: 1024 → 64

What they control: Channel buffer sizes for packet queues.

Safety: ✅ Safe. If queues fill up, backpressure applies (blocks until space available). Tests don't generate sustained high packet rates. 64 is still a generous buffer for test scenarios.

---

3. MaxSegmentSize: 65535 → 2048

What it controls: Maximum UDP datagram size. This is the size of [MaxMessageSize]byte arrays allocated for packet buffers.

Memory impact: This is the dominant factor. Each message buffer is MaxSegmentSize bytes. With 65535 bytes each, these quickly consume gigabytes across many connections.

Safety: ✅ Safe - and here's why:

Platform	MaxSegmentSize
Default (Linux)	65535
Windows	2016 (2048-32)
iOS	1700
Android	2200
testsmallbuffers	2048

Critical insight: Windows and iOS run in production with values ≤2048. Our test value of 2048 is:

• Equal to Windows production value
• Larger than iOS production value (1700)
• Larger than standard MTU (1500) + WireGuard overhead (~80 bytes)

Since iOS devices successfully run WireGuard in production with MaxSegmentSize=1700, our test value of 2048 is definitively safe.

---

4. PreallocatedBuffersPerPool: 0 (unchanged)

Already 0 in default config, meaning sync.Pool handles allocation dynamically. No change needed.

---

Memory Calculation

Per-device with default settings:

• Message buffers: ~65KB each, pooled
• Container slices: 128 pointers × multiple pools

Per-device with testsmallbuffers:

• Message buffers: ~2KB each, pooled
• Container slices: 4 pointers × multiple pools

Result: ~32x reduction in per-buffer memory, compounding across hundreds of test connections.

---

Verdict

All changes are justified and safe:

1. IdealBatchSize=4: Affects only throughput, not correctness
2. Queue sizes=64: Provides backpressure, doesn't fail
3. MaxSegmentSize=2048: Production-proven on Windows and iOS - this is the key insight that validates the choice