router: model the channel as a continuous carrier (never drop mid-transmission)

[M0LTE]

Problem

The per-link audio queue (internal/router/router.go linkQueue) was a fixed ~3 s buffered channel that dropped the oldest block on overflow. A TNC bursts a whole keyup's worth of audio with no real-time pacing, so any transmission longer than ~3 s overran the buffer and net-sim dropped blocks mid-transmission.

Dropping mid-transmission gaps the receiver's audio, which collapses its DCD / carrier sense — so the far end thinks the channel is idle and keys up on top of the in-progress transmission. That overlap is a collision (the mixer's MixCollision → silence), the frame is lost, and the loss kicks off a go-back-N retransmit storm that overruns the buffer further. A real continuous-carrier channel has no such limit.

Evidence

Measured on a real LinBPQ ↔ packet-radio-node link over the sim, driving a chat help-text dump (back-to-back long I-frames):

• the receiver decoded only ~26 % of the sender's I-frames;
• 36–63 % of the receiver's own transmissions collided with the sender's (composite-WAV L/R energy-overlap analysis);
• the collisions began a median ~3.3 s into each transmission — exactly when the 3 s buffer saturated and started dropping;
• reproduced identically under both the samoyed and direwolf backends, so it's the shared layer (net-sim), not the modem.

Fix

Model the channel as a continuous carrier: linkQueue becomes a non-dropping FIFO that grows as needed. The real-time rxFeeder still meters it out at the channel sample rate, so channel timing is unchanged — but the receiver now hears a gap-free carrier for the full transmission and its carrier sense holds. Memory is bounded in practice (a keyup is finite; the backing array is released once drained); only a pathological runaway past a 60 s safety cap ever drops, logged as the anomaly it is.

Result

Same link, after the fix:

	before (3 s drop)	after
audio queue overflow	2,876	0
receiver-TX-over-sender collisions	36–63 %	0 %
receiver I-frame delivery	26 %	~90 %
help text delivered	38 bytes	full (1,050)

The go-back-N storm disappears (the sender stops retransmitting). Existing TestCompositeRealTimePacing / TestCompositeConcurrentFeed still pass (metering unchanged); adds TestLinkQueueDeliversLongBurstGapFree — a ~10 s burst (well past the old 3 s cap) must be delivered FIFO and gap-free.

🤖 Generated with Claude Code