Address review feedback #2 on PR #442: detect silently-broken pooled
connections proactively instead of waiting for the next real request
to discover them via code=1006.

Track lastActivityAt — updated by ANY pong or data message from the
peer (both prove the connection is alive). On each ping tick, if
now - lastActivityAt > livenessTimeoutMs, markDead the WS. Default
threshold is 2.5x pingIntervalMs (~62.5s with default 25s ping):
tolerates one missed pong (network blip) but evicts before a third
would tick, at which point the connection is almost certainly dead
and reusing it would cost a real-request cache miss.

Counter-based "missed pings" alternative was rejected: it would
false-positive on healthy streaming sessions where the server sends
data but no separate pong, dragging working connections offline.

E2E verified end-to-end from device a (Mac mini, 192.168.10.2) →
proxy 192.168.10.6:8080 → chatgpt.com via a 10-turn pinned-session
load script with a 70s idle gap between turns 5 and 6. Turn 6 stayed
on the same pooled WS as turns 1-5 and hit 99.6% cache (matching
pre-gap turn 5), with zero liveness-timeout markDead events — the
keepalive pings carried the connection across the LB idle window
unharmed.

WsLike interface gains `on("pong", listener)`. Real ws.WebSocket
already emits "pong" per RFC 6455 §5.5.3.

Tests added (6 new):
- liveness > marks dead when peer stays silent past timeout
- liveness > pong resets the clock
- liveness > data message resets the clock
- liveness > livenessTimeoutMs=0 disables
- liveness > default multiple keeps healthy WS alive across many cycles
- ping > skips while busy (active stream keeps LB alive)

…442)

* perf(proxy): WebSocket keepalive ping prevents middlebox idle drops

Pooled WSes were silently RST'd by upstream LB / NAT / firewall idle
timeouts after ~30-60s with no traffic, surfacing as code=1006 on the
next turn. Each drop forced a fresh WebSocket against a different
backend instance, losing the prompt cache prefix and dragging hit rates
back to 5-9%.

Send a ws-level ping frame every 25s (configurable, 0 disables) so the
middlebox NAT/connection-tracker keeps the mapping alive. Real-traffic
verification: single pooled WS sustained 22+ consecutive turns at
88-94% hit, vs the prior pattern of single-use WS dying after one
request.

* perf(proxy): skip keepalive ping while WS is busy + harden tests

Address review feedback on PR #442:

- sendKeepalivePing returns early when this.busy is true. The active
  stream's data frames already keep the upstream LB / NAT idle timers
  fresh, so emitting a ping during streaming would be redundant
  bandwidth on chatty sessions.
- Strengthen the error-swallow test to assert pingCount=1 after the
  swallowed throw — a bare not.toThrow() would have missed a
  regression that crashes the interval loop after one bad ping.
- Add a regression test for the busy-skip behavior.
- Inline comment on WsLike.ping() to flag the narrowed signature
  versus real ws.WebSocket.ping(data?, mask?, callback?).

* perf(proxy): add WS liveness check (pong/message tracking)

Address review feedback #2 on PR #442: detect silently-broken pooled
connections proactively instead of waiting for the next real request
to discover them via code=1006.

Track lastActivityAt — updated by ANY pong or data message from the
peer (both prove the connection is alive). On each ping tick, if
now - lastActivityAt > livenessTimeoutMs, markDead the WS. Default
threshold is 2.5x pingIntervalMs (~62.5s with default 25s ping):
tolerates one missed pong (network blip) but evicts before a third
would tick, at which point the connection is almost certainly dead
and reusing it would cost a real-request cache miss.

Counter-based "missed pings" alternative was rejected: it would
false-positive on healthy streaming sessions where the server sends
data but no separate pong, dragging working connections offline.

E2E verified end-to-end from device a (Mac mini, 192.168.10.2) →
proxy 192.168.10.6:8080 → chatgpt.com via a 10-turn pinned-session
load script with a 70s idle gap between turns 5 and 6. Turn 6 stayed
on the same pooled WS as turns 1-5 and hit 99.6% cache (matching
pre-gap turn 5), with zero liveness-timeout markDead events — the
keepalive pings carried the connection across the LB idle window
unharmed.

WsLike interface gains `on("pong", listener)`. Real ws.WebSocket
already emits "pong" per RFC 6455 §5.5.3.

Tests added (6 new):
- liveness > marks dead when peer stays silent past timeout
- liveness > pong resets the clock
- liveness > data message resets the clock
- liveness > livenessTimeoutMs=0 disables
- liveness > default multiple keeps healthy WS alive across many cycles
- ping > skips while busy (active stream keeps LB alive)

---------

Co-authored-by: icebear0828 <icebear0828@users.noreply.github.com>

…442)

* perf(proxy): WebSocket keepalive ping prevents middlebox idle drops

Pooled WSes were silently RST'd by upstream LB / NAT / firewall idle
timeouts after ~30-60s with no traffic, surfacing as code=1006 on the
next turn. Each drop forced a fresh WebSocket against a different
backend instance, losing the prompt cache prefix and dragging hit rates
back to 5-9%.

Send a ws-level ping frame every 25s (configurable, 0 disables) so the
middlebox NAT/connection-tracker keeps the mapping alive. Real-traffic
verification: single pooled WS sustained 22+ consecutive turns at
88-94% hit, vs the prior pattern of single-use WS dying after one
request.

* perf(proxy): skip keepalive ping while WS is busy + harden tests

Address review feedback on PR #442:

- sendKeepalivePing returns early when this.busy is true. The active
  stream's data frames already keep the upstream LB / NAT idle timers
  fresh, so emitting a ping during streaming would be redundant
  bandwidth on chatty sessions.
- Strengthen the error-swallow test to assert pingCount=1 after the
  swallowed throw — a bare not.toThrow() would have missed a
  regression that crashes the interval loop after one bad ping.
- Add a regression test for the busy-skip behavior.
- Inline comment on WsLike.ping() to flag the narrowed signature
  versus real ws.WebSocket.ping(data?, mask?, callback?).

* perf(proxy): add WS liveness check (pong/message tracking)

Address review feedback #2 on PR #442: detect silently-broken pooled
connections proactively instead of waiting for the next real request
to discover them via code=1006.

Track lastActivityAt — updated by ANY pong or data message from the
peer (both prove the connection is alive). On each ping tick, if
now - lastActivityAt > livenessTimeoutMs, markDead the WS. Default
threshold is 2.5x pingIntervalMs (~62.5s with default 25s ping):
tolerates one missed pong (network blip) but evicts before a third
would tick, at which point the connection is almost certainly dead
and reusing it would cost a real-request cache miss.

Counter-based "missed pings" alternative was rejected: it would
false-positive on healthy streaming sessions where the server sends
data but no separate pong, dragging working connections offline.

E2E verified end-to-end from device a (Mac mini, 192.168.10.2) →
proxy 192.168.10.6:8080 → chatgpt.com via a 10-turn pinned-session
load script with a 70s idle gap between turns 5 and 6. Turn 6 stayed
on the same pooled WS as turns 1-5 and hit 99.6% cache (matching
pre-gap turn 5), with zero liveness-timeout markDead events — the
keepalive pings carried the connection across the LB idle window
unharmed.

WsLike interface gains `on("pong", listener)`. Real ws.WebSocket
already emits "pong" per RFC 6455 §5.5.3.

Tests added (6 new):
- liveness > marks dead when peer stays silent past timeout
- liveness > pong resets the clock
- liveness > data message resets the clock
- liveness > livenessTimeoutMs=0 disables
- liveness > default multiple keeps healthy WS alive across many cycles
- ping > skips while busy (active stream keeps LB alive)

---------

Co-authored-by: icebear0828 <icebear0828@users.noreply.github.com>