fix(dispatcher): stale keep-alive burst on local egress

[VIPKaiser]

Problem

On macOS Docker Desktop, every *.internal / *.local hostname the
container reaches through host.docker.internal (every local LLM
provider wired through the proxy: oMLX, lightning-mlx, mtplx, Breeze TTS,
Parakeet STT, Kokoro TTS) can hit a 30-second ECONNREFUSED burst when
Docker Desktop's NAT silently drops idle keep-alive sockets inside the
direct round-robin pool. Triggers: a request lands on a stale pooled
socket, retries, gets a fresh one, hits the same dead path, then
cooldown/lockout kicks in.

Three concrete amplifiers:

• proxyDispatcher.ts:73-76 sets autoSelectFamilyAttemptTimeout: 1000,
  but the IPv6 form of host.docker.internal
  (fdc4:f303:9324::254) is ENETUNREACH inside the container (the
  IPv4 form 192.168.65.254 is the working address). Every healthy
  request waits 1s on the dead family before getting ECONNREFUSED on v4.
• getDefaultDispatcher() caches the round-robin pool in
  globalThis and never reaps on PROXY_UNREACHABLE — a single stale
  socket poisons the pool for all subsequent requests until process
  restart.
• The default keepAliveMaxTimeout (4s) sits squarely inside the NAT's
  silent-drop window for local egress.

Fix

Three coordinated changes, all backward-compatible:

1. Hostname-aware dispatcher options —
getDispatcherOptions(hostname?) shortens
keepAliveMaxTimeout to 1000ms and autoSelectFamilyAttemptTimeout to
200ms when the hostname matches *.internal / *.local. New
exported isLocalEgressHostname() helper.

2. Parallel cache for local-egress dispatchers —
proxyDispatcherCache.ts adds LOCAL_DEFAULT_DISPATCHER_KEY /
LOCAL_RETRY_DISPATCHER_KEY with matching accessors.
getDefaultDispatcher(hostname?) / getRetryDispatcher(hostname?)
route to them when the hostname is local-egress. The cloud-upstream
pool keeps its wider keep-alive.

3. Cache invalidation on PROXY_UNREACHABLE for local egress —
proxyFetch.ts calls clearDispatcherCache() before _nativeFallback
when the error is PROXY_UNREACHABLE and the target hostname is
local-egress, forcing the next request to rebuild the pool with fresh
sockets.

Files

• open-sse/utils/proxyDispatcher.ts (+63 / −8)
• open-sse/utils/proxyDispatcherCache.ts (+29)
• open-sse/utils/proxyFetch.ts (+22 / −1)

Backward compatibility

Every existing caller of getDefaultDispatcher() /
getRetryDispatcher() / __getDefaultDispatcherOptionsForTest() keeps
working — the hostname? parameter is optional and defaults to the
existing cloud behaviour. No config / env / DB changes.

Verification

Scope	Pass	Fail	Skipped
Targeted proxy/dispatcher/executor suite (39 files)	742	0	1
proxyfetch-direct-response-start-timeout-10214	✔
direct-dispatcher-pipelining-4580	✔
proxy-egress-isolation-bdd	✔
proxy-fetch	✔
proxy-concurrency-keepalive-regression	✔
web-cookie-validation-proxy-7058	✔
socks-connect-timeout-e2e	✔

Pre-existing failures on the cycle (uc-video persona-timeout,
models-catalog Jina/GLM-5.2, tunnel-routes sanitize, zai-stream
error code) reproduce on main unchanged — none introduced by this PR.

Notes

Cherry-picked onto release/v3.8.51 from a separate working branch.
One conflict in proxyFetch.ts resolved by taking upstream's rename of
the inline tlsProfileForProvider return type to a named
TlsProfileResult alias (no semantic change).

The companion AgentBridge MITM autostart and call_logs.id UNIQUE race
fixes are intentionally not included here — they're independent fixes
and will land separately.

Related work (not duplicates)

These recently merged neighbours touch nearby layers but do not
supersede this PR — included here to head off the "is this duplicate?"
review question:

• fix(db): rotate proxy pools on the chat path like the registry does #14044 fix(db): rotate proxy pools on the chat path like the registry does — DB-layer resolveProxyForConnection cache-freeze fix for the chat-path proxy resolver. Does not touch undici dispatcher cache, keep-alive, or local-egress hostnames.
• fix(sse): stop direct fetch retry reusing pooled flat response-start budget (#13703) #14047 fix(sse): stop direct fetch retry reusing pooled flat response-start budget — directHeadersTimeoutMs per-attempt budget for the response-start timeout. Different layer (response-start timeout, not idle socket lifetime).
• fix(sse): pin DNS on the three public-only image fetch sites (#13883) #14032 fix(sse): pin DNS on the three public-only image fetch sites — pinDns: true for outbound image fetches. Different layer (DNS, not undici dispatcher).
• fix(proxy): skip bare TCP health probe for SOCKS5 data plane #13571 fix(proxy): skip bare TCP health probe for SOCKS5 data plane — SOCKS5-only health probe change. No dispatcher cache work.
• feat(proxy): support multiple local core endpoints, one per line #13923 feat(proxy): support multiple local core endpoints, one per line — proxy-tunnel loopback endpoint registration. Different problem surface (proxy pool rotation in the DB layer).
• feat(sse): per-egress pacing + fleet-wide backoff for opencode rotation (opt-in) #14180 / feat(sse): per-egress pacing + fleet-wide backoff for opencode rotation (opt-in) #14290 feat(sse): per-egress pacing + fleet-wide backoff for opencode rotation (opt-in) — opt-in via OPENCODE_EGRESS_THROTTLE_ENABLED=1. OpenCode-only IP-bucketed burst throttle, behind a flag. Different layer.

None of the above touches the three properties this PR changes: (1) hostname-keyed getDispatcherOptions, (2) parallel LOCAL_*_DISPATCHER_KEY cache slots, (3) clearDispatcherCache() invocation on PROXY_UNREACHABLE for local-egress.