fix(channel): fail workspace calls fast when the workspace host stops answering

[iamnbutler]

Fixes #372.

A frozen or unreachable workspace host kept its socket attached to the hosted channel, so the instructions prompt section waited on AGENTS.md reads and blocked model-only runs until the host returned.

• Liveness: hosts send {"ping":true} on the workspace socket every 5s. The object answers {"pong":true} through setWebSocketAutoResponse, so hibernation is unaffected.
• Link fails fast: Link treats a socket whose getWebSocketAutoResponseTimestamp is older than 15s (three intervals) as not responding. New calls fail immediately with "The workspace is not responding". Calls already in flight fail on a watchdog that runs only while calls are pending, and each one gets a cancel for when the host resumes. Desktop and browser actions are refused without being dispatched.
• Instructions degrade: instructions now treats only a missing file or directory as absent. Any other read failure throws, so pi keeps the instructions it last showed and reports the error, where before it silently dropped them.
• Host reconnects: once a service has answered a ping, a host that then misses three pongs closes the socket and reconnects. It counts unanswered pings rather than elapsed time, so a host waking from sleep keeps its live socket.

Compatibility

• An older host never pings, so its socket has no auto-response timestamp and is treated as answering, the same as today. I checked this: a host built from main sat idle for 30s against the new Worker and its tools still worked.
• A new host against an older Worker: the ping is a JSON object that the old Link.receive ignores. No pong comes back, so the host never starts watching for pongs. This was not run, only read from the code.
• ace-channel-study needs a redeploy from main to get liveness.

Smoke test (wrangler dev plus a source host on its own port and data dir, openai/gpt-4.1-mini, with an AGENTS.md instruction the model has to follow):

• Host live: the run answered in 1.7s and followed AGENTS.md.
• Ask sent right after SIGSTOP of the host: answered in 16s. The pending read failed at the stale mark, and the kept instructions still applied.
• Ask sent 20s into the freeze: answered in 1.5s. A shell tool reported "The workspace is not responding" within 2s.
• After SIGCONT: tools worked again on the same socket, with no reconnect.
• After SIGSTOP of workerd: the host logged workspace.drop (stale) about 16s later and reconnected after SIGCONT. Tools worked afterward.

Built in Ace

No — built with Claude Code outside Ace by an overnight issue watcher, because this Mac's Ace Canary is a participant in the running two-person study (#218); dogfooding gap tracked in #5.