fix(codex): bound bridge app-server stalls

[yui-stingray]

Summary

• add bounded Codex bridge app-server connect/upgrade and JSON-RPC request timeouts
• apply request timeout cleanup to both stdio and direct WebSocket app-server clients
• stop after configurable consecutive real watch-once failures while keeping exit 2 as normal re-arm behavior
• clear failed watch spawn state and fail the bridge when an app-server request times out during event handling
• isolate Codex bridge Bats project roots under each test temp dir

Closes #195.

Behavior notes

New knobs:

• --connect-timeout-ms, AGMSG_CODEX_BRIDGE_CONNECT_TIMEOUT_MS, default 10000
• --request-timeout-ms, AGMSG_CODEX_BRIDGE_REQUEST_TIMEOUT_MS, default 30000
• --watch-failure-limit, AGMSG_CODEX_BRIDGE_WATCH_FAILURE_LIMIT, default 3

0 disables the corresponding timeout/limit.

A request timeout inside an app-server event handler now intentionally terminates the bridge instead of only logging and continuing. With the new timeout behavior, continuing after a timed-out process/spawn or turn/start can leave the bridge alive but unable to monitor correctly, for example with a non-null watchHandle and no actual watch process. Failing fast gives a clear error instead of a silent pseudo-monitor stall.

Validation

• node --check scripts/drivers/types/codex/codex-bridge.js
• git diff --check
• bats --print-output-on-failure tests/test_codex_bridge.bats -> 22/22
• bats --print-output-on-failure -f 'codex' tests/test_delivery.bats -> 10/10
• timeout 240s bats --print-output-on-failure tests/ -> reached 168/393 before the outer timeout; the changed Codex bridge section passed (33-54), Codex delivery checks passed (157-163), and the only observed failure before timeout was the existing unrelated delivery set monitor: existing settings with single-quoted hook commands stays valid JSON (#134) malformed JSON case.

Review notes

This was checked with separate read-only review passes for approach, test design, implementation diff, and final readiness. The remaining practical risk is live Codex app-server variance; the added tests use fake stdio/WebSocket app-servers to cover the protocol stall/failure paths without touching real db/, teams/, or run/ state.