fix(docker): harden egress proxy forwarding

[lavaman131]

Summary

Hardens the Docker restricted-egress Squid sidecar so long installed-agent runs are less likely to wedge allowed provider CONNECT tunnels and so proxy diagnostics persist with the trial directory.

Fixes #15.

Changes

• Persist allowed_domains.txt, squid_access.log, and squid_cache.log under egress-proxy/logs/ via a bind mount instead of leaving them in container /tmp.
• Replace the TCP-open healthcheck with a Squid PID healthcheck to avoid per-second NONE_NONE/000 transaction-end-before-headers access-log spam.
• Configure explicit DNS resolvers, shorter negative DNS TTL, and longer/retried connect/forward timeouts for the Squid sidecar.
• Preserve the existing authenticated domain allowlist behavior.

Validation

• python3 -m py_compile src/pier/environments/agent_setup.py
• uv run ruff check src/pier/environments/agent_setup.py
• uv run ruff format --check src/pier/environments/agent_setup.py
• Built the generated Squid sidecar locally and verified a proxied request to https://openrouter.ai/api/v1/models returned HTTP 200 with TCP_TUNNEL/200 in squid_access.log.
• In a DeepSWE/Atomic workflow validation run using this patch, OpenRouter CONNECT traffic has remained TCP_TUNNEL/200 so far, with no repeated TCP_TUNNEL/503 ... HIER_NONE/- burst observed.

[lavaman131]

Hey @Ynng, any updates on this? I talked with James from your team and he mentioned you would be able to look at this. Thanks.