test(e2e): cover multi-node p2p across sandbox churn

[WindSongShaoGe]

What

• Add a dedicated multi-node P2P compose E2E suite.
• Start one Firecracker sandbox on each AgentENV node.
• Verify sandbox veth creation and cleanup are observed.
• Exercise P2P layer transfers in both directions before and after sandbox deletion.
• Disable the origin path to prove that transfers use P2P rather than HTTP fallback.
• Verify HTTP 206 responses, content size, and SHA256 integrity.
• Add a dedicated Make target and CI workflow step.
• Make the E2E template build timeout configurable.

Validation

• Multi-node P2P E2E: 33 assertions passed
• E2E summary: 1 passed, 0 failed
• Shell syntax checks passed
• make fmt passed
• git diff --check passed

Refs #138

[github-actions]

🔍 OpenCodeReview found 3 issue(s) in this PR.

• ✅ Successfully posted inline: 3 comment(s)

[github-actions]

This new regression step has no execution timeout, while the target includes Docker operations and node-local API calls whose helpers do not consistently set network/process deadlines. A stuck docker compose exec or sandbox API request can therefore occupy the runner until GitHub's much larger default job limit. Add a step-level timeout-minutes (or a bounded job-level timeout) appropriate for the two E2E runs.

Suggestion:

Suggested change

	- name: P2P multi-node regression test
	if: ${{ matrix.e2e_name == 'docker-compose' }}
	run: SKIP_BUILD=1 make test-e2e-p2p PROFILE=debug
	- name: P2P multi-node regression test
	if: ${{ matrix.e2e_name == 'docker-compose' }}
	timeout-minutes: 20
	run: SKIP_BUILD=1 make test-e2e-p2p PROFILE=debug

[github-actions]

The new E2E target is not included in the existing test-e2e-all dependency list, so callers expecting that aggregate target to run every E2E suite will silently miss this regression. Add test-e2e-p2p to test-e2e-all (unless the aggregate target is intentionally no longer exhaustive).

[github-actions]

This creation-side check uses a fixed 5-second delay, unlike the deletion-side check below which polls for up to 20 seconds. On a loaded runner, sandbox state can become running before the host-side interface observation settles, causing an otherwise healthy P2P run to fail here. Poll until the veth set differs from the baseline (with a bounded timeout), then perform the assertion; the subsequent P2P fetch loop already follows this asynchronous polling pattern.