feat(sight): add ECS security group guide to dashboard command

[chengshuyi]

Description

ECS users deploying AgentSight often cannot access the Dashboard remotely because port 7396 is not open in the instance security group. Currently there is no guidance — users see a timeout and have no idea what to do.

This PR adds ECS metadata probing to the agentsight dashboard command so that when running on an ECS instance, it automatically detects the instance and outputs a direct console link for configuring the security group. It also shows a hint at agentsight serve startup when binding to 0.0.0.0.

The implementation extracts shared ECS metadata primitives (metadata_agent(), read_plain(), METADATA_BASE) into a new ecs_metadata module, which genai::instance_id now reuses — eliminating ~30 lines of duplicate agent-builder and URL code.

Related Issue

no-issue: new UX feature for ECS deployment workflow

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional change)
• Performance improvement
• CI/CD or build changes

Scope

• sight (agentsight)

Key Changes

• src/agentsight/src/ecs_metadata.rs (new): shared ECS metadata client — metadata_agent(), read_plain(), probe_ecs_metadata() with mpsc::recv_timeout(2s) hard deadline, IMDSv2 token fetched once and reused across all fields
• src/agentsight/src/bin/cli/dashboard.rs: probe ECS metadata on dashboard command, display instance console URL with security group hint when on ECS; added --no-open and --skip-sg-guide flags
• src/agentsight/src/server/mod.rs: eprintln security group hint at serve startup when binding to 0.0.0.0 or ::
• src/agentsight/src/genai/instance_id.rs: reuse metadata_agent() and read_plain() from ecs_metadata, remove local agent builder and hardcoded URL (~30 lines deleted)
• src/agentsight/src/lib.rs: register ecs_metadata module (shared, not server-gated)
• src/agentsight/src/bin/agentsight.rs: update doc comments for dashboard subcommand

Checklist

• I have read the Contributing Guide
• cargo fmt --check pass
• cargo clippy --all-targets -- -D warnings pass (pre-existing warnings in unified.rs, not from this change)
• cargo test pass (5 ecs_metadata + 2 instance_id tests, all green)
• I have added tests that prove my fix is effective or that my feature works
• I have updated the documentation accordingly
• Lock files are up to date (Cargo.lock)

Testing

1. Non-ECS environment: agentsight dashboard --no-open completes within ~2s (probe_ecs_metadata returns None via recv_timeout), shows fallback message. Verified probe_returns_none_when_not_on_ecs unit test.

2. ECS environment (cn-shenzhen, gn7i instance): deployed release binary, ran agentsight serve --host 0.0.0.0 (shows security group hint at startup), then agentsight dashboard --no-open — output includes instance console URL with region.

3. Unit tests: 5 tests in ecs_metadata (public_ip preference, instance_url format, probe timeout) + 2 tests in instance_id (CachedUid caching, MAX_RETRIES cap) — all pass.

[jfeng18]

Reviewed in depth (independent sweep + adversarial pass, local build/test on the CI toolchain 1.89.0, and a real-ECS metadata E2E). The core is solid — approving. A few non-blocking notes below.

Verified:

• Builds clean (let-chains fine on 1.89); 25 new unit tests pass (14 ecs_metadata + 11 dashboard); off-ECS probe returns None in ~1.6s.
• The instance_id refactor is behavior-equivalent — old and new both use plain IMDSv1 (no token, no XFF), same URL/timeout/trim/empty-handling and the same CachedUid/OnceLock semantics, so the SLS owner-account-id uid path is unchanged. The token/XFF machinery lives only in the new probe path, which instance_id never calls.
• Module registration (ecs_metadata as cross-cutting) and the arch-boundary update are correct; the module has no outward crate:: edges.

Non-blocking notes:

1. [low] On ECS, dashboard prints the public IP with the live token. resolve_display_url prefers the ECS public IP (EIP → public-ipv4), and the auth line prints URL with token: {display_url}/?token={token}. Verified on a real instance: the NIC carries a private IP (172.x) while eipv4 returns the routable public IP over IMDSv1, so the printed URL becomes http://<public-eip>:7396/?token=<token> — whereas the previous code paired the token with the private NIC IP. It's the operator's own terminal, but it's now a directly-usable public credential if that output is pasted into a log/screenshot/issue, and the same command's SG guide steers the user toward opening 7396 publicly. Consider masking the token on the public-IP line (or printing the token separately).

2. [nit] Unexplained X-Forwarded-For: China header in the IMDSv2 path (ecs_metadata.rs, read_with_token). On a non-hardened instance the token PUT /latest/api/token returns 400 → ureq treats it as Err → token = None → the v2 path (and this header) is skipped entirely, IMDSv1 handles everything. So it's dormant here; I couldn't exercise the hardened-IMDS path. Worth a comment on why the header is there, or dropping it if it isn't needed.

3. [nit] instance_id fetch logs dropped the {e} cause (owner-account-id / instance-id warn/debug lines) in the refactor — mild loss of diagnostic detail when the SLS uid comes back empty. Behavior otherwise identical.

4. [nit, informational] clippy note in the PR body is imprecise. cargo clippy --all-targets -- -D warnings @1.89.0 does fail, but with 9 pre-existing uninlined_format_args across five files (probes/codex_offsets.rs, probes/elf_buildid.rs, health/checker.rs, server/token_savings.rs ×5, unified.rs) — not just unified.rs, and none in this PR's changed files. The new code is clippy-clean and the CI Test agentsight gate is green. Could be a nice separate cleanup.

Nice UX improvement for the ECS deployment path. LGTM.