fix(xtask): set GridNetwork gatewayRefs[].localSiteName in E2E fixtures

[jordigilh]

Summary

xtask env verify-operator-reconcile failed deterministically with:

error: scoring order check failed: op-e2e-api-fallback (pos 0) must appear after site-a (pos 1); expected local > api_provider

Root-caused live on a fresh local kind cluster (see comment thread on #60 for the full investigation). Fixes #60.

Root cause

network_fixture_json (the shared GridNetwork E2E fixture builder) never set gatewayRefs[].localSiteName. render_routing_overlay's local_site therefore fell back to the network name (op-e2e-net), which matches no candidate's site field. Every candidate tied at LocalityTier::Unknown, so GeographyFirst ordering (the production default) fell through past locality to score.

Under the noMetrics scoring strategy (#38, now the default), both the local and API-provider candidates correctly score identically — "both providers must have equal dynamic scores under noMetrics" is asserted intentionally in routing_overlay.rs's own unit tests. With score tied too, the sort fell to the final deterministic tiebreak: (site, name, cluster) ascending — and "op-e2e-api-fallback" < "site-a" alphabetically.

This was previously masked by the old combined-weight scorer, which gave local backend_kind a static score advantage over api_provider even with no live metrics — so the tie-through-to-score path happened to still land on the right answer by coincidence, not because locality was genuinely exercised. noMetrics correctly removed that bias (per its own design/tests), which is what exposed the dormant fixture gap.

Confirmed pre-existing and unrelated to any of my recent grid#58/grid#59 changes: reproduces identically with those changes stashed, on main.

Fix

Thread the existing routing_cluster parameter ("site-a" by default) through to network_fixture_json as gatewayRefs[].localSiteName, so the healthy/degraded/metrics fixtures — which all use routingClusterRef = routing_cluster — genuinely resolve to LocalityTier::SameSite, and the API-fallback/invalid fixtures (no matching routingClusterRef) correctly resolve to Unknown.

apply_multi_provider_fixtures has no single "local" site across multiple provider sites, so it passes the network name explicitly, preserving its exact prior fallback behavior (that check validates per-site candidate presence, not locality ordering).

Verification

• cargo test -p xtask: 459 passed, 0 failed
• cargo clippy --workspace --all-targets -- -D warnings: clean
• cargo +nightly-2026-03-28 fmt --all -- --check: clean
• cargo machete: no unused deps
• Live local kind cluster run:

  cargo run -p xtask -- env up --config tests/env/config.toml
  cargo run -p xtask -- env verify-operator-reconcile --config tests/env/config.toml
  ...
    [OK] scoring order: site-a (pos 0) before op-e2e-api-fallback (pos 2)
  verify-operator-reconcile: PASS
  

Test plan

• verify-operator-reconcile passes locally end-to-end
• Existing xtask unit tests for verify_scoring_order untouched and still pass (that function itself was never the bug)
• apply_multi_provider_fixtures behavior intentionally unchanged (documented in code comment)

[praxis-bot]

PR Review

Summary: Fixes E2E fixture wiring — network_fixture_json now receives an explicit localSiteName parameter so GeographyFirst locality-tier ordering is genuinely exercised instead of silently falling through to alphabetical tiebreak.

Overall: Clean, well-scoped fix with clear root-cause analysis. The two new tests pin the fixture contract at the unit level. Two medium findings below.

Severity	Count
Critical	0
Large	0
Medium	2

Findings without inline placement

[Medium] The PR's central thesis is that fixture wiring regressions should be caught by cargo test -p xtask in milliseconds rather than only by a live-cluster E2E run. The two new tests pin the single-provider path (apply_test_fixtures_for_cluster), but apply_multi_provider_fixtures (line 5540) has no corresponding contract test verifying that its localSiteName equals TEST_NETWORK (intentionally matching no candidate). If someone later changed that call to pass a real site name, locality ordering would silently re-engage in multi-provider validations with no unit-level signal. Add a test analogous to network_fixture_json_sets_local_site_name_from_argument that asserts apply_multi_provider_fixtures's network fixture uses TEST_NETWORK as localSiteName.

[jordigilh]

Addressed both findings in a290858:

• Extracted apply_multi_provider_fixtures's network fixture construction into a new multi_provider_network_fixture_json helper so it's directly unit-testable without a live cluster, and added apply_multi_provider_fixtures_network_uses_test_network_as_local_site_name pinning that it stays TEST_NETWORK.
• Removed the test-body comment flagged inline.