fix(scenarios): provision rpc fleet via provision-node, not provision-snd

[bdchatham]

Coral-xreview blocker on the WS-C nightly migration (platform #1165): the externally-fetched load-test/release-test scenarios still provisioned the RPC fleet with provision-snd --role=rpc on a kind: SeiNode template — but provision-snd creates a SeiNetwork (strict-unmarshal), so the RPC step render-fails the first time the nightly cron fetches these scenarios → load + release dead on arrival.

Change

Migrate the RPC provision step in both scenarios provision-snd → provision-node (standalone SeiNode followers), mirroring the in-repo chaos scenarios already migrated in #1165:

• --replicas (load-test=2 — seiload drives all N via RPC_EVM_RPC_LIST; release-test=1 — mocha hits a single RPC_TM_RPC/EVM/REST)
• --network=$SEI_CHAIN_ID — genesis peer auto-wire (sei.io/seinetwork=<chain>)
• --running-timeout (was --ready-timeout; SeiNode has no Ready phase)

The validator steps intentionally stay provision-snd (genesis SeiNetwork). Also refreshed two stale RPC-publishing comments.

Note

provision-node exists in the pinned seitask image (adca2d5), so no image bump needed; platform #1191 bumps SCENARIO_REF to this commit. Validated YAML.

🤖 Generated with Claude Code

[cursor]

PR Summary

Medium Risk
Touches core nightly chaos workflow provisioning for load and release validation; misconfiguration could still break benchmarks, but the change corrects an already-failing path and mirrors in-repo patterns.

Overview
Fixes load-test and release-test nightly workflows that failed at the RPC step because provision-snd only provisions SeiNetwork resources while the RPC templates are SeiNode followers.

The provision-rpc-fleet task now calls provision-node with --replicas (2 for load-test, 1 for release-test), --network=$SEI_CHAIN_ID for genesis peer wiring, and --running-timeout instead of --ready-timeout. Validator provisioning stays on provision-snd.

Task deadlines are raised (42m / 32m) with inline budgeting for sequential catch-up and EVM readiness plus Chaos Mesh scheduling/image-pull slack. Comments are updated for provision-node endpoint publishing, the redundant wait-rpc-caught-up gate, and single-node RPC routing in release-test.

^{Reviewed by Cursor Bugbot for commit 8956eb2. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d29f3d3. Configure here.}

[bdchatham]

/xreview ledger (Class: cross-component, T2)

Slate requested: systems-engineer (esp.), sei-network-specialist (dissenter), platform-engineer, prose-steward.
Ran: sei-network-specialist ✅, platform-engineer ✅, prose-steward ✅. systems-engineer ❌ blocked (account out of usage credits, resets Jun 30).

Boundary	Lens	Verdict	Resolution
key-name contract (RPC_* vars)	sei-network, platform	COMPATIBLE	provision-node publishes RPC_EVM_RPC_LIST/RPC_EVM_RPC/RPC_TM_RPC/RPC_REST/CHAIN_ID byte-identical to provision-snd
peer-wiring (--network)	sei-network	COMPATIBLE	synthesized sei.io/seinetwork=<chain> LabelPeerSource resolves; followers genuinely sync
replica counts (load=2, release=1)	sei-network, platform	COMPATIBLE	load feeds EVM_RPC_LIST; release collapses to one node for stateful mocha
flag contract @ adca2d5	platform	COMPATIBLE	all flags exist; --ready-timeout→--running-timeout rename mandatory
workflow-vars CM name/ns + ownerRef cascade	platform	COMPATIBLE	workflow-vars-<wf> matches downstream envFrom; ownerRef intact
wait-rpc-caught-up tautology / stale "aggregate" comment	sei-network (correctness), prose (correctness), platform (advisory)	RESOLVED	RPC_TM_RPC is node-0 (not aggregate); provision-node gates caught-up inline → comment rewritten to label the step an honest redundant re-confirm (2be6a6c)
Task deadline 25m < worst-case readiness budget	sei-network (correctness), platform (advisory)	RESOLVED	bump provision-rpc-fleet deadline (load→40m, release→30m) past 18m + N×2×5m so typed exit beats the Chaos-Mesh kill (2be6a6c)

Prose addendum (advisory, deferred): SND-nomenclature in both file headers + provision-snd in release-test's primitives list + the was --ready-timeout history breadcrumb — accurate-enough divergences, worth a sweep but non-gating.

State: OPEN — systems-engineer lens (operator-requested) pending on credit reset. Dissenter ran and its correctness findings are resolved; platform COMPATIBLE; prose correctness resolved. Holding merge for the systems pass per operator request (or operator accept-with-risk: sei-network + platform independently covered the systems brief — the readiness/budget/reliability boundary).

[bdchatham]

/xreview ledger — UPDATED (systems-engineer lens closed)

The previously credit-blocked systems-engineer lens has now run. State: RESOLVED.

Boundary	Lens	Verdict	Resolution
deadline-vs-budget math + typed-exit-beats-kill invariant	systems-engineer	COMPATIBLE	budget verified against provision.go: waits are strictly sequential (one shared 18m waitForRunning + a serial per-node loop of WaitCaughtUp→WaitEVMServing, 5m each), so N=2→38m / N=1→28m is the true worst-case wall clock. Invariant holds.
sequential-vs-concurrent (N× multiplier)	systems-engineer	COMPATIBLE	confirmed serial for loop, no goroutines; 18m is a flat shared term, only the readiness probes are ×N — comment is accurate.
provision-snd→provision-node failure modes	systems-engineer	COMPATIBLE	no regression; stronger semantics — no silent partial-fleet, first-non-ready-node fails the whole Task with the node name, idempotent on Task restart (AlreadyExists tolerated, fresh outer deadline per attempt).
failure observability	systems-engineer	COMPATIBLE (net improvement)	typed, node-localized signal (SeiNode name + gate /status caught-up vs eth_blockNumber; Infra vs Task exit class) — strictly better than provision-snd's aggregate-fleet wait.
2m outer slack vs pod-scheduling + cold image-pull	systems-engineer (advisory)	RESOLVED	the Chaos-Mesh deadline clock starts at Task admission, so the slack must also absorb scheduling + a cold $SEITASK_IMAGE pull (outside the inner budget). Bumped deadline 40m→42m (load) / 30m→32m (release) for 4m headroom (8956eb2).

All 4 requested lenses have now run (sei-network dissenter ✅, platform ✅, prose ✅, systems-engineer ✅); all correctness findings resolved; the one systems advisory hardened rather than accepted. Merge gate clear pending CI re-green on 8956eb2.