Scenario contract enforcement: build-time guards + single-sourced CM name

Start with the Layer 1 tests added in sei-k8s-controller#339, the wrapper bash, and the files under scenarios/. Read how the current envsubst inputs and scenario YAML are wired, then determine whether Layer 2 shape (a) or (b) is selected before designing the validation work. Done means the chosen CM-name approach is migrated, scenario validation exists, and the stated CI and pre-apply checks run.

Problem

The seitask workstream has surfaced a recurring failure pattern: contract drift between the seitask binary's internal helpers (`WorkflowVarsName`, scheme registration, downward-API env contract) and the scenario YAML / RBAC layer that has to mirror them manually. Each of the last four PRs (sei-protocol/sei-k8s-controller#334, #337, #339, plus the in-flight #339 build-time tests) addressed a different facet of the same shape: an internal helper has a convention, the scenario author has to mirror it manually in YAML, no test catches the drift, the bug surfaces only at first cluster fire ~10 minutes into the run.

Platform-engineer (cross-review on #339):

"The scenario YAML is the integration contract between three things (the runtime binary, the chaos-mesh CR shape, the wrapper's envsubst inputs) and none of them validate it. Each bug surfaced at first cluster fire."

Impact

• Slow feedback loop. Each contract bug costs ~10–30 min of manual-fire + investigation + fix-PR + image rebuild + SCENARIO_REF bump + re-fire. We've done this loop four times in the last hour to get the harness past keygen.
• Compounds with scenario count. Adding a second/third scenario will repeat the contract surface from scratch. Without enforcement, each new scenario brings its own #337-class bugs.
• Build-time enforcement is cheap. Two narrow tests added in sei-protocol/sei-k8s-controller#339 already catch the two highest-frequency classes (scheme registration + CM-name drift) at `go test`. There's more we could enforce; this issue tracks the broader pattern.

Proposed approach

Three reinforcing layers, deferred-ranked by effort:

Layer 1 (already partially in #339): unit tests for internal contracts

• ✅ Scheme round-trip test for every typed CR provision-snd / keygen / upload-report constructs
• ✅ CM-name validation for scenario YAMLs that opt in
• ⏳ RBAC vs kubebuilder-marker reconciliation — defer; un-defer when we hit a third RBAC-class bug

Layer 2: single-source the CM name across YAML + binary

Two candidate shapes (both reviewers raised independently):

(a) Wrapper exports `SEI_WORKFLOW_VARS_CM=workflow-vars-${WORKFLOW_NAME}` env var; scenarios reference `$SEI_WORKFLOW_VARS_CM` via envsubst allow-list. Single string-builder lives in the wrapper bash. No new templating dependency.

(b) Render-time template helper `{{ workflowVarsCM }}` exposed in a scenario template engine. Aligns with how the runner subcommand already templates SeiNodeTask CRs and how provision-snd templates SND specs. Requires a scenario rendering engine the wrapper invokes (vs current envsubst).

Platform-engineer recommends (a) as the MVP; kubernetes-specialist recommends (b) longer-term. Both eliminate the manual-mirror failure mode.

Layer 3 (longer-term): `seitask scenario validate` subcommand

A schema-validator subcommand that:

• Parses scenario YAML
• Checks every `configMapRef.name` matches `WorkflowVarsName(metadataName)`
• Checks every `--var=KEY=...` flag matches a documented input on the target subcommand
• Checks every `$(VAR)` reference has a producer step earlier in the Serial
• Run pre-commit (Husky/lefthook), in CI, and pre-apply in the wrapper

Catches more bugs than Layer 1 unit tests because it has access to the full scenario semantics (DAG ordering, var producer/consumer matching), not just the YAML structure.

Source: platform-engineer cross-review on sei-protocol/sei-k8s-controller#339.

Relevant experts

• platform-engineer — owns the wrapper bash + envsubst contract; (a) lives entirely in their territory
• kubernetes-specialist — owns the operator-pattern alignment for (b); also the rbac-marker reconciliation in Layer 1
• product-engineer — should weigh in on which Layer 2 shape fits the longer-term scenario authoring DX

Acceptance criteria

This issue resolves when:

• Layer 1 partially done (scheme + CM-name tests in #339)
• Layer 1 RBAC reconciliation test added (when justified)
• Layer 2: pick (a) or (b) and migrate release-test + future scenarios to it
• Layer 3: `seitask scenario validate` subcommand exists, runs in sei-k8s-controller CI on every PR that touches `scenarios/`, runs pre-apply in the wrapper

Out of scope

• Workflow engine swap (sei-protocol/sei-k8s-controller#332 tracks longer-term Argo evaluation)
• Status-check shared template library (sei-protocol/sei-k8s-controller#330)
• Chaos-mesh fail-fast (sei-protocol/sei-k8s-controller#340)

References

• sei-protocol/sei-k8s-controller#334 — first contract bug class (downward-API UID assumption)
• sei-protocol/sei-k8s-controller#337 — second (CM name mismatch)
• sei-protocol/sei-k8s-controller#339 — third (scheme + RBAC) + Layer 1 partial implementation
• Memory: `feedback_prototype_first.md` survey-checkpoint pattern (this issue's recurrence count, 4, justifies hardening)

🤖 Generated with Claude Code