feat(testing): storyboard runner matrix entry for examples/v3_reference_seller

[bokelley]

Context

The CI storyboard runner currently targets examples/seller_agent.py only. The v3 reference seller (examples/v3_reference_seller/src/app.py) now has all 9 sales-non-guaranteed methods, projection guard, and traffic counters (per #405) — but ships unverified at the storyboard level. End-to-end coverage is missing for the seller adopters are most likely to fork.

The Postgres service required by the v3 ref seller is already provisioned for the postgres-conformance job, so the same service can be reused.

Acceptance Criteria

• Dedicated GHA matrix job that boots examples/v3_reference_seller/src/app.py
• Storyboard runner executes against the booted seller
• Asserts at least one sales-non-guaranteed storyboard passes end-to-end
• Asserts /_debug/traffic shows the expected per-method counts (the anti-façade contract from PR feat(decisioning): MockAdServer Protocol + /_debug/traffic counters (#383) #405)
• Reuses Postgres service from postgres-conformance job

Refs

• PR feat(decisioning): MockAdServer Protocol + /_debug/traffic counters (#383) #405 (traffic counters / MockAdServer Protocol)
• PR feat(v3-ref-seller): broaden sales surface + sync_accounts + invoice_recipient #408 (broadened sales surface + sync_accounts + invoice_recipient)
• PR feat(v3-ref-seller): enable server validation + declare account.supported_billing #402 (account.supported_billing)

[bokelley]

Triage

Classification: Feature request
Bucket(s): examples, ci (neither label exists in this repo — gap flagged in run summary)
Status: ready-for-human

What the experts said:

• dx-expert: Approach is sound; asyncpg missing from all extras is a startup-fatal blocker; SubdomainTenantMiddleware will 400 every storyboard request because no tenant row is seeded in CI — seed.py must run or tenant routing must be bypassed for loopback; "reuses Postgres service" is imprecise — GHA jobs don't share service containers, new job needs its own services: block.
• code-reviewer: Confirms asyncpg blocker (also sqlalchemy[asyncio] — verify it's present transitively); GHA services-sharing is architecturally impossible; /_debug/traffic assertion is stable and non-flaky; new job should start with continue-on-error: true matching the existing storyboard job pattern.

My take: The overall design (dedicated job, Postgres service, storyboard runner + /_debug/traffic anti-façade assertion) is exactly right and worth shipping — the traffic-counter assertion prevents the facade failure mode. Two blockers need resolving before the workflow is written: asyncpg must be added to pyproject.toml (the [pg] extra currently only covers psycopg3), and a seed.py step (or a loopback-tenant fixture) must be added to the job so SubdomainTenantMiddleware can resolve requests.

Blockers to resolve (in implementation order):

1. Add asyncpg>=0.29 (and verify sqlalchemy[asyncio]>=2.0 is not relying solely on a transitive dep) to a pyproject.toml extra — the job install step needs it.
2. Add a seed step that runs python seed.py (or equivalent minimal fixture) after Postgres readiness, before booting app.py — without a seeded tenant row, SubdomainTenantMiddleware returns no tenant and all handler calls fail.
3. New job needs its own services: postgres: block (same spec as pg-conformance) — GHA service containers are job-scoped and cannot be shared.
4. Start with continue-on-error: true + promotion criteria comment (matching the existing storyboard job pattern).

This triage routine cannot edit .github/workflows/** or pyproject.toml, so implementation is for a human maintainer.

---

Triaged by Claude Code. Session: https://claude.ai/code/cse_01CMZYds8ZeAkNiynphffwUp

---

Generated by Claude Code