fix(compliance): thread brand through session-scoped storyboard steps (#2236)

[bokelley]

Summary

• Three media_buy_seller scenarios had brand: { domain: "acmeoutdoor.example" } on the opening get_products / create_media_buy step but omitted it on the follow-up get_media_buys / update_media_buy / sync_creatives steps. That made each run land in open:default for the later calls while create_media_buy wrote to open:acmeoutdoor.example, so any seller scoping session state by brand (spec-required for multi-tenant isolation) correctly refused to return the buy — and the storyboard scored it as a failure.
• Added brand: { domain: "acmeoutdoor.example" } to every session-scoped step across inventory_list_targeting, invalid_transitions, pending_creatives_to_start (9 steps, matching the already-passing inventory_list_no_match pattern).
• Left sync_accounts / sync_governance / comply_test_controller steps alone — their handlers are global / test-control and don't derive from sessionKeyFromArgs.

Why

Closes the verification gap in #2236. Training-agent handlers for get_collection_list / update_collection_list / create_media_buy with targeting.collection_list are correct (see server/src/training-agent/inventory-governance-handlers.ts and task-handlers.ts:138-149). The remaining failures on live test-agent.adcontextprotocol.org were entirely driven by the storyboard harness asking the wrong session.

@adcp/client 5.2.0 added applyBrandInvariant as a runner-layer safety net (PR adcp-client#586), but it only fires when the caller sets options.brand — the CLI doesn't currently expose that flag, so each step's sample_request.brand is authoritative in the typical invocation. Fixing the YAMLs is the right layer until the CLI lands a passthrough.

Protocol review

ad-tech-protocol-expert confirmed per-request brand is spec-correct — AdCP has no session-login primitive, MCP is stateless per call, and relying on prior-step brand would break horizontal scaling. Task classification (which tasks need tenant scoping vs. which don't) matches sessionKeyFromArgs semantics in server/src/training-agent/state.ts:307-330.

Follow-ups (filed, non-blocking)

• Storyboard lint: require brand/account on every session-scoped step #2527 — storyboard lint rule: any stateful: true step on a tenant-scoped task must carry brand or account.brand.
• Storyboard consistency: standardize buyer-side steps on account { brand, operator } #2528 — standardize buyer-side storyboards on account { brand, operator } (currently mixed with top-level brand).
• storyboard CLI: --file path lands in positionalArgs, trips "Cannot combine ID with --file" adcp-client#637 — CLI --file <path> lands in positional args, blocks local-YAML verification.
• storyboard CLI: expose --brand so applyBrandInvariant fires from adcp storyboard run adcp-client#639 — CLI --brand passthrough so applyBrandInvariant fires from adcp storyboard run (defense-in-depth behind the storyboard lint).

Close #2236 once the next @adcp/client release pulls this into the compliance cache and inventory_list_targeting scores green end-to-end.

Test plan

• npm run build:compliance succeeds with the edited YAMLs.
• node -e parses all three files via js-yaml; every session-scoped step's sample_request.brand.domain === "acmeoutdoor.example".
• Precommit hook passed (587 tests, typecheck clean).
• After release + cache sync, re-run npx @adcp/client storyboard run test-mcp media_buy_seller/inventory_list_targeting against test-agent.adcontextprotocol.org and confirm 5/5 steps pass.

🤖 Generated with Claude Code