fix(compliance): flow-derived id captures, context-echo cleanup, format_id integrity step

[bokelley]

Summary

Closes #2585, #2589, #2597 at the storyboard-YAML layer. Each of these issues has a spec/doc side that already landed (#2585 + #2589 via #2593, #2597 is net-new); this PR completes the storyboard migration so the training-agent conformance suite stops tripping on hardcoded fixtures and missing context blocks.

#2588 is not touched — test:error-codes lint + alias file landed in #2595 + #2596; issue stayed open because the close-on-merge marker didn't fire.

What changed

#2585 — flow-derived IDs via $context.media_buy_id. Seven storyboards hardcoded mb_* literals against media buys created earlier in the same run. Each create_media_buy step now declares context_outputs: [{ name: media_buy_id, path: "media_buy_id" }], and downstream get_media_buys / update_media_buy / get_media_buy_delivery / report_plan_outcome steps resolve "$context.media_buy_id". Storyboards covered: protocols/media-buy/index.yaml, protocols/governance/index.yaml (governance_escalation phase), and specialisms sales-guaranteed / sales-non-guaranteed / sales-proposal-mode / sales-broadcast-tv / sales-catalog-driven / creative-generative/generative-seller. creative-ad-server and governance-delivery-monitor keep their literals — those storyboards reference prerequisite media buys (Pattern A / fixtures: block) with no creating step to capture from.

#2589 — two context-echo authoring bugs. Both storyboards asserted context.correlation_id on the response but did not send context: on the sample_request, relying on runner-synthesized context that #2593 now explicitly forbids.

• specialisms/content-standards/index.yaml:calibrate_after_policy_change — sends context.correlation_id explicitly (assertion value was also stale — it pointed at the prior step's correlation id).
• protocols/media-buy/scenarios/governance_denied_recovery.yaml:create_media_buy_denied — sends context: explicitly and tightens field_present: context to a field_value: context.correlation_id echo assertion. Exercises the echo-on-error contract directly (the step has expect_error: true).

#2597 — list_formats_integrity step on media_buy_seller. New step in product_discovery phase between get_products_brief and create_buy. Filters list_creative_formats by format_ids: ["$context.product_format_id"] (captured FormatID object from get_products) and asserts:

• formats[0] resolves (stale/typo'd format_id → empty array → fail).
• formats[0].format_id equals $context.product_format_id verbatim (agent cannot substitute a different format under the filter).

Implements the issue's Option B (baseline check for every seller agent). Cross-agent dispatch — calling the creative agent at format_ids[].agent_url directly rather than the sales agent under test — is a runner follow-up: for single-agent sellers the current step is an integrity check; for true cross-agent sellers the sales agent either proxies the lookup or fails with an empty result, which still catches the failure mode the issue describes.

Review addressed

• ad-tech-protocol-expert + docs-expert: narrative/expected on new step no longer claims "asks the creative agent" (dispatch is to the sales agent under test); dropped redundant field_present checks on schema-required subfields; added field_value equality on the round-tripped FormatID so the narrative claim is actually enforced by a validation (previously only field_present: formats[0]).
• code-reviewer: capture shape verified end-to-end (name → key alias in loader.js:31, explicit outputs applied after convention extractors in runner.js:842). $context.<name> substitution confirmed to walk nested objects and array elements and resolve to captured values including FormatID objects. test:storyboard-scoping, test:error-codes, test:schemas, test:examples, test:composed, test:unit, typecheck, build:compliance all green. Kept explicit context_outputs: on create_media_buy (even though the convention extractor at context.js:63 produces the same key) to match the Pattern B example in docs/contributing/storyboard-authoring.md — authoring-doc consistency over terseness.
• security-reviewer: no Must-Fix or Should-Fix findings. $context.<k> substitution is anchored whole-string regex (no prompt-injection surface). Context accumulator is run-scoped via a per-run object identity (no cross-run leakage). Correlation_id values are non-secret storyboard identifiers already committed to the repo.

Test plan

• npm run test:storyboard-scoping — 3/3 pass
• npm run test:error-codes — clean (0 errors, 1 pre-existing YAML-parse warning on runner-output-contract.yaml)
• npm run test:schemas — 7/7 pass
• npm run test:examples — 31/31 pass
• npm run test:composed — 15/15 pass
• npm run test:unit — 627/627 pass
• npm run typecheck — clean
• npm run build:compliance — clean
• Runner-integration verification in @adcp/client against the training agent (follow-up; this repo has no runner-integration test suite)
• Follow-up issue for cross-agent list_creative_formats dispatch in list_formats_integrity

🤖 Generated with Claude Code