@adcp/sdk 15.8.1 → 16.0.0-beta.1

package/compliance/cache/3.2.4/universal/storyboard-schema.yaml

@@ -0,0 +1,2749 @@
+# Storyboard Definition Schema
+#
+# Storyboards are narrative test workflows that walk agent builders through
+# the sequence of calls their agent will receive, with context at each step.
+#
+# Each storyboard targets a specific agent interaction model and
+# describes the flow from a caller's perspective.
+
+# --- Schema definition ---
+
+# A storyboard file must conform to this structure:
+#
+# id: string (unique identifier, e.g., "creative_template")
+# version: string (semver, e.g., "1.0.0")
+# title: string (human-readable title)
+# category: string — matches the storyboard's `id` top-level segment.
+#
+#   For *specialism* storyboards, this is the snake_case form of an entry in
+#   /schemas/enums/specialism.json (the wire-protocol enum declared by agents
+#   on `get_adcp_capabilities.specialisms[]`). The protocol enum is
+#   authoritative — when adding a new specialism storyboard, add the kebab-case
+#   form to specialism.json first. The current specialism categories
+#   (snake_case form) are:
+#
+#     Sales:       sales_guaranteed | sales_non_guaranteed | sales_proposal_mode
+#                  | sales_catalog_driven | sales_broadcast_tv | sales_social
+#     Creative:    creative_ad_server | creative_generative | creative_template
+#     Governance:  content_standards | property_lists | collection_lists
+#                  | governance_aware_seller | governance_delivery_monitor
+#                  | governance_spend_authority
+#     Signals:     signal_marketplace | signal_owned
+#     Brand:       brand_rights
+#     Audiences:   audience_sync
+#     Buyer:       buyer_discovery | buyer_activation | buyer_negotiation
+#                  | buyer_monitoring | buyer_recovery
+#     Orchestrator: orchestrator_multi_agent
+#
+#   For *universal / domain-level* storyboards, the category is descriptive
+#   only (not on the wire). Examples in current use: capability_discovery,
+#   schema_validation, error_compliance, security_baseline, si_baseline,
+#   media_buy_seller, media_buy_governance_escalation, signed_requests,
+#   idempotency, webhook_emission, deterministic_testing,
+#   runner_output_contract, v3_envelope_integrity. New universal storyboards
+#   pick a descriptive snake_case identifier; see existing files under
+#   static/compliance/source/universal/ and protocols/ for the conventions.
+#
+#   Scenario variants use <category>/<variant> form, e.g.
+#   governance_spend_authority/denied, creative_generative/seller,
+#   brand_rights/governance_denied, media_buy_seller/provenance_enforcement.
+# summary: string (one-line description for listings)
+# track: enum (optional — compliance track this storyboard contributes to:
+#   core | products | media_buy | media_buy_buyer | creative | reporting | governance |
+#   campaign_governance | signals | si | audiences | error_handling | brand | security)
+# required_tools: string[] (optional — tool names the storyboard uses; empty for protocol-level tests.
+#   Lenient any-of: missing all listed tools triggers a coverage-gap skip
+#   (skip_result.reason: missing_tool at step scope). See `required_any_of_tools`
+#   below for the strict per-family load-time gate that grades requirement_unmet.)
+# required_any_of_tools: array of OR-family objects (optional — declarative one-of-N
+#   tool-advertisement gate. Each entry is { tools: string[], rationale?: string };
+#   at least one tool in each family MUST be advertised by the agent under test.
+#   Multiple entries are AND-combined: every family must be satisfied independently.
+#   Distinct from `required_tools` (which is the lenient any-of for storyboard
+#   applicability — missing all listed tools triggers a coverage-gap skip).
+#   See the dedicated `required_any_of_tools` block below for full semantics.
+# narrative: string (paragraph explaining the overall flow)
+#
+# requires_scenarios: string[] (optional — scenario IDs from storyboards/scenarios/ that must pass alongside this storyboard.
+#   Scenarios are small, focused behavior tests (e.g., "media_buy_seller/accepts_governance").
+#   The compliance engine resolves and runs them alongside the main storyboard.
+#
+#   Flag flow across `requires_scenarios`:
+#     A `branch_set` flag contributed in a step of storyboard X is
+#     considered asserted if EITHER (a) a step in X's own phases asserts
+#     it via `assert_contribution` with `check: any_of, allowed_values:
+#     [<flag>]`, OR (b) a scenario Y listed in `X.requires_scenarios`
+#     asserts it via the same check. The `orphan_contribution` lint
+#     honors this flow.
+#
+#     Reverse flow (a contribution in scenario Y asserted in parent X)
+#     is NOT supported — scenarios MUST be internally self-grading.
+#     The restriction preserves standalone lintability of scenarios
+#     and prevents cycles in the composition graph. Scenarios that
+#     contribute a flag without asserting it in their own phases will
+#     surface as `orphan_contribution` when linted standalone.
+#
+#     Scenario IDs in `requires_scenarios` MUST match the referenced
+#     file's top-level `id:` exactly, and the referenced scenario file
+#     MUST exist in the source tree at build time. Duplicate IDs
+#     across files are a build-time error.)
+#
+# default_agent: string (optional — logical agent key used by multi-agent
+#   storyboard runners to route steps that do not have a unique specialism
+#   claimant in the runtime agents map. Resolved by the runner against the
+#   `agents` option passed to `runStoryboard({ agents: { sales: …, governance: …, … } })`.
+#
+#   Key shape: free-form non-empty string, matched verbatim against the
+#   runtime `agents` map's keys. The spec does NOT constrain the key to the
+#   specialism enum (`sales`, `signals`, `governance`, `creative`, `brand`)
+#   because production multi-agent topologies legitimately fan out per-property
+#   (`nyt_sales`, `wsj_sales`), per-region (`sales_eu`, `sales_us`), or
+#   per-brand-rights-holder. Authors choose the convention that matches their
+#   operator's CI invocation; portability across operators is the author's
+#   concern, not the spec's.
+#
+#   When to set it. Storyboards that exercise cross-domain tools — e.g.,
+#   `sync_creatives`, `list_creative_formats` — do not name a single specialism
+#   the runner can match. The author has the most context for which logical
+#   tenant should receive these calls (usually "wherever the seller is":
+#   `sales`). Encoding `default_agent: sales` once in the YAML beats
+#   re-asserting it on every CI invocation. (Note: `comply_test_controller`
+#   is routed via `prerequisites.controller_seeding`, not this field — the
+#   controller is a back-channel, not a specialism claimant.)
+#
+#   Resolution order (runner contract — see adcp-client#1066, adcp-client#1355):
+#     1. Step-level `agent:` override (if declared on the step).
+#     2. Specialism-claimant match against the runtime agents map (matched
+#        via each agent's `get_adcp_capabilities.supported_protocols`):
+#          - Exactly one agent claims the step's task's specialism → route there.
+#          - Zero claimants → fall through to slot 3.
+#          - Two or more claimants → runner MUST grade the step `unrouted_step`
+#            rather than picking arbitrarily. Multi-claim is an operator-config
+#            error; the storyboard cannot disambiguate it. Slot 3/4 do NOT
+#            rescue this case — silently picking one would mask the misconfig.
+#     3. Storyboard-level `default_agent` (THIS field) resolved against the
+#        runtime agents map. When the field is set but the key is absent from
+#        the map, the runner MUST grade the step `default_agent_unresolved`
+#        and MUST NOT fall through to slot 4 — silent fallback would invisibly
+#        override the storyboard author's encoded intent. Slot 4 fires only
+#        when the storyboard does NOT declare this field.
+#     4. Run-options `default_agent` passed to `runStoryboard({ default_agent })`.
+#        Same key-resolution rule as slot 3: set-but-unmatched grades
+#        `default_agent_unresolved`; unset falls through to slot 5.
+#     5. Fail-fast — runner raises `unrouted_step` and grades the step failed.
+#
+#   Single-agent runs ignore this field entirely — there is no map to resolve
+#   against, every step routes to the only configured agent. Authors SHOULD
+#   still declare it for clarity; multi-agent runners treat it as advisory and
+#   single-agent runners as a no-op.
+#
+#   Validation. The `<key>` MUST be a non-empty string. Validation that the
+#   key resolves to a configured tenant is a runtime concern (per the slot 3
+#   rule above), not a schema concern — the same storyboard runs against
+#   different topologies.)
+#
+# Applicability order: storyboard-level `requires_capability` and
+#   `requires_all_capabilities` predicates MUST be evaluated before the runner
+#   evaluates `requires`. When any capability predicate is not satisfied, the
+#   runner emits `not_applicable` and MUST NOT report a missing runtime
+#   requirement. Only an agent that selected the storyboard by satisfying every
+#   authored capability predicate can receive `requirement_unmet`. This keeps
+#   optional capability storyboards out of the coverage totals of agents that
+#   never claimed the capability.
+#
+# requires: string[] (optional — storyboard-level runtime requirement gate.
+#   Known values: controller | seeded_state | real_wire | webhook_receiver | multi_agent |
+#   trusted_match_publisher_auth_runner (forward-compat:
+#   unrecognized values are NOT schema violations — runners emit
+#   requirement_unmet at runtime; see Forward compatibility below).
+#   Default when absent: [real_wire] — untagged storyboards run unchanged.
+#   minItems: 1 — runners MUST fail-load on requires: [] (empty array) because
+#   no distinct semantic exists between the empty set and omitting the field;
+#   the error message SHOULD read "use requires: [real_wire] for the default
+#   behavior, or omit the field entirely."
+#
+#   Values and availability:
+#
+#     controller   — gate on agent advertising comply_test_controller.
+#                    Available when the agent's get_adcp_capabilities response
+#                    includes comply_test_controller in its tool list. When
+#                    unmet, the runner MUST emit skip_result.reason:
+#                    requirement_unmet with detail naming "controller".
+#                    Orthogonal to prerequisites.controller_seeding: the two
+#                    fields are independent. controller_seeding is an execution-
+#                    phase instruction ("inject a fixtures phase before the main
+#                    phases run"); requires: [controller] is a load-phase gate
+#                    ("skip the whole storyboard if the capability is absent").
+#                    Storyboards may declare both: the gate fires first, and
+#                    controller_seeding takes effect only if the gate passes.
+#                    A storyboard that declares prerequisites.controller_seeding:
+#                    true MUST also declare requires: [controller] (enforced by
+#                    tests/lint-storyboard-controller-requires.test.cjs). The
+#                    gate coalesces what would otherwise be a
+#                    missing_test_controller grade on every seeded step into a
+#                    single whole-storyboard requirement_unmet skip at load
+#                    time. Whether a given class of seller is expected to
+#                    implement comply_test_controller is a separate policy
+#                    question — see docs/building/by-layer/L3/comply-test-controller.mdx.
+#                    Mid-run transient loss: if the load-phase gate passes
+#                    (the agent advertised comply_test_controller at load
+#                    time) but the tool becomes unavailable mid-run, the
+#                    runner MUST grade the affected step missing_test_controller
+#                    per the existing per-step path, NOT a retroactive
+#                    requirement_unmet at storyboard scope.
+#
+#     seeded_state — gate on operator asserting out-of-band state provisioning.
+#                    Available when the operator declares (e.g., via a CLI flag
+#                    or runner option) that prerequisite state has been seeded
+#                    externally before the run. When unmet, the runner MUST emit
+#                    skip_result.reason: requirement_unmet with detail naming
+#                    "seeded_state". Scenarios fail naturally on the first
+#                    stateful step if state is not actually present.
+#                    ANTI-GAMING: trust signals (verified-live badges, Tier-2
+#                    production scores) MUST NOT consume requires: [seeded_state]
+#                    pass results without independent verification. The field
+#                    widens the graded set; it does not attest the state.
+#                    This is an operator-governance norm; the runner cannot
+#                    enforce it on behalf of downstream certification dashboards.
+#
+#     real_wire    — reserved for a future --mock-only execution mode. Current
+#                    behavior: no runtime effect (gate is always met). Authors
+#                    SHOULD omit this value rather than listing it explicitly,
+#                    since the default when the field is absent is semantically
+#                    equivalent to [real_wire] today (this equivalence holds
+#                    only while --mock-only is not active; the --mock-only
+#                    inversion may break it — see below).
+#                    IMPORTANT: the future --mock-only semantics INVERT the
+#                    skip direction compared to other values — this value will
+#                    cause a storyboard to skip when --mock-only is active (i.e.,
+#                    skip-when-mocked, pass-when-live), opposite to the
+#                    skip-when-absent logic of controller and seeded_state.
+#                    Whether the default-when-absent (omitting `requires`)
+#                    also triggers the skip under --mock-only is to be defined
+#                    when --mock-only ships; until then, authors MUST NOT rely
+#                    on any particular behavior for omitted vs. explicit
+#                    real_wire under a mock-only runner.
+#                    Composing real_wire with other values (e.g., requires:
+#                    [controller, real_wire]) is syntactically valid but the
+#                    combined behavior under --mock-only is unspecified; authors
+#                    SHOULD NOT mix real_wire with other values until --mock-only
+#                    semantics are defined.
+#                    Harnesses that consume this value MUST account for the
+#                    skip-direction inversion when --mock-only support lands.
+#
+#     webhook_receiver
+#                  — gate on the runner hosting a webhook receiver for this run.
+#                    Available when the runner is configured with the webhook
+#                    receiver contract (for example
+#                    `webhook_receiver: { mode: "loopback_mock" }` or an
+#                    operator-supplied proxy URL). When unmet, the runner MUST
+#                    emit skip_result.reason: requirement_unmet with detail
+#                    naming "webhook_receiver" rather than sending unresolved
+#                    `{{runner.webhook_url:...}}` placeholders to the agent.
+#                    This value may be declared explicitly on storyboards such
+#                    as webhook-emission; runners may also infer it from
+#                    webhook URL substitutions in sample_request payloads.
+#
+#     multi_agent  — gate on the runner being configured with every logical
+#                    agent key required by this storyboard's `default_agent`
+#                    and step-level `agent:` declarations. Available when the
+#                    runner receives an `agents` map containing at least two
+#                    logical agents and all authored route keys used by the
+#                    storyboard resolve against that map. When unmet, the
+#                    runner MUST emit skip_result.reason: requirement_unmet
+#                    with detail naming "multi_agent" rather than failing a
+#                    routed step with default_agent_unresolved or unrouted_step.
+#                    Storyboards SHOULD declare this when the scenario's setup
+#                    and execution require distinct logical tenants, such as a
+#                    seller agent plus a governance agent.
+#
+#     trusted_match_publisher_auth_runner
+#                  — gate on runner-owned, credential-profile-neutral Trusted
+#                    Match publisher-authentication probes. Available when the
+#                    runner has exact HTTPS `contextEndpoint` and
+#                    `identityEndpoint` values for the publisher-facing ingress,
+#                    plus an operator adapter exposing
+#                    `preparePublisherAuthProbe({ operation, credentialState })`.
+#                    The adapter supplies only declarative credential headers
+#                    and optional PEM-encoded client-certificate, private-key,
+#                    passphrase, and CA-certificate material for `absent` and
+#                    `invalid` credential states. It MUST NOT supply a URL,
+#                    dispatcher, fetch implementation, method, body, response,
+#                    SNI override, proxy, or TLS verification override. The
+#                    runner rejects Host, content-framing, hop-by-hop, forwarding,
+#                    routing-override, and proxy headers from the adapter and
+#                    constructs its own HTTP client/dispatcher. The runner owns
+#                    the POST, endpoint SSRF
+#                    validation, redirect rejection, timeout, bounded response
+#                    read, and output redaction. The adapter, not storyboard YAML
+#                    or the generic AdCP `auth` override, is responsible for
+#                    removing ambient deployment credentials or supplying a
+#                    credential known to be invalid for that deployment. When
+#                    unmet, the runner emits
+#                    skip_result.reason: requirement_unmet naming
+#                    "trusted_match_publisher_auth_runner"; it MUST NOT route
+#                    this pseudo-task through MCP or A2A as a fallback.
+#
+#   Forward compatibility: runners encountering a requires value they do not
+#   recognize MUST emit skip_result.reason: requirement_unmet (not fail-load)
+#   and MUST name the unrecognized value in skip_result.detail so dashboards
+#   can surface it. This matches runner-output-contract.yaml > requirement_unmet
+#   and is consistent with the forward-compat posture throughout the harness —
+#   an unrecognized gate is not a coverage gap (not_applicable would signal
+#   "agent did not claim the protocol"), it is an unmet requirement. Only
+#   requires: [] (empty array) is a hard fail-load — that is a schema
+#   constraint, not a semantic gate, and schema violations are always loader
+#   errors.
+#
+#   Interaction with per-step requires_tool: the two fields are independent.
+#   requires gates the whole storyboard at load time; requires_tool gates
+#   individual steps at execution time. Both may be declared on the same
+#   storyboard.)
+#
+# requires_capability: object (optional — storyboard-level applicability gate
+#   evaluated against the agent's `get_adcp_capabilities` response before
+#   executing phases. When the predicate is not satisfied, the runner skips the
+#   storyboard with skip_result.reason: not_applicable and detail naming the
+#   predicate path and declared value.
+#
+#   Shape:
+#     path: string (dot path into get_adcp_capabilities, e.g.
+#       "media_buy.supports_proposals" or "supported_protocols")
+#     equals: JSON value (optional — predicate passes when the path value is
+#       strictly equal to this value)
+#     contains: JSON value (optional — predicate passes when the path value is
+#       an array containing this value)
+#     not_contains: JSON value (optional — predicate passes when the path value
+#       is an array that does not contain this value; missing/non-array values
+#       do not satisfy the predicate)
+#     present: boolean (optional — predicate passes when the path exists when
+#       true, or is absent when false)
+#
+#   Exactly one of equals, contains, not_contains, or present SHOULD be declared. Runners
+#   that support stricter validation MAY fail-load a gate with no matcher or
+#   multiple matchers. The same field is also legal on a phase; phase-level
+#   gates skip only that phase and are useful when a universal storyboard
+#   contains protocol-specific phase families.)
+#
+# requires_all_capabilities: object[] (optional — storyboard-level AND gate
+#   for two or more capability predicates. Each entry has the same shape and
+#   matcher rules as `requires_capability`. Every entry MUST pass before the
+#   storyboard is selected. When this field and `requires_capability` are both
+#   present, all predicates across both declarations MUST pass.
+#
+#   minItems: 2 — use `requires_capability` for a single predicate. Runners MUST
+#   fail-load empty or single-entry arrays so an incomplete compound gate cannot
+#   silently widen a storyboard's applicability.
+#
+#   Compound gates fail closed when the raw `get_adcp_capabilities` payload is
+#   unavailable. Tool advertisement or auto-registration MUST NOT substitute
+#   for an explicit capability declaration. Supported by @adcp/sdk starting in
+#   14.0.0-beta.4.)
+#
+# required_any_of_tools: array (optional — declarative one-of-N tool-advertisement
+#   gate at storyboard load time. Each array entry is an OR-family object:
+#
+#     required_any_of_tools:
+#       - tools: [list_accounts, sync_accounts]
+#         rationale: "AdCP 3.0.9 §accounts/overview"
+#
+#   Per-entry object shape:
+#     tools: string[] — tool names; at least one MUST be advertised by the agent
+#                       under test. minItems: 2 (a single-tool family collapses
+#                       to `required_tools`; encode it there instead).
+#     rationale: string (optional) — human-readable spec citation surfaced in
+#                       skip messages so dashboards can attribute the gate to
+#                       the underlying normative requirement. Storyboard-author-
+#                       controlled string; runners that render this value into
+#                       LLM-facing surfaces MUST fence per
+#                       runner-output-contract.yaml > rendered_output_fencing
+#                       (the value flows into skip_result.detail, which is
+#                       already on the fenced list — this is defense-in-depth
+#                       for renderers that consume rationale directly).
+#
+#   Multiple entries are AND-combined: every family in the array MUST be
+#   satisfied independently. Empty array is fail-load — omit the field entirely
+#   for "no requirement" (same posture as `requires`).
+#
+#   Distinct from `required_tools`. `required_tools` is the lenient any-of for
+#   storyboard applicability — missing all listed tools triggers a coverage-gap
+#   skip. `required_any_of_tools` is the strict variant: missing the family is
+#   still a skip (not a fail-load — see "Runner behavior" below), but with
+#   distinct `detail` attribution so the skip aggregates against the spec rule
+#   rather than the storyboard scope. The two compose cleanly on the same
+#   storyboard.
+#
+#   Distinct from `provides_state_for`. `provides_state_for` is a per-step
+#   substitution mechanism that waives a downstream cascade when a peer step
+#   establishes equivalent state at runtime. `required_any_of_tools` is a
+#   storyboard-load-time gate on tool *advertisement* — it fires before any
+#   step executes and does not interact with step-level substitution. The two
+#   operate on different axes (storyboard scope vs step scope; advertisement-
+#   time vs run-time) and may coexist on the same storyboard.
+#
+#   Runner behavior. When an agent advertises NONE of the tools in any family,
+#   the runner MUST emit at storyboard scope:
+#
+#     skip_result.reason  = requirement_unmet
+#     skip_result.detail  = "missing_required_tool_family: needs <tool_a> or <tool_b>[ or <tool_c>...] (<rationale>)"
+#
+#   The `missing_required_tool_family:` sub-reason prefix in `detail` is the
+#   canonical attribution string (see runner-output-contract.yaml >
+#   `requirement_unmet` for the full enumeration of recognized detail
+#   sub-reasons and the canonical wire shape for separators when more than one
+#   gate fails). This matches the same `requirement_unmet`-with-detail pattern
+#   already used for `requires:` gates — no new top-level reason enum is added
+#   to `runner-output-contract.yaml`, so this field ships without bumping the
+#   contract version.
+#
+#   Per-family OR semantics. A family is satisfied as soon as the agent
+#   advertises ANY tool in `tools[]`. The runner MUST NOT additionally require
+#   that the named tools are exercised by storyboard steps — this field gates
+#   on tool *advertisement*, not on step execution. Per-step gating remains
+#   the job of `requires_tool` on individual steps.
+#
+#   Forward compatibility. Runners encountering this field on a storyboard
+#   they otherwise support MUST honor it. Runners that pre-date this field
+#   will load and execute the storyboard as if the field were absent; this
+#   is acceptable because the universal account-discovery rule (the first
+#   use case) is also enforced by the SDK's runner-level conformance gate,
+#   which fires independently of per-storyboard tagging. Future runners that
+#   migrate to per-storyboard attribution (see adcp-client#1642) consume this
+#   field for richer skip-cause aggregation.
+#
+#   Composition with other gates. `required_any_of_tools` is evaluated at
+#   load time alongside `requires`. Either gate failing produces
+#   skip_result.reason: requirement_unmet at storyboard scope. Runners SHOULD
+#   emit a single sub-reason in `detail` when only one gate is unmet. When
+#   multiple gates fail, runners MAY aggregate sub-reasons per the wire shape
+#   pinned in runner-output-contract.yaml > `requirement_unmet`; aggregated
+#   `detail` strings are intended for human display and SHOULD NOT be
+#   parsed beyond the first sub-reason by automated consumers. Dashboards
+#   that need exhaustive multi-gate attribution SHOULD surface a multi-gate
+#   indicator separately rather than splitting `detail` themselves.)
+#
+# agent:
+#   interaction_model: enum (stateless_transform | stateful_preloaded | stateful_push | stateless_generate | media_buy_seller | media_buy_buyer | marketplace_catalog | owned_signals | si_platform | brand_rights_holder | governance_agent | tmp_router)
+#     Descriptive authoring metadata, not a runner selection gate: agents do not
+#     publish one canonical interaction_model value on the wire. Authors MUST
+#     express executable scope through required_tools, requires_capability,
+#     requires_contract, supported_protocols, or specialisms instead. Runners
+#     MUST resolve those declared gates before context extraction or setup.
+#   capabilities: string[] (legacy descriptive capability labels; bundle selection is driven by get_adcp_capabilities.supported_protocols and specialisms)
+#   examples: string[] (real-world examples: "Celtra", "Innovid")
+#
+# caller:
+#   role: string (who initiates the calls: "buyer_agent", "orchestrator", "dsp")
+#   example: string (e.g., "Scope3", "Pinnacle Agency")
+#
+# prerequisites:
+#   description: string (what must be true before running this storyboard)
+#   test_kit: string (reference to a test kit file, e.g., "test-kits/acme-outdoor.yaml").
+#     Always a single path — test kits are not composed through this field. See
+#     "Test kit flavors" below for the two shapes a test kit may take and how
+#     storyboards compose brand identity with harness coordination when they need
+#     both.
+#   controller_seeding: boolean (optional — when true, the runner auto-injects a
+#     fixtures phase that seeds every entry in the storyboard's top-level
+#     `fixtures:` block via `comply_test_controller` before the main phases run.
+#     Storyboards that hardcode fixture IDs in sample_request payloads SHOULD
+#     declare this and populate `fixtures:`. Missing `seed_*` support on the
+#     agent grades the storyboard `not_applicable`, not failed.)
+#
+# fixtures: object (optional — declarative prerequisite state the runner seeds via
+#   comply_test_controller before executing phases. Structure:
+#     fixtures:
+#       products:
+#         - product_id: "test-product"
+#           delivery_type: "non_guaranteed"
+#           pricing_options:
+#             - pricing_option_id: "test-pricing"
+#               pricing_model: "cpm"
+#           # Optional seller-internal calendar state for availability-window
+#           # scenarios. Consumed by seeding only — never emitted as a product
+#           # response field. Windows are half-open RFC3339 instants.
+#           availability:
+#             min_bookable_days: 3
+#             booked_windows:
+#               - start_time: "2099-09-20T00:00:00Z"
+#                 end_time: "2099-09-27T00:00:00Z"
+#       accounts:
+#         - account_id: "acct_test"
+#           fixture:
+#             brand: { domain: "acme.example" }
+#             operator: "buyer.example"
+#             billing: "operator"
+#             sandbox: true
+#       creatives:
+#         - creative_id: "campaign_hero_video"
+#           status: "approved"
+#           format_kind: "video_hosted"
+#       plans:
+#         - plan_id: "gov_acme_q2_2027"
+#           budget: { total: 30000, currency: "USD" }
+#       media_buys:
+#         - media_buy_id: "mb_acme_q2_2026_auction"
+#           status: "active"
+#   Each top-level key maps to a seed_* scenario (accounts → seed_account,
+#   products → seed_product, pricing_options → seed_pricing_option,
+#   creatives → seed_creative, plans → seed_plan, media_buys → seed_media_buy).
+#   Storyboards that need an
+#   external measurement vendor catalog SHOULD add an explicit
+#   comply_test_controller step with scenario: seed_measurement_catalog. Product
+#   fixtures may carry `measurement_catalogs[]` entries only as a compatibility
+#   fallback when a storyboard needs to keep that snapshot beside the product
+#   contract. Foreign-key dependency DAG the runner MUST honor when auto-seeding:
+#
+#     account ───→ product
+#     product ──┬─→ pricing_option
+#               ├─→ plan
+#               └─→ media_buy
+#     creative ────→ media_buy
+#     plan ────────→ media_buy
+#
+#   Products seed before pricing_options and plans that reference them;
+#   products, creatives, and plans all seed before media_buys that reference
+#   them. See docs/building/implementation/comply-test-controller.mdx for the
+#   full seeding semantics and per-scenario param shapes.
+#
+#   Storyboards SHOULD prefer the `fixtures:` block over hardcoded sample_request
+#   IDs whose existence the runner cannot verify.
+#
+#   Hardcoded-literal fixtures in sample_request (e.g., product_id: "test-product"
+#   without a matching `fixtures:` entry) are permitted through 3.x for existing
+#   storyboards that pre-seed matching IDs out of band, and SHOULD migrate to the
+#   `fixtures:` block before 4.0. New storyboards authored after this spec edit
+#   MUST NOT hardcode fixture IDs without a corresponding `fixtures:` block or a
+#   test-kit substitution.)
+#
+# fixture_resolution: object (optional, introduced in 3.2 — runner-owned rules
+#   for satisfying fixture handles through the seller's normal AdCP surface).
+#
+#   Identity fields in `fixtures:` are run-scoped handles. They remain the
+#   caller-selected literal IDs for the `seed` strategy, but a seller is not
+#   required to persist those literals when another explicitly enabled strategy
+#   resolves the handle. Keeping this block separate from `fixtures:` is
+#   intentional: 3.1 runners continue to seed the unchanged fixture payload and
+#   never forward resolution metadata to the seller.
+#
+#   Structure:
+#
+#     fixture_resolution:
+#       products:
+#         - handle: "sports_display_auction"
+#           strategies: [seed, discover]
+#           match:
+#             - path: /delivery_type
+#               operator: equals
+#               value: non_guaranteed
+#             - path: /channels
+#               operator: contains_all
+#               value: [display]
+#             - path: /format_options
+#               operator: any_match
+#               where:
+#                 - path: /format_kind
+#                   operator: equals
+#                   value: image
+#                 - path: /params/width
+#                   operator: equals
+#                   value: 300
+#                 - path: /params/height
+#                   operator: equals
+#                   value: 250
+#       pricing_options:
+#         - product_handle: "sports_display_auction"
+#           handle: "cpm_auction"
+#           strategies: [seed, discover]
+#           match:
+#             - path: /pricing_model
+#               operator: equals
+#               value: cpm
+#             - path: /currency
+#               operator: equals
+#               value: USD
+#             - path: /floor_price
+#               operator: present
+#
+#   handles and parent scope:
+#     - `handle` MUST equal the corresponding identity value in `fixtures:`.
+#     - A pricing-option entry MUST also declare `product_handle`; the pair
+#       `(product_handle, handle)` is its identity because pricing_option_id is
+#       scoped to a product and may repeat across products.
+#     - Two handles MUST NOT bind the same seller entity unless both resolution
+#       entries set `allow_reuse: true`. The default is false so a storyboard
+#       requiring two contrasting products does not accidentally test one row
+#       twice.
+#
+#   strategies: non-empty array containing an ordered subset of:
+#     - seed: call the existing seed_* scenario with the authored literal ID and
+#       fixture exemplar.
+#     - discover: enumerate the normal read surface and match seller-owned
+#       entities. Product and pricing-option discovery uses a fully paginated
+#       `get_products` wholesale request, scoped to the run's resolved sandbox
+#       account when one exists. The request omits `brief`, filters, and `fields`
+#       so matching operates on the complete returned product objects.
+#     - construct: execute the entity-type-specific buyer lifecycle declared by
+#       the runner contract, then apply required force_* transitions. A runner
+#       MUST NOT invent a generic construct path for an entity type whose setup
+#       lifecycle has not been specified.
+#
+#   Backward compatibility: when a fixture has no matching
+#   `fixture_resolution` entry, its effective strategies are `[seed]`. Runners
+#   MUST NOT infer discoverability by treating the fixture object as a generic
+#   subset predicate. Existing fields include illustrative values and
+#   adversarial multi-entity relationships for which that inference would
+#   weaken or change the authored test.
+#
+#   match: non-empty array required when `strategies` includes `discover`.
+#   Every clause has an RFC 6901 `path` relative to the candidate entity and an
+#   operator from the following closed set:
+#     - equals: candidate value is JSON-deep-equal to `value`.
+#     - present: path resolves to a non-null value; `value` and `where` omitted.
+#     - contains_all: candidate array contains every JSON-deep-equal item in
+#       the `value` array; order is irrelevant.
+#     - any_match: candidate path is an array and at least one member satisfies
+#       every nested clause in non-empty `where`; `value` omitted.
+#     - canonical_format_satisfies: candidate format declaration at `path`
+#       satisfies the canonical format requirement in `value` using the same
+#       directional algorithm as the `canonical_format_satisfaction` check.
+#   Unknown operators are an authoring error, not a forward-compatible skip.
+#
+#   strategy disposition:
+#     - unavailable: an unadvertised seed scenario, no qualifying discovery
+#       candidate, or an unavailable construct prerequisite advances to the next
+#       declared strategy.
+#     - failed: a strategy the seller advertises or begins returns a transport,
+#       schema, protocol, pagination, or setup failure. The fixture resolution
+#       fails immediately; the runner MUST NOT fall back and hide the failure.
+#     - resolved: the handle binds to the seller identity and is pinned for the
+#       full run.
+#     - unsatisfied: every declared strategy was unavailable. The storyboard
+#       grades `fixture_unsatisfied`, not passed.
+#
+#   Discovery selection is deterministic. After excluding candidates already
+#   pinned to non-reusable handles, the runner sorts matches by UTF-8 bytewise
+#   seller identity (product_id; then product_id + U+0000 + pricing_option_id
+#   for pricing options) and selects the first. Response order is never the
+#   tie-breaker. An empty, valid result is an unavailable strategy; a malformed,
+#   rejected, incomplete-without-a-usable-continuation, or transport-failed
+#   discovery response is a conformance failure.
+#
+#   handle substitution: after the whole fixture dependency graph resolves,
+#   runners replace authored handles only in request fields whose resolved JSON
+#   Schema carries the matching `x-entity` annotation. Replacement is exact —
+#   handles embedded in larger strings are not rewritten. Product pricing-option
+#   handles resolve against the sibling/ancestor product binding; an ambiguous
+#   unscoped occurrence is an authoring failure. This translation applies to
+#   normal task calls and comply_test_controller force/simulate calls. The agent
+#   never sees the authored handle when discovery or construction supplied a
+#   seller-issued identity.
+#
+#   See runner-output-contract.yaml > fixture_resolution_result for evidence and
+#   `fixture_unsatisfied` grading. See governance decision DR-0010.
+#
+# --- Test kit flavors ---
+#
+# Files under `static/compliance/source/test-kits/` come in two shapes. The
+# distinction is not enforced by a separate field today — it's identified by
+# which fixture data the kit carries.
+#
+#   Brand kit — carries a brand identity and authentication fixture. Declares:
+#     id                 string (short identifier)
+#     auth.api_key       string (`demo-<kit>-v1` pattern; the demo Bearer the
+#                        runner sends on positive api-key probes)
+#       OR
+#     auth.basic         object (the demo HTTP Basic credential the runner sends
+#                        on positive Basic probes). Shape is either:
+#                          username: string
+#                          password: string
+#                        or:
+#                          credentials: string
+#                        where `credentials` is the unencoded `username:password`
+#                        pair used to build the Authorization header.
+#     auth.probe_task    string (the protected read the runner calls under
+#                        `auth: none` and with random-invalid credentials)
+#     brand              object (house + brand_id + names + logos + colors +
+#                        fonts + tone — full brand.json payload)
+#     + optional products, creatives, pricing_options, destinations, etc.
+#
+#   Today's brand kits: acme-outdoor, bistro-oranje, nova-motors, osei-natural,
+#   summit-foods. Used by storyboards that exercise brand-specific AdCP protocol
+#   flows (creative, media-buy, signals, governance, etc.).
+#
+#   Runner contract — carries a harness coordination contract. Declares:
+#     id                 string (short identifier)
+#     applies_to         object or list (which storyboards/specialisms consume
+#                        this contract, e.g., `{ universal_storyboard: signed-requests }`
+#                        or `{ specialism: sales-catalog-driven }`)
+#     + contract-specific fields describing how the runner should pre-configure
+#       the agent under test (e.g., `receiver_urls:`, `retry_replay_contract:`
+#       for webhook-receiver-runner; `endpoint_modes:` for signed-requests)
+#
+#   Today's runner contracts: signed-requests-runner, substitution-observer-
+#   runner, webhook-receiver-runner. They carry no credentials today — the
+#   runner injects no auth for these storyboards because the storyboards test
+#   harness infrastructure (signature verification, webhook receiver
+#   coordination, URL substitution observability), not brand-specific AdCP
+#   flows. A future runner contract could legitimately carry its own test-
+#   coordination principal (e.g., a callback-authenticating receiver), so the
+#   "no credentials" property is temporal, not structural.
+#
+# Enforcement. `scripts/lint-storyboard-test-kits.cjs` fails the build if a
+# kit under `test-kits/` declares none of `auth.api_key`, `auth.basic`, or
+# `applies_to` — that's the bimodal partition this section describes. The lint
+# tolerates kits that declare both a credential marker and `applies_to` (the
+# future-branded-runner case above).
+#
+# Composition. Storyboards that need BOTH a brand identity AND a harness
+# contract declare the brand kit in `prerequisites.test_kit` and opt into the
+# runner contract through the `requires_contract: <runner_id>` field on
+# specific assertion tasks that need it — today that's `expect_substitution_safe`
+# and `expect_webhook*` (see their task definitions below for the per-task
+# field shape). When the named contract is not in scope for the runner, the
+# task grades as `not_applicable` rather than failing. Example:
+# `specialisms/sales-catalog-driven/index.yaml` declares
+# `prerequisites.test_kit: "test-kits/acme-outdoor.yaml"` and its
+# `expect_substitution_safe` step carries `requires_contract:
+# substitution_observer_runner`.
+#
+# Storyboards that test pure harness infrastructure without a specific brand
+# (e.g., `universal/signed-requests.yaml`,
+# `universal/webhook-emission.yaml`) point `prerequisites.test_kit` directly
+# at the runner contract and fall back on `task_default:` when any
+# `$test_kit.operations.<name>` reference resolves to null (see Step field
+# docs).
+#
+# `test_kit=<path>` in the contradiction-lint fingerprint disambiguates the
+# two flavors — `auth=kit_default` resolves to the kit's static credential for
+# brand kits and to no-credential for runner contracts, but the two never
+# collide because the `test_kit=` component separates them.
+#
+# Functional request signing composes differently from assertion contracts.
+# A runner that implements `test-kits/signed-requests-runner.yaml >
+# functional_dispatch` loads that shared signing identity at run level and
+# applies it to ordinary functional task dispatch after capability discovery.
+# Storyboards KEEP their brand test kit and do not add `requires_contract` to
+# each signed step. This avoids duplicating the public conformance key across
+# every fictional-brand kit and lets signing compose with an assertion step's
+# existing webhook/substitution contract.
+#
+# Dispatch rules:
+#   1. This path is sandbox-only. The public conformance private scalar MUST
+#      NOT be used when the run targets production or disables sandbox routing.
+#   2. Bootstrap `get_adcp_capabilities` discovery remains unsigned so the
+#      runner can learn the seller's profile. For subsequent ordinary task
+#      calls, sign operations in `request_signing.required_for` and, when the
+#      contract declares `sign_supported_for: true`, operations in
+#      `supported_for` as well.
+#      Signing the supported set covers request-body-dependent MUSTs such as a
+#      webhook registration carrying `authentication` without hard-coding
+#      those predicates in the runner.
+#   3. Build the complete transport request first (context substitutions,
+#      generated idempotency key, protocol/version envelope, and exact JSON
+#      bytes), then add RFC 9421 headers. Resolve `content-digest` coverage from
+#      the advertised `request_signing.covers_content_digest` policy.
+#   4. Preserve independently configured transport authentication. A signed
+#      request may also carry bearer/API-key credentials; signing is not a
+#      replacement for the storyboard's principal selection.
+#   5. Mint a fresh nonce and signature for every dispatch or retry. Never
+#      replay previously generated signature headers onto changed bytes.
+#   6. Never fall back to unsigned dispatch after signer setup or invocation
+#      fails. Report a runner failure. Older runners that cannot implement the
+#      functional-dispatch contract use the declared `unavailable_behavior`
+#      compatibility disposition (`not_applicable`) rather than blaming the
+#      seller for `request_signature_required`.
+#
+# This is NOT a sandbox-mode verifier bypass. The seller processes the signed
+# request through its normal RFC 9421 authentication path. `endpoint_scope:
+# sandbox` scopes the deliberately public test key; it does not weaken seller
+# enforcement.
+#
+# Future: if a storyboard ever needs to compose TWO assertion runner contracts
+# simultaneously (e.g., substitution-observer + webhook-receiver), the step-
+# level `requires_contract:` field will need to accept a list. Functional
+# signing does not consume that slot and therefore does not trigger this case.
+#
+# See `docs/contributing/storyboard-authoring.md` for the author-workflow
+# side of kit selection.
+#
+# context: object (optional — initial values seeded into the runner's context
+#   accumulator before any step executes. Keys become $context.<key> references
+#   available throughout the storyboard run. Use this field to provide
+#   per-storyboard defaults for values that vary by operator topology (e.g.,
+#   a governance agent URL the runner can override at invocation time once the
+#   CLI supports --context on `storyboard run`).
+#
+#   Example:
+#     context:
+#       governance_agent_url: "https://test-agent.adcontextprotocol.org"
+#
+#   Precedence in the context accumulator (highest to lowest):
+#     1. Step `context_outputs:` captures (populated after a step passes)
+#     2. THIS field (storyboard-root `context:` block, fixed at run start)
+#     3. Test-kit substitutions where explicitly scoped into context
+#
+#   Values MUST be scalars (strings, numbers, booleans — JSON scalar types).
+#   Object and array values are not supported here — capture structured values
+#   from step responses via `context_outputs:` instead.
+#
+#   See "Context accumulator and substitution" below for the full precedence
+#   contract, runner substitution rules, and unresolved-substitution grading.)
+#
+# phases: array of Phase objects
+#
+# --- Phase ---
+#
+# id: string (unique within storyboard)
+# title: string (human-readable phase title)
+# narrative: string (paragraph explaining this phase from the caller's perspective)
+#
+# requires_capability: object (optional — same shape as the storyboard-level
+#   field above, but scoped to this phase. When the predicate is not satisfied,
+#   the runner skips the phase as not_applicable and MUST NOT treat the skipped
+#   phase as a prerequisite failure for independent sibling phases that declare
+#   `depends_on: []`.)
+#
+# depends_on: string[] (optional — overrides the cross-phase cascade default)
+#   Default semantics (field absent or undefined): "all prior phases." Every
+#   phase implicitly depends on every prior phase, so when ANY prior phase
+#   trips its stateful cascade (a stateful step fails or skips for a
+#   missing-state reason — see `runner-output-contract.yaml > cascade_rules.
+#   default_cascade`), every stateful step in this phase cascade-skips with
+#   `prerequisite_failed`. This default preserves the F6 round-2 pattern
+#   where setup phases establish state consumed by a later phase.
+#
+#   Two override forms:
+#
+#     - Independent phase: `depends_on: []` — explicitly declares the phase
+#       has no upstream dependencies and runs even if every prior phase
+#       tripped. Use for phases whose state derives entirely from in-phase
+#       steps (e.g., the phase creates its own media buy / session / account
+#       via a `comply_test_controller`-gated step) and that consume no
+#       `$context.*` value produced by an earlier phase. The
+#       `sole_stateful_step_exemption` in `runner-output-contract.yaml`
+#       addresses a related-but-narrower case at the step level; `depends_on:
+#       []` is the phase-level escape hatch and applies regardless of how
+#       many stateful steps the phase contains.
+#
+#     - Targeted dependency: `depends_on: ['phase_id', ...]` — only the
+#       named phases gate this phase's cascade. Other phases tripping is
+#       irrelevant. Listed phase IDs MUST exist in the same storyboard and
+#       MUST be declared earlier in the `phases[]` array.
+#
+#   Validation rules (enforced at storyboard load; fail at parse time):
+#     - Value MUST be an array of non-empty strings (or absent).
+#     - Self-references are rejected.
+#     - Forward references and unknown phase IDs are rejected.
+#     - Empty list is legal and means "no upstream dependencies."
+#
+#   Authoring guidance: prefer the default. Use `[]` only when the phase
+#   verifiably builds its own state in-phase — read the phase's
+#   `context_inputs` and `$context.*` references to confirm no cross-phase
+#   coupling exists. Use the targeted form when a phase depends on a
+#   specific subset of prior phases (e.g., consumption phase depends only
+#   on setup phases, not on sibling consumption phases that may legitimately
+#   skip on different adopters).
+#
+#   Branch-set peers are exempt from this cross-phase cascade: phases sharing a
+#   `branch_set.id` with `any_of` semantics do NOT cascade `prerequisite_failed`
+#   onto one another (they are mutually-exclusive alternatives, not a chain),
+#   independent of whether the dependency is the implicit default or an explicit
+#   `depends_on`. See `runner-output-contract.yaml >
+#   cascade_rules.branch_set_cascade_exemption`.
+#
+#   References: adcp-client#1161 (field introduction); adcp-client#1711
+#   follow-up + adcp#4750 (deterministic_testing scope tightening that
+#   surfaced the documentation gap).
+#
+# steps: array of Step objects
+#
+# --- Step ---
+#
+# id: string (unique within phase)
+# title: string (human-readable step title)
+# narrative: string (what's happening and why)
+# task: string (AdCP task name: list_creative_formats, preview_creative, build_creative, get_products,
+#               create_media_buy, sync_accounts, etc. May reference a test-kit field, e.g.
+#               "$test_kit.auth.probe_task", in which case `task_default` provides the fallback.)
+#   Runner-owned Trusted Match probes:
+#     `trusted_match_missing_auth_context_probe`,
+#     `trusted_match_invalid_auth_context_probe`,
+#     `trusted_match_missing_auth_identity_probe`, and
+#     `trusted_match_invalid_auth_identity_probe` send the step's schema-valid
+#     sample_request to the matching operator-supplied publisher-facing endpoint.
+#     TMP does not prescribe router-ingress paths; the endpoint MUST NOT be
+#     inferred from provider-facing `/context` and `/identity` paths. The runner
+#     obtains declarative credential-header/TLS configuration from
+#     `preparePublisherAuthProbe`, then owns the guarded HTTP request and bounded
+#     response read. It MUST NOT implement these tasks through the generic
+#     step-level `auth` override because that path is scoped to AdCP task
+#     transports. The result exposed to ordinary validations includes HTTP
+#     status, headers, and any body. These pseudo-tasks are valid only in a
+#     storyboard gated by
+#     `requires: [trusted_match_publisher_auth_runner]` and MUST NOT be
+#     dispatched through an AdCP task transport.
+# task_default: string (optional — default task name when `task` is a test-kit reference that resolves to null)
+# schema_ref: string (path to request schema, e.g., "creative/list-creative-formats-request.json" or "media-buy/get-products-request.json")
+# response_schema_ref: string (path to response schema)
+# doc_ref: string (path to documentation page)
+# comply_scenario: string (maps to @adcp/client testing scenario, e.g., "creative_sync")
+# expected: string (human-readable description of expected behavior)
+# stateful: boolean (does this step depend on state from a previous step?)
+#
+# sample_request: object (optional — example request payload for this step)
+#   Placeholders: `$context.<name>` and `{{prior_step.<id>.<field>}}`. See
+#   the "Substitution" section below for the full semantics. For the narrow
+#   case where the SDK's per-task request builder discards a body field
+#   from `sample_request`, use `context_inputs:` (above) as the
+#   post-builder injection path.
+#
+#   Creative asset expansion: an `assets` object MAY contain the reserved
+#   `$build_assets_from_format` key with a `$context.<name>` value. The context
+#   value MUST be either canonical `FormatOption.params` captured from
+#   `get_products` or a legacy `FormatId` resolved against formats captured from
+#   `list_creative_formats`. Before request-schema validation or transport, the
+#   runner MUST replace the entire directive-bearing `assets` object with an
+#   asset map keyed by each required slot's canonical `asset_group_id` or legacy
+#   `asset_id`. Values come from the storyboard's test kit and MUST satisfy the
+#   slot contract. Image slots use an `assets.images[]` fixture satisfying the
+#   declared dimensions; URL slots use `assets.click_url`; text slots use
+#   `assets.text` only through an explicit runner-maintained mapping from the
+#   exact declared `asset_group_id` or legacy `asset_id` to a fixture category
+#   (for example, `headlines`, `descriptions`, or `cta`). Runners MUST NOT infer
+#   text semantics from a slot-id substring or populate an unmapped slot (for
+#   example, `legal`, `disclaimer`, or `alt_text`) from an arbitrary fallback
+#   category.
+#
+#   Asset-fixture coverage exception: when the captured format is valid but the
+#   runner cannot populate a required slot from the test kit — including an
+#   unsupported asset type, an unknown text-slot semantic, or no image fixture
+#   satisfying the declared dimensions — the runner MUST stop before transport
+#   and grade the storyboard `not_applicable` with `skip_result.reason:
+#   fixture_unavailable`. As soon as the referenced format context is captured,
+#   the runner MUST preflight every remaining `$build_assets_from_format`
+#   directive that uses it before executing any intervening side-effecting
+#   step. On this disposition, the runner MUST terminate the storyboard without
+#   executing its remaining phases or emitting `prerequisite_failed` cascades.
+#   This early whole-storyboard preflight prevents a later fixture gap from
+#   leaving behind state such as an active media buy. The diagnostic detail MUST
+#   begin with
+#   `creative_asset_fixture_unavailable:` and name the slot id, asset type, and
+#   any unsatisfied constraint. This is a runner/test-kit coverage gap, not a
+#   grading signal about the agent, and MUST NOT contribute to failed steps.
+#   A malformed directive, or missing format context caused by a prior agent
+#   step failing or being skipped, follows the ordinary harness-error or
+#   prerequisite-failure path instead; this exception MUST NOT conceal agent
+#   behavior that the storyboard actually observed.
+# sample_response: object (optional — example expected response; also surfaces in published docs)
+#
+# auth: "none" | object (optional — overrides the transport's default credentials for this step)
+#   auth: none                                            → strip any configured credentials before sending
+#   auth: { type: api_key, from_test_kit: true }          → use the API key at `auth.api_key` in the test kit
+#   auth: { type: api_key, from_test_kit: "auth.principals.<name>.api_key" }
+#                                                         → select a named principal within a multi-principal kit
+#                                                           (forward-compatible shape; no kit exposes multiple
+#                                                           principals today)
+#   auth: { type: api_key, value_strategy: random_invalid }
+#                                                         → use a per-run random bogus API key (for invalid-key probes).
+#                                                           Runner generates `invalid-<32 random bytes>` per run.
+#   auth: { type: basic, from_test_kit: "auth.basic" }   → use the Basic credential at `auth.basic` in the test kit
+#   auth: { type: basic, value_strategy: random_invalid }
+#                                                         → use a per-run random bogus Basic credential (for invalid-Basic probes).
+#                                                           Runner generates random username/password bytes per run.
+#   auth: { type: oauth_bearer, value_strategy: random_invalid_jwt }
+#                                                         → use a per-run random JWT-shaped bogus Bearer token.
+#                                                           Runner generates `<base64url random>.<base64url random>.<base64url random>` per run.
+#
+#   Positive credential checks SHOULD use the runner's normal transport auth
+#   and initialized protocol session rather than step-level `auth:`. In MCP
+#   terms: authenticate the Streamable HTTP session, call `initialize`, then
+#   call the protected task. `auth:` overrides are intended for negative probes
+#   that deliberately strip credentials or send invalid values, where a raw
+#   unauthenticated-session rejection is acceptable evidence of rejection.
+#
+# DISALLOWED: `auth: { type: ..., value: "<literal>" }` — literal credentials in
+# storyboard YAML are a code smell (they bind the storyboard to a specific value,
+# can't rotate without rewriting, and leak plaintext identity into source).
+# `lint-storyboard-auth-shape.cjs` (rule `literal_value`) fails the build on this
+# shape. Use `from_test_kit` or `value_strategy` instead. See #2720.
+#
+# omit_idempotency_key: boolean (optional — when true on a task step, the runner
+#   skips `applyIdempotencyInvariant` and signals the SDK client via
+#   `skipIdempotencyAutoInject` so the request reaches the agent WITHOUT an
+#   `idempotency_key`. Use for missing-key rejection vectors — without this flag
+#   the SDK may auto-inject a key and the agent never sees the missing-key path.
+#   Read tools leave the key optional, so omission probes SHOULD also set this
+#   flag to prevent current or future runner/client auto-injection from masking
+#   the vector. Supplying a key on a pure read does not by itself require replay.)
+#
+# contributes_to: string (optional — marks this step as contributing a named flag
+#                         that a later assert_contribution step can require)
+# contributes: boolean (optional — shorthand for `contributes_to: <enclosing phase's branch_set.id>`
+#                       introduced alongside first-class `branch_set:` declarations.
+#                       Legal only inside a phase that declares `branch_set:`; current
+#                       SDK runners resolve `contributes: true` to the phase's branch_set.id
+#                       and raise a storyboard-load error on `contributes: true` outside
+#                       a branch_set phase or on a step that also declares `contributes_to`.
+#                       Use explicit `contributes_to` when a storyboard must remain
+#                       compatible with older runners that grade before loader
+#                       normalization. See adcp-client#693 and
+#                       `lint:storyboard-branch-sets` for the full rule set.)
+# contributes_if: string (optional — runner-evaluated expression gating the contribution)
+#   Supported grammar (single form; not a general-purpose DSL):
+#     "prior_step.<step_id>.passed"   → true if the named prior step in the same storyboard passed
+#   Storyboard authors should prefer expressing gating through step ordering when possible.
+#
+# provides_state_for: string | string[] (optional — declares that this step's pass
+#   establishes equivalent state for the named peer step(s) in the same phase. Used
+#   to rescue the cascade-skip default when one of two interchangeable stateful
+#   steps is missing on the agent: e.g., explicit-mode sellers that pre-provision
+#   accounts out-of-band declare `sync_accounts` as missing_tool but expose
+#   `list_accounts` as the canonical alternative — declaring
+#   `provides_state_for: sync_accounts` on `list_accounts` tells the runner that
+#   the substitute's pass satisfies the same downstream state contract.
+#
+#   Grammar parallels `contributes_to`: declared on the substitute step, names
+#   the target peer(s) whose state it provides. The two fields are independent —
+#   `contributes_to` flows branch-set flags into a later `assert_contribution`;
+#   `provides_state_for` waives a peer's missing-tool / missing-test-controller
+#   cascade so downstream stateful steps still run.
+#
+#   Array semantics: `provides_state_for: [A, B]` is ALL-OF — one substitute pass
+#   establishes state for both A and B simultaneously. ANY-OF semantics are not
+#   supported; if a substitute only conditionally satisfies a peer, model it as
+#   two separate substitute steps.
+#
+#   Validation rules (enforced by lint:storyboard-provides-state-for, fail at
+#   parse time):
+#     - Same-phase only. The target step MUST live in the same phase as the
+#       substitute step. Cross-phase substitution is rejected; a cross-phase
+#       state contract belongs in `context_outputs` / `context_inputs`, not
+#       this field.
+#     - Target step must exist in the same phase by `id`.
+#     - Target step MUST declare `stateful: true`. Stateless peers don't
+#       carry a state contract to substitute for.
+#     - Substitute step MUST declare `stateful: true`. A stateless step
+#       cannot establish equivalent state on the agent side.
+#     - Self-references (`provides_state_for: <own_id>`) are rejected.
+#     - Cycles (A→B and B→A in the same phase) are rejected. The peer-graph
+#       per phase MUST be acyclic.
+#
+#   Runner behavior (see runner-output-contract.yaml > skip_result.reasons.peer_substituted):
+#     - When the substitute step passes AND the target peer would otherwise
+#       grade `missing_tool` / `missing_test_controller`, the runner MUST
+#       grade the target peer with skip reason `peer_substituted` (not
+#       `missing_tool`) and MUST NOT cascade `prerequisite_failed` to
+#       downstream stateful steps in subsequent phases.
+#     - When the substitute step itself fails or is skipped, the cascade
+#       proceeds as if `provides_state_for` were absent — the substitute
+#       did not establish the state it claimed.
+#
+#   See adcontextprotocol/adcp#3734 (mechanism rationale) and
+#   adcp-client#1130 (the cascade-skip default this field rescues).
+#
+# context_outputs: array of ContextOutput objects (optional — capture values from
+#   this step's response into the runner's context accumulator so later steps can
+#   reference them via $context.<name>). Each entry:
+#     - name: string (the context variable to populate; MUST be unique within the
+#             storyboard run)
+#       path: string (JSON path against this step's response body, e.g.
+#             "media_buy_id", "accounts[0].account_id", "plans[0].plan_id")
+#       generate: string (alternative to `path`: mint a value at run start
+#             rather than capture from a response. Supported generator names:
+#               - uuid_v4   — RFC 4122 v4 UUID
+#               - opaque_id — RFC 4122 v4 UUID (alias; the two names exist for
+#                             spec-vs-implementation framing — use uuid_v4 when
+#                             the threaded value will appear as a UUID on the
+#                             wire, opaque_id when callers treat it as an
+#                             opaque correlation token).
+#             Exactly one of `path` or `generate` MUST be set on each entry.
+#             Setting both is a storyboard-load failure. Generated values are
+#             materialized once per storyboard run and stored under
+#             `context.<name>` so later steps' `$context.<name>` substitutions
+#             resolve to the same value. Use for deterministic correlation
+#             IDs (e.g. threading task_id through force_create_media_buy_arm
+#             register → create_media_buy → force_task_completion → tasks/get
+#             poll) without authors copy-pasting magic strings. Requires
+#             runner >= adcp-client v6.x with `generate` support (see
+#             adcp-client#1006); older runners reject the field at load.
+#
+#             Example — deterministic task_id threading:
+#               context_outputs:
+#                 - name: forced_task_id
+#                   generate: uuid_v4
+#               # Subsequent steps reference $context.forced_task_id
+#               # in sample_request / context_inputs.
+#
+#   Special path prefix `task_completion.<inner>`: when the immediate response
+#   carries a non-terminal AdCP response (status `submitted` / `working` /
+#   `input-required`, carrying a `task_id`), the runner polls the AdCP task
+#   until it reaches a terminal state and resolves `<inner>` against the
+#   completion result instead of the immediate response. On the AdCP v3 A2A profile,
+#   this means fresh `get_task_status` invocations; MCP may use its advertised
+#   AdCP polling task, including the legacy `tasks/get` alias. Use
+#   for captures whose value only exists on the completion artifact — e.g.
+#   the seller-assigned `media_buy_id` on an IO-signing / async-signed HITL
+#   flow where `create_media_buy` returns `submitted` and the ID lands on
+#   the completion artifact. Without the prefix, the literal key
+#   `task_completion` is looked up on the immediate response, which fails
+#   as `capture_path_not_resolvable`. Requires runner >= adcp-client v6.7;
+#   older runners treat the prefix as a literal key. See adcp-client#1417
+#   (rationale) and adcp-client#1426 (implementation).
+#
+#   Runner behavior:
+#   - Captures occur AFTER the step's validations pass. A failed step MUST NOT
+#     populate the context accumulator — downstream $context.<name> references
+#     resolve to an unresolved_substitution error, which grades the dependent
+#     step as failed rather than the storyboard passing on fabricated state.
+#   - Paths that do not resolve to a usable value in the response body (absent,
+#     null, or "") are a runner-side grading failure on THIS step
+#     (capture_path_not_resolvable), not on the downstream reader — the capture
+#     declared a contract that the response did not meet. The capturing step
+#     grades as failed; downstream steps that depended on the captured value
+#     grade as skipped with reason: prerequisite_failed citing this step.
+#     See runner-output-contract.yaml > validation_result for the output shape.
+#   - The accumulator is storyboard-run-scoped; values do not leak across runs.
+#   - Special path prefix `task_completion.<inner>`: when the immediate
+#     response carries a non-terminal AdCP status (submitted/working,
+#     with a task_id), the runner polls the AdCP task until it exits
+#     {submitted, working}. On `completed`: resolves <inner> against the
+#     `result` payload; a missing path grades this step as
+#     capture_path_not_resolvable. On any other status (failed, canceled,
+#     rejected, auth-required, input-required, unknown): `result` is absent;
+#     the capture grades as capture_path_not_resolvable on this step;
+#     downstream $context.<name> references grade as prerequisite_failed
+#     citing this step. See adcp-client#1426 for the implementation.
+#
+# context_inputs: array of ContextInput objects (optional — inject captured
+#   context values into the request at a specific path AFTER the per-task
+#   request builder runs).
+#
+#   Shape:
+#     context_inputs:
+#       - key: product_format_id        # name populated by a prior context_outputs capture
+#         inject_at: "format_ids[0]"    # JSON path inside the request body
+#
+#   Do NOT use `context_inputs` unless `$context.<name>` inside
+#   `sample_request:` has been tried and verified not to reach the wire.
+#   The default substitution path is `$context.<name>` placeholders inside
+#   `sample_request:` — keep the request body honest about what the agent
+#   will receive. Reasons the default is preferred:
+#   - The published sample_response/sample_request pair is what seller
+#     engineers read to implement against the storyboard; anything the
+#     runner injects out-of-band is invisible there.
+#   - `$context.<name>` produces a pointed `unresolved_substitution` failure
+#     when the referenced capture is missing. `context_inputs` is
+#     silently a no-op on a missing `key` (runner:
+#     `if (input.key in context) setPath(...)`) — a typo in `key` or a
+#     failed capture produces an empty field rather than a loud error.
+#
+#   Legitimate uses:
+#   - The @adcp/client per-task request builder for a given tool discards
+#     body fields from `sample_request` (the runner's post-builder merge
+#     only forwards envelope fields: `context`, `ext`, `idempotency_key`,
+#     `push_notification_config`). The observable symptom is that the
+#     agent answers with an unfiltered/default result and a round-trip
+#     or substitution-observer validation grades on fabricated state.
+#     `context_inputs` runs AFTER the builder, so it round-trips the
+#     captured value to the wire regardless.
+#   - Dependency-aware multi-instance dispatch, where the runner needs to
+#     route a step to a specific replica URL derived from a prior step's
+#     response (not applicable to single-instance storyboards).
+#
+#   Value types are preserved. The captured value is written to `inject_at`
+#   as-is — object and numeric captures are NOT stringified, matching the
+#   `$context.<name>` behavior described below.
+#
+#   Document every SDK-gap-workaround entry inline. The comment MUST name
+#   the SDK version that carries the gap and the upstream PR that closes
+#   it, so the entry can be removed when we bump past that release. See
+#   `static/compliance/source/protocols/media-buy/index.yaml`
+#   (`list_formats_integrity` step) for a worked example tied to
+#   adcontextprotocol/adcp-client#789.
+#
+# validations: array of Validation objects (optional)
+#
+# expect_error: boolean (optional, default false — when true, the runner expects
+#   the agent to return an error response for this step. Paired with `negative_path`
+#   to control whether the step's `sample_request` is schema-validated.)
+#
+# negative_path: "schema_invalid" | "payload_well_formed" (optional, default "schema_invalid")
+#   Disambiguates two categories of negative-path steps when `expect_error: true`:
+#   - schema_invalid (default): the `sample_request` payload is intentionally malformed
+#     (missing required field, wrong type, bad enum value, etc.) to verify the agent's
+#     validation response. Schema validation is SKIPPED on this step.
+#   - payload_well_formed: the `sample_request` payload is schema-valid but the agent
+#     rejects it at runtime — wrong state-machine transition, resource not found,
+#     governance denial, auth failure, temporal constraint, etc. Schema validation
+#     RUNS even though the agent is expected to return an error. Shape drift on these
+#     steps is caught statically.
+#   Omitting `negative_path` when `expect_error: true` behaves as `schema_invalid`
+#   (backwards-compatible default).
+#
+# sample_request_skip_schema: boolean (optional — explicit opt-out of schema
+#   validation for a step regardless of `expect_error` / `negative_path`. Use
+#   sparingly; prefer `negative_path: schema_invalid` for intentionally malformed
+#   negative-path fixtures.)
+#
+# expected_arm: string (optional — names the discriminator value of the `oneOf`
+#   arm this step expects the agent to return. When present, both the
+#   context-output-paths and validations-paths lints restrict path resolution
+#   to the matching arm only. Default behavior (no annotation) accepts a path
+#   that resolves through ANY oneOf arm — the right call for "no prior info,"
+#   but it lets a step capture or assert on a path that's defined on, say, the
+#   Acquired arm even when the storyboard semantically expects the Rejected
+#   arm. The annotation tightens that.
+#
+#   Match shape: the lint walks each `oneOf` / `anyOf` branch and looks for
+#   ANY property with `const: "<expected_arm>"`. The first matching branch
+#   is the expected arm; subsequent path resolution restricts to it.
+#
+#   Examples:
+#     expected_arm: "acquired"      # matches AcquireRightsAcquired (status.const = "acquired")
+#     expected_arm: "rejected"      # matches AcquireRightsRejected
+#     expected_arm: "pending_approval"  # matches AcquireRightsPendingApproval
+#
+#   When `expect_error: true` is set without `expected_arm`, the lint also
+#   tries to find an Error arm (one whose `required` list includes `errors`
+#   and which has no `status` const) and restricts to it. This covers
+#   `acquire_rights` / `creative_approval` / similar response shapes whose
+#   Error arm isn't const-discriminated.
+#
+#   When the named arm doesn't match any branch, the lint emits an
+#   `unknown_expected_arm` violation — an authoring bug to fix at storyboard
+#   author time, not at runtime.
+#
+# --- Phase (optional fields for conditional execution) ---
+#
+# optional: boolean (default false — see semantics below)
+# skip_if: string (expression — phase is skipped when the expression evaluates to true,
+#                  e.g., "!test_kit.auth.api_key")
+# branch_set: object (optional — first-class declaration of branch-set membership;
+#                     see "Branch sets" section below)
+#
+# Optional phase semantics:
+#   - An `optional: true` phase always runs unless `skip_if` evaluates to true.
+#     `skip_if` (e.g., "!test_kit.auth.api_key") means "skip this phase when the
+#     expression is true"; without `skip_if`, the phase runs every time.
+#   - Failures inside an optional phase do NOT fail the overall storyboard on
+#     their own — the phase is for accumulating contributions toward a later
+#     `assert_contribution` check. The final assert is what fails the storyboard
+#     when no optional path succeeded.
+#   - A non-optional phase that fails fails the storyboard unconditionally.
+#
+# Branch sets:
+#   A branch set is a group of peer optional phases exercising mutually
+#   exclusive agent behaviors for the same trigger — e.g., an operation whose
+#   spec permits either an immediate success response or a pending-review
+#   response depending on seller policy. A later `assert_contribution` step
+#   over the set's `branch_set.id` fails the storyboard only when NO branch
+#   contributed.
+#
+#   Minimal example — two peer phases and the assertion that grades them:
+#
+#     phases:
+#       - id: immediate_success_path
+#         optional: true
+#         branch_set:
+#           id: operation_handled
+#           semantics: any_of
+#         steps:
+#           - id: submit_immediate_success
+#             task: example_operation
+#             contributes_to: operation_handled
+#             # ... expect status: "accepted" ...
+#
+#       - id: pending_review_path
+#         optional: true
+#         branch_set:
+#           id: operation_handled
+#           semantics: any_of
+#         steps:
+#           - id: submit_pending_review
+#             task: example_operation
+#             contributes_to: operation_handled
+#             # ... expect status: "pending_review" ...
+#
+#       - id: operation_enforcement
+#         steps:
+#           - task: assert_contribution
+#             validations:
+#               - check: any_of
+#                 allowed_values: [operation_handled]
+#
+#   `@adcp/client` 5.8.0+ also accepts the boolean shorthand
+#   `contributes: true` on a step inside a branch_set phase — the runner's
+#   loader resolves it to the enclosing phase's `branch_set.id`. The two
+#   forms are semantically identical in current SDK runners. Use explicit
+#   `contributes_to: <branch_set.id>` for storyboards that need to run against
+#   older graders that may inspect the field before loader normalization;
+#   use `contributes: true` only when the target runner line is known to
+#   support shorthand normalization. The lint enforces both forms.
+#
+#   New storyboards MUST declare branch-set membership with a `branch_set:`
+#   object. The implicit-detection fallback below exists only so pre-#2633
+#   storyboards keep running unchanged.
+#
+#   `branch_set:` field shape:
+#     - id:        non-empty string. MUST equal the `contributes_to` value
+#                  on every contributing step inside the phase and MUST
+#                  appear in the `assert_contribution` step's any_of
+#                  allowed_values.
+#     - semantics: `any_of`. No other value is supported today.
+#
+#   Per-step contribution signalling is REQUIRED on each contributing step —
+#   the phase-level `branch_set:` declares membership but does not replace
+#   the step-level flag. Runners accumulate contributions at the step level.
+#   Two equivalent forms in current SDK runners:
+#     - `contributes_to: <branch_set.id>` (compatibility-safe string form;
+#       required outside a branch_set phase and must equal `branch_set.id`
+#       inside one)
+#     - `contributes: true` (shorthand inside a branch_set phase; resolves
+#       to the enclosing phase's `branch_set.id` on runners that support
+#       shorthand normalization)
+#
+#   Authoring rules (enforced by `lint:storyboard-branch-sets`):
+#     - A phase with `branch_set:` MUST set `optional: true`. A
+#       non-optional phase's failure would fail the storyboard
+#       unconditionally and defeat the any_of semantics.
+#     - All phases in a storyboard sharing the same `branch_set.id` MUST
+#       share the same `branch_set.semantics`.
+#     - A storyboard with a `branch_set:` declaration MUST contain an
+#       `assert_contribution` step whose `validations[].check: any_of`
+#       includes the `branch_set.id` in `allowed_values`. Any
+#       `contributes_to:` / `contributes: true` whose resolved flag has no
+#       matching assert_contribution is flagged as `orphan_contribution` —
+#       nothing grades it.
+#     - Any step inside a branch-set phase that declares `contributes_to`
+#       MUST use the value `<branch_set.id>`.
+#     - A step MUST NOT declare both `contributes` and `contributes_to`
+#       (ambiguous). `contributes: true` is legal only inside a phase
+#       that declares `branch_set:`.
+#     - If any phase declares `branch_set: { id: X, ... }`, every optional
+#       phase in the same storyboard whose steps contribute to X MUST also
+#       declare `branch_set: { id: X, ... }`. A mixed-mode storyboard (one
+#       peer declared, one peer relying on implicit detection) would be
+#       graded as a single-member set by a runner preferring the explicit
+#       declaration.
+#
+#   Runner grading:
+#     Runners MUST prefer the explicit `branch_set:` declaration when
+#     present and fall back to implicit detection otherwise (see below).
+#     - If at least one peer branch contributed the `branch_set.id`, the
+#       non-contributing branch's failing steps MUST be graded with skip
+#       reason `peer_branch_taken` (see
+#       runner-output-contract.yaml > skip_result.reasons.peer_branch_taken).
+#       They MUST NOT surface as `failed` in the storyboard summary, and
+#       MUST NOT be conflated with `not_applicable` (reserved for coverage
+#       gaps — agent didn't declare the protocol).
+#     - If no branch contributed, the final `assert_contribution` step
+#       fails the storyboard. Individual branch step failures still grade
+#       `failed` so the implementor sees what each branch observed.
+#
+#   When to use a branch set:
+#     Only when the AdCP spec permits multiple conformant behaviors for
+#     the same trigger. If the spec mandates one behavior, write a
+#     required phase — a branch set where only one branch is legal
+#     weakens the test to any_of and hides regressions. Branch sets are
+#     not for optional features (use `skip_if` and coverage flags) or
+#     for progressive disclosure (use sequential phases).
+#
+#   Authoring guidance: keep branch sets flat — two or three peer phases
+#   is typical. Nested branch sets (a branch set whose member is itself a
+#   branch set) are out of scope for runners. If a scenario needs deeper
+#   branching, restructure into separate storyboards or a linear sequence.
+#
+#   Implicit branch-set detection (DEPRECATED — pre-#2633 storyboards only):
+#     When no phase in a storyboard declares `branch_set:`, runners fall
+#     back to the legacy rule: a branch set is the set of `optional: true`
+#     phases whose steps declare the same `contributes_to: <flag>` value
+#     as a later `assert_contribution` step's
+#     `validations[].check: any_of, allowed_values: [<flag>]` target.
+#     Fragile against typos in `contributes_to:` — new storyboards MUST
+#     use the explicit form above. This fallback exists only so
+#     pre-#2633 storyboards keep running during migration.
+#
+# Cross-storyboard contradiction lint:
+#   The `lint:storyboard-contradictions` script groups every step-with-
+#   assertions across every storyboard by (task, canonicalized request
+#   fingerprint, prior-state fingerprint, env fingerprint) and flags
+#   groups whose asserted outcomes disagree in a way no conformant agent
+#   can satisfy (success ↔ error, or disjoint error-code sets).
+#
+#   Env fingerprint includes the storyboard's top-level `id:`, so two
+#   separate storyboards exercising different controller-seeded states
+#   are legitimately independent test suites and do not collide. To make
+#   two different storyboards' assertions collide — e.g., to catch
+#   "storyboard A assumes media_buy_id X is active, storyboard B
+#   assumes it is canceled" — use matching `comply_scenario:` values or
+#   shared `prerequisites.controller_seeding`.
+#
+#   Branch-set peers (two optional phases in the same storyboard sharing
+#   a `branch_set.id`) are exempt from contradiction flagging by design:
+#   any_of semantics intentionally asserts mutually exclusive outcomes.
+#
+#   Mutating-task detection: the lint's prior-state fingerprint partitions
+#   steps by the ordered list of prior MUTATING tasks. A task is mutating
+#   iff its request schema declares `"x-mutates-state": true` at the top
+#   level. The annotation is consumed by
+#   `scripts/lint-storyboard-contradictions.cjs` (which builds the
+#   mutating-task set at lint time) and is the single source of truth for
+#   this concern.
+#
+#   Decidability rule (apply to any new task):
+#     A task declares `"x-mutates-state": true` iff a conformant compliance
+#     suite could write an assertion — on any later `get_*`, `list_*`, or
+#     follow-up call — whose outcome depends on this request having run.
+#     Equivalently: "this task changes observable server state a later
+#     conformant call may assert against."
+#
+#   Mutating task classes (illustrative; the decidability rule above is
+#   the actual test):
+#     - Writes (create/update/delete/sync/build) — resource lifecycle.
+#     - State transitions (cancel, approve, reject) — status moves.
+#     - Session-scoped primitives (si_initiate_session, si_send_message,
+#       si_terminate_session) — session lifecycle.
+#     - Reporting/audit writes (log_event, report_usage,
+#       report_plan_outcome, provide_performance_feedback) — mutating
+#       because `get_plan_audit_logs` and billing-summary tasks can assert
+#       against their recorded outputs.
+#
+#   Non-mutating task classes:
+#     - Read-only tasks (get_*, list_*, check_*, validate_*, preview_*) do
+#       not declare `x-mutates-state` UNLESS the read itself produces
+#       observable state a later task can assert against (e.g., a view
+#       counter that `get_analytics` can read) — those are mutating
+#       regardless of prefix.
+#     - Pure-compute tasks (no state read or write — estimates,
+#       calculations) MUST NOT declare `x-mutates-state`.
+#
+#   Relationship to idempotency_key (decoupled, do not unify):
+#     `x-mutates-state` declares mutation semantics. `required:
+#     [idempotency_key]` declares the idempotency mechanism. The sets
+#     overlap ~95% but legitimately diverge for naturally-idempotent
+#     mutations. Example: `comply_test_controller` declares
+#     `x-mutates-state: true` but omits `idempotency_key` from `required`
+#     because the `scenario` enum is the dedup boundary — replaying
+#     `force_media_buy_status=active` converges to the same observable
+#     state without a key. `si_terminate_session` is the same shape
+#     (session_id is the dedup boundary). `scripts/build-compliance.cjs`
+#     reads `idempotency_key` for its own enforcement; do not try to share
+#     the two reads.
+#
+# --- Validation ---
+#
+# CANONICAL CHECK ENUM
+# --------------------
+# The authored check kinds storyboards MAY declare in `step.validations[].check`
+# live in `runner-output-contract.yaml > authored_check_kinds` (structured
+# YAML; this file is comment-only documentation). The build-time lint
+# `scripts/lint-storyboard-check-enum.cjs` reads from there; runtime
+# forward-compat (unknown kinds → not_applicable) is documented in
+# runner-output-contract.yaml > validation_result. Add new kinds to that
+# enum AND document semantics here in the same PR.
+#
+# id: string (optional stable machine-readable identifier for this validation.
+#                Use when a step declares multiple validations that need a
+#                durable cross-reference in runner output, dashboards, support
+#                requests, or SDK adapters. The value is author-chosen and
+#                stable across description/path edits; it SHOULD be unique
+#                within the containing step's `validations` array. Runners
+#                echo this value in `validation_result.id`; see
+#                runner-output-contract.yaml > validation_result.)
+# check: string (what to validate: "response_schema", "field_present",
+#                "field_pattern" (path MUST resolve to a string matching
+#                `pattern`, a JavaScript regular expression source),
+#                "envelope_field_present" / "envelope_field_absent" (walks
+#                protocol-envelope.json instead of response_schema_ref — use for
+#                top-level envelope fields like `status`, `task_status`, and
+#                `response_status`),
+#                "envelope_field_pattern" (same as field_pattern, but scoped to
+#                protocol-envelope.json / version-envelope.json instead of
+#                response_schema_ref),
+#                "field_value",
+#                "field_value_or_absent",
+#                "field_contains" (wildcard-aware membership check; `path` may
+#                use `[*]` segments and passes when any resolved value matches
+#                `value` or one of `allowed_values`; object/array expectations
+#                are matched as deep subsets),
+#                "array_length" (asserts exact array cardinality with `value`,
+#                or an inclusive range with `min` / `max`),
+#                "status_code", "http_status", "http_status_in",
+#                "error_code", "on_401_require_header", "resource_equals_agent_url",
+#                "oauth_metadata_graph",
+#                "any_of", "refs_resolve", "a2a_submitted_artifact",
+#                "field_greater_than" (strict finite-number comparison against
+#                a literal threshold),
+#                "field_less_than" / "field_equals_context" /
+#                "field_in_context_array" (compare a field on this
+#                step's response against a value captured by a prior step — see
+#                "Cross-step comparison" section below),
+#                "upstream_traffic" (asserts side-effects via comply_test_controller's
+#                query_upstream_traffic scenario — see "upstream_traffic" section below),
+#                "canonical_format_satisfaction" (asserts create_media_buy package
+#                selectors against prior get_products format_options[] context — see
+#                "Canonical format satisfaction" section below))
+# severity: string (optional, default "required"): "required" | "advisory".
+#                A "required" validation failure fails the step and contributes
+#                to steps_failed; this is the default. An "advisory" failure
+#                surfaces in the validation_result with passed: false but does
+#                NOT fail the step — the step grades on its remaining required
+#                validations. Advisory failures contribute to a distinct
+#                `validations_advisory_failed` counter on run_summary so they
+#                stay visible without polluting conformance verdicts.
+#
+#                Use case: a storyboard declares a check whose runner support
+#                is rolling out (e.g., upstream_traffic during the months
+#                between the contract spec landing and the @adcp/sdk runner
+#                shipping it). Authors flip severity: advisory while
+#                instrumentation matures, then drop the field (or set
+#                "required") once adoption is stable. Without this escape
+#                valve, the choice was binary: declare the check and break
+#                conformance for adopters who haven't caught up, or omit it
+#                entirely and lose the validation surface during rollout.
+#                Distinct from the runtime forward-compat default (unknown
+#                check kinds → not_applicable): forward-compat handles
+#                runner-spec version skew; severity handles author-managed
+#                rollout gating.
+# expires_after_version: string (optional, only meaningful when severity:
+#                advisory): semver gating runtime promotion of this advisory
+#                to "required". Closes the advisory-drift problem:
+#                storyboards declared with severity: advisory during an
+#                adoption window get force-promoted once runners catch up,
+#                so authors don't need to track every advisory entry across
+#                every PR.
+#
+#                Versioning anchor: the comparison is against the runner's
+#                self-declared `runner_capability_version` (see
+#                runner-output-contract.yaml > run_summary), NOT against
+#                any specific implementation's package version. The runner
+#                chooses what its capability version corresponds to —
+#                @adcp/sdk semver, the AdCP spec version it implements, or
+#                a runner-binary version — and declares it in run_summary.
+#                Storyboard authors target the spec capability they need,
+#                not a specific SDK package; runners self-report the
+#                capability they offer.
+#
+#                Runner behavior: at storyboard-load time the runner
+#                compares its own runner_capability_version to
+#                expires_after_version using semver.gte. If
+#                runner_capability_version >= expires_after_version, the
+#                runner MUST promote severity to "required" — advisory
+#                failures then fail the step. The validation_result MUST
+#                carry severity_promoted_from_advisory: true to surface the
+#                promotion on rendered reports.
+#
+#                Forward-compat ordering: when a storyboard declares both
+#                an unknown check kind (graded `not_applicable` per #3816's
+#                forward-compat clause) AND severity: advisory +
+#                expires_after_version, forward-compat resolves FIRST. The
+#                check grades not_applicable and severity_promoted_from_advisory
+#                MUST be absent — no promotion was evaluated because no
+#                grading occurred. Promotion only applies to checks the
+#                runner actually grades.
+#
+#                Pre-release tags: comparison uses standard semver including
+#                pre-releases (RFC §11). `6.5.0-rc.3 < 6.5.0`. Authors
+#                wanting "after 6.5.0 stable" write `6.5.0`; authors
+#                wanting "after any 6.5.0 pre-release" write `6.5.0-0`.
+#                The runner MUST use semver.gte semantics, not a custom
+#                comparator. Authors targeting a specific minor without
+#                regard to pre-release status SHOULD write the lowest
+#                pre-release form (`X.Y.0-0`) when they want to include
+#                early adopters running RC builds.
+#
+#                Permanent advisories — see `permanent_advisory` field
+#                below for declaring an advisory that's deliberately not
+#                gated on runner version.
+#
+#                Format: a semver-valid string. The build-time lint at
+#                scripts/lint-storyboard-advisory-expiry.cjs MUST validate
+#                the value via Node's `semver.valid()` and reject malformed
+#                values; this prevents typos AND bounds the trust surface
+#                on the value when it's interpolated into rendered reports.
+#
+# permanent_advisory: object (optional, only meaningful when severity:
+#                advisory and expires_after_version is absent): structured
+#                marker declaring this advisory is deliberately not gated
+#                on runner version. Use case: experimental signals where
+#                the spec authors explicitly want the advisory grade to
+#                persist regardless of runner adoption.
+#
+#                Shape:
+#                  permanent_advisory:
+#                    reason: "<text>"
+#
+#                Replaces the earlier `# advisory-permanent: <reason>` YAML
+#                comment marker (which broke under YAML round-tripping in
+#                editor formatters and downstream tooling). Structured
+#                field is greppable, schema-validatable, and the reason is
+#                machine-readable for dashboards.
+#
+#                Schema requirement: when severity: advisory is declared,
+#                exactly one of `expires_after_version` or
+#                `permanent_advisory` MUST be present. The build-time lint
+#                surfaces violations (warning, not error — drift is a
+#                judgment call).
+# path: string (JSON path to the field, e.g., "formats[0].format_id")
+# value: any (expected value — string, number, or boolean; required when check is "field_value";
+#             accepted (optional) when check is "field_value_or_absent";
+#             required on "field_greater_than"; accepted (optional) on
+#             "field_less_than" as a literal comparand when
+#             `context_key` is not set; a non-negative integer exact count for
+#             "array_length" when range bounds are not used)
+# min: integer (optional inclusive lower bound for "array_length"; non-negative)
+# max: integer (optional inclusive upper bound for "array_length"; non-negative)
+# allowed_values: array (acceptable values for "field_value", "field_value_or_absent",
+#                        "field_contains", "http_status_in", "error_code", "any_of")
+# context_key: string (only meaningful for "field_less_than" /
+#                      "field_equals_context" / "field_in_context_array"
+#                      cross-step comparison checks — names a key populated by a prior
+#                      step's `context_outputs`. Required on
+#                      "field_equals_context" and "field_in_context_array";
+#                      optional on "field_less_than" (falls back to the literal `value`
+#                      when omitted). Ignored on every other check kind. When set and
+#                      the key is absent from the storyboard accumulator, the check
+#                      passes with a `context_key_absent` observation — see
+#                      "Cross-step comparison" section below.)
+# description: string (human-readable description of validation)
+#
+# field_value_or_absent check:
+#   Passes when the field is absent OR present and equal to `value` / contained in
+#   `allowed_values`; fails only when the field is present with a disallowed value.
+#   Use for fields that the spec permits to be omitted but MUST NOT carry a wrong value
+#
+# field_greater_than check:
+#   Requires the value at `path` and the literal `value` threshold to be finite
+#   numbers, and passes only when the observed value is strictly greater.
+#   when present (e.g., a `replayed` flag that MAY be omitted on a fresh path but MUST
+#   NOT be true). Accepts both `value` (single expected value) and `allowed_values` (set
+#   membership); at least one of the two MUST be provided. A `field_value_or_absent` check
+#   that declares neither `value` nor `allowed_values` is a storyboard-load error; runners
+#   MUST reject the storyboard before execution begins.
+#   For fields that are required by the response schema, pair this check with a separate
+#   `check: field_present` to avoid silently accepting absent fields that are required.
+#
+# array_length check:
+#   Resolves `path` to one array and asserts its cardinality. Authors choose
+#   exactly one configuration: `value: N` for an exact count, or `min: N` /
+#   `max: N` for inclusive bounds (either bound may be omitted). Counts MUST be
+#   non-negative integers and `min` MUST NOT exceed `max`. A missing path or a
+#   non-array value fails; this check has no optional-field arm. Wildcard paths
+#   are not supported. Failure output uses the array path as `json_pointer`, the
+#   requested count or bounds as `expected`, and the observed cardinality as
+#   `actual`. For a missing/non-array path, failure output instead uses
+#   `expected: "array"` and the observed value as `actual` (null when absent).
+#
+# OAuth discovery graph — use `check: oauth_metadata_graph`:
+#   Applies only to a `protected_resource_metadata` runner task. Starting from
+#   the RFC 9728 document returned by that task, the runner validates the full
+#   public OAuth discovery graph described by `test-kits/oauth-setup-runner.yaml`:
+#   the protected-resource identifier, a non-empty `authorization_servers`
+#   array, every referenced RFC 8414 document, resource equality under the
+#   shared `resource_equals_agent_url` rule (scheme and host lowercased,
+#   default ports elided, path case-sensitive), exact issuer string equality
+#   after JSON decoding (without URL normalization),
+#   conditional RFC 8414 required fields, HTTPS URL constraints, and passive
+#   reachability of advertised authorization, token, and JWKS endpoints.
+#
+#   The check follows EVERY authorization server named in a graph that is within
+#   the runner contract's fixed limits; an oversized graph fails with
+#   `oauth_graph_limit_exceeded` before unbounded traversal. A valid,
+#   explicitly named authorization server on another origin is conformant;
+#   origin equality with the protected resource MUST NOT be imposed. Every URL
+#   is independently subject to the runner contract's SSRF, redirect, size, and
+#   timeout policy before it is fetched.
+#
+#   `registration_endpoint` is syntax-checked when advertised, but this check
+#   MUST NOT submit an RFC 7591 registration request. Active registration has
+#   credential-lifecycle and cleanup semantics and belongs to a separately
+#   capability-gated future grade.
+#
+#   Failure output MUST use one of the stable diagnostic codes in the test-kit
+#   contract and include the offending URL/issuer/field when available. The
+#   deterministic fixtures at `test-vectors/oauth-setup/vectors.json` are the
+#   conformance corpus for runner implementations of this check.
+#
+# Error-shape validation — use `check: error_code`:
+#   The `error_code` check is shape-agnostic. The runner resolves the code from
+#   any of the transport-binding locations defined by the client detection
+#   order (see docs/building/implementation/transport-errors#client-detection-order):
+#   - `adcp_error.code` (MCP structuredContent, A2A artifact DataPart, JSON-RPC
+#     error.data, MCP text-fallback content JSON)
+#   - `errors[0].code` (task-payload errors array, top-level or under payload)
+#   A storyboard SHOULD assert error shape via `check: error_code` rather than
+#   `check: field_present, path: "errors"` or `path: "adcp_error"` — pinning to
+#   a specific shape makes the validator flakey against conformant agents that
+#   surface errors on the other layer.
+#
+# Error-code vocabulary:
+#   Every code referenced in a `value:` or `allowed_values:` under `check: error_code`
+#   MUST exist in the canonical enum at `static/schemas/source/enums/error-code.json`
+#   (or be registered as a deprecation alias). The `lint:error-codes` script
+#   enforces this at build time — references to undefined codes fail the build.
+#   Sellers MAY emit vendor-specific codes outside the enum, but storyboards
+#   MUST NOT assert on them (that couples the conformance suite to a specific
+#   vendor's taxonomy). When a test needs to accept multiple possible codes,
+#   use `allowed_values: [...]` rather than a single-code `value:` assertion.
+#
+# Canonical format satisfaction — use `check: canonical_format_satisfaction`:
+#   Applies only to create_media_buy steps. The runner compares each package's
+#   format selector against the target product captured from prior get_products
+#   context. When multiple get_products responses exist, the runner MUST use
+#   the most recent prior get_products response containing the package's
+#   product_id; if no such response exists, the check fails as an authoring or
+#   setup error. `value: true` means the request should be accepted; `value:
+#   false` means the request should be rejected for a format-selector cause.
+#
+#   The check evaluates each present selector route independently before
+#   applying precedence:
+#   1. Resolve `format_option_refs[]`, normalize legacy `format_ids[]` through
+#      `v1_format_ref[]` and the v1-to-canonical registry, and materialize the
+#      direct `format_kind`/`params` declaration. If any option reference is
+#      unresolved or any legacy ID is unprojectable, local_satisfied is false;
+#      a negative step asserting this path pairs the check with an
+#      `error_code: UNSUPPORTED_FEATURE` validation.
+#   2. For every resolved route, derive the target product's selected
+#      `format_options[]` entries using directional product gating for fixed
+#      dimensions, duration declarations, canonical params, and range
+#      containment. Legacy compatibility uses the asymmetric v2-narrows-v1
+#      relation from canonical-formats.mdx, not raw declaration equality.
+#   3. When multiple routes are present, their selected product-option sets
+#      MUST match. Different format shapes, option sets, or incompatible
+#      dimensions make local_satisfied false; a negative step asserting this
+#      path pairs the check with `error_code: CONFLICTING_SELECTORS`.
+#   4. Only after every route resolves and the selected sets agree does the
+#      runner apply package precedence: format_option_refs, then direct
+#      format_kind/params, then format_ids.
+#
+#   Negative cases only pass when the observed rejection is attributable to
+#   formats, so unrelated auth or tenant failures do not mask selector bugs.
+#
+#   Fields:
+#     check: canonical_format_satisfaction
+#     value: boolean                # true = expected accepted; false = expected rejected.
+#     path: string                  # optional; selects one package object or package array.
+#                                   # default is the request payload's packages[] array.
+#
+#   Example:
+#     - check: canonical_format_satisfaction
+#       value: false
+#       description: "Bare image selector lacks the dimensions required by the fixed product"
+#
+# Cross-step integrity — use `check: refs_resolve`:
+#   `refs_resolve` asserts every ref in a source set resolves to a member of a
+#   target set, matched on declared keys. Use it when one step's response
+#   contains references that MUST exist in another step's response. In 3.2,
+#   prefer direct canonical declarations and use this primitive for remaining
+#   cross-resource identifiers.
+#   Without this primitive, broken references are silent until a later call
+#   (e.g., `sync_creatives`) fails at runtime, after commitments are already
+#   made.
+#
+#   Fields:
+#     source:
+#       from: current_step | context
+#           # `current_step` reads the step's task-result data;
+#           # `context` reads values captured by prior steps.
+#       path: <string>              # supports `[*]` wildcards — see below.
+#     target:
+#       from: current_step | context
+#       path: <string>
+#     match_keys: [<key>, ...]      # keys compared on each ref — e.g. [agent_url, id].
+#                                   # A ref missing any declared key is NEVER a match
+#                                   # (agents that drop a key don't fuzzy-match others
+#                                   # that also dropped it).
+#     scope:                        # optional — restrict integrity to in-scope refs.
+#       key: <string>               # e.g. `agent_url`.
+#       equals: <string>            # literal value, or `$agent_url` for the
+#                                   # runner target URL. Keys ending in `url`
+#                                   # get trailing-slash / case normalization
+#                                   # on both sides before compare. Transport
+#                                   # path canonicalization (`/mcp`, well-known
+#                                   # A2A card path) for `$agent_url` is tracked
+#                                   # in adcp-client#710 — until it lands, MCP
+#                                   # agents should expect refs to fall
+#                                   # out-of-scope when format_ids name the
+#                                   # bare agent URL.
+#     on_out_of_scope: warn | ignore | fail
+#                                   # how refs outside `scope` are graded.
+#                                   # default `warn`: pass the check, attach
+#                                   # observations naming the skipped refs.
+#                                   # `ignore`: silent. `fail`: promote to
+#                                   # missing so reports name them.
+#
+#   Path wildcards:
+#     `[*]` in a `source.path` or `target.path` flattens over arrays. So
+#     `products[*].format_ids[*]` walks every product and flattens every
+#     `format_ids` array into a single list of refs. Runners cap terminal
+#     fan-out at 10,000 values to bound malicious agent responses; paths
+#     realistic for any catalog size are well under the cap.
+#
+#   Grading output:
+#     A failing `refs_resolve` check names the specific unresolved ref tuples
+#     in `actual.missing` (projected to `match_keys`). Duplicates are collapsed
+#     on the projected tuple so one broken ref across 50 products shows up
+#     once — compliance reports stay readable.
+#
+#   Legacy 3.x example — every named format reference resolves:
+#     - check: refs_resolve
+#       description: "Every deprecated format_id resolves in the compatibility catalog"
+#       source:
+#         from: context
+#         path: "products[*].format_ids[*]"
+#       target:
+#         from: current_step
+#         path: "formats[*].format_id"
+#       match_keys: [agent_url, id]
+#       scope:
+#         key: agent_url
+#         equals: $agent_url
+#       on_out_of_scope: warn
+#
+# Cross-step comparison — use `check: field_less_than` /
+# `field_equals_context` / `field_in_context_array` /
+# `all_fields_in_context_array`:
+#   These checks compare a field in THIS step's response against a value
+#   captured by a PRIOR step via `context_outputs`. Without them, every existing
+#   `check:` is single-step (`field_value` resolves the comparand at YAML
+#   authoring time via `$context.<name>` substitution; the runtime value the
+#   prior step actually returned is opaque to the validation).
+#
+#   Use cases:
+#     - Round-trip identity: assert that `media_buys[0].media_buy_id` on
+#       `get_media_buys` equals the id returned by the earlier `create_media_buy`
+#       step. Today's `field_value` + `$context.media_buy_id` substitution
+#       resolves the placeholder against the request, not the response — a
+#       seller could return a different id on the read and still pass.
+#     - Filtered-vs-baseline: assert that a filtered `get_rights` returns a
+#       strictly smaller `rights.length` than the unfiltered baseline call.
+#     - Idempotency replay: assert that the replayed response's body hash
+#       equals the originally captured hash (byte-for-byte cache hit).
+#     - Capability membership: assert that an applied account billing value is
+#       a member of the supported_billing array captured during discovery.
+#
+#   Fields:
+#     check: field_less_than | field_equals_context | field_in_context_array | all_fields_in_context_array
+#     path: string                  # JSON path on this step's response.
+#     context_key: string           # name set by a prior step's
+#                                   # `context_outputs[].name`. Required for
+#                                   # `field_equals_context` and
+#                                   # `field_in_context_array` and
+#                                   # `all_fields_in_context_array`. Optional for
+#                                   # `field_less_than` — when omitted, the
+#                                   # check falls back to the literal `value`
+#                                   # (so authors can compare against a
+#                                   # static threshold without context).
+#     value: number                 # Literal comparand for `field_less_than`
+#                                   # when `context_key` is not set. Ignored on
+#                                   # `field_equals_context` and
+#                                   # membership checks (which always
+#                                   # read from context).
+#     description: string
+#
+#   Type semantics:
+#     `field_less_than` requires both operands to be finite numbers; non-numeric
+#     or absent values fail the check with a type error in the validation_result.
+#     `field_equals_context` deep-equals the response value against the captured
+#     comparand using the same equality semantics as `field_value`.
+#     `field_in_context_array` requires the captured comparand to be an array and
+#     a scalar path; it passes when the response value deep-equals one member.
+#     `all_fields_in_context_array` requires an array comparand and a wildcard
+#     path; it passes only when EVERY resolved terminal value belongs to the
+#     captured array. A path resolving no terminal values passes vacuously; pair
+#     it with field_present when an empty response is not allowed. Older runners
+#     that do not implement this additive check grade it not_applicable under the
+#     runner forward-compatibility contract rather than producing a false failure.
+#
+#   Absent context_key — context_key_absent observation:
+#     When `context_key` is set but the named entry is missing from the
+#     storyboard accumulator (typically because the prior capturing step was
+#     legitimately skipped on a branch-set path), the check PASSES with a
+#     `context_key_absent` observation rather than failing. This preserves the
+#     "skipped prerequisite cascades to skip, not fail" semantics that
+#     `context_outputs` already establishes for `$context.<name>` substitution.
+#     A genuinely missing capture (the prior step ran but its capture path did
+#     not resolve) is graded `capture_path_not_resolvable` ON THE PRIOR STEP —
+#     downstream cross-step checks on a never-populated key are not the place
+#     to surface that failure.
+#
+#   Examples:
+#     # Round-trip: get_media_buys returns the id we just created.
+#     # Pair with `context_outputs` on the earlier create_media_buy step:
+#     #   context_outputs:
+#     #     - name: created_media_buy_id
+#     #       path: "media_buy_id"
+#     - check: field_equals_context
+#       path: "media_buys[0].media_buy_id"
+#       context_key: "created_media_buy_id"
+#       description: "get_media_buys echoes the id returned by create_media_buy"
+#
+#     # Filtered call returns fewer rights than baseline.
+#     # Capture the baseline length on the unfiltered step:
+#     #   context_outputs:
+#     #     - name: baseline_rights_count
+#     #       path: "rights.length"
+#     - check: field_less_than
+#       path: "rights.length"
+#       context_key: "baseline_rights_count"
+#       description: "Filtered rights array is smaller than unfiltered baseline"
+#
+#     # Idempotency replay: replayed body deep-equals the originally-captured one.
+#     # Pair with `context_outputs` on the original call:
+#     #   context_outputs:
+#     #     - name: original_create_response
+#     #       path: ""              # capture the entire response body
+#     - check: field_equals_context
+#       path: ""
+#       context_key: "original_create_response"
+#       description: "Replay returned the cached response body, deep-equal to the original"
+#
+#     # Capability membership: returned billing is one advertised value.
+#     - check: field_in_context_array
+#       path: "accounts[0].billing"
+#       context_key: "supported_billing"
+#       description: "Applied billing belongs to the advertised set"
+#
+#     # Exhaustive routing scope: every returned product channel was advertised.
+#     - check: all_fields_in_context_array
+#       path: "products[*].channels[*]"
+#       context_key: "portfolio_channels"
+#       description: "Returned product channels stay inside portfolio scope"
+#
+#     # Static threshold (single-step form — no context_key).
+#     # NOTE: `field_less_than` is primarily a cross-step check; the literal-
+#     # `value` form is a convenience for numeric thresholds. Authors who
+#     # need same-step equality should reach for `field_value` instead.
+#     - check: field_less_than
+#       path: "delivery.spend"
+#       value: 10000
+#       description: "Spend stays under the buy's budget cap"
+#
+# A2A wire-shape — use `check: a2a_submitted_artifact`:
+#   Asserts the A2A envelope invariants for AdCP `submitted` arms: `Task.id` /
+#   `Task.contextId` non-empty, `Task.state` normalised to `completed` (A2A 0.3
+#   wire value `"completed"`; A2A 1.0 enum `"TASK_STATE_COMPLETED"`), and the
+#   last non-null, object-typed DataPart's `data.status === 'submitted'`
+#   preserving the AdCP discriminator. Grades `not_applicable` on non-A2A
+#   transports so storyboards can include it alongside MCP-shape assertions
+#   without forking by transport.
+#
+# Upstream side-effects — use `check: upstream_traffic`:
+#   Asserts that the agent caused observable upstream traffic with a payload
+#   carrying the storyboard-supplied identifiers. The mechanism raises the bar
+#   against UNINTENTIONAL façades — adapters (often LLM-generated) that satisfy
+#   AdCP schema requirements with synthetic placeholders without forwarding the
+#   payload upstream. It is NOT an adversarial integrity check: adopters
+#   self-report their own traffic, and a determined façade could fabricate
+#   recorded_calls. Spec consumers MUST NOT treat upstream_traffic passing as
+#   cryptographic proof of adapter behavior; the value is "raises the bar
+#   against accidental façades," not "stops malicious façades."
+#
+#   The runner queries the adopter's comply_test_controller with scenario
+#   `query_upstream_traffic`, scoped to traffic recorded since the storyboard
+#   step's request timestamp, and applies the assertion declared on the
+#   storyboard step. Adopters who do not advertise `query_upstream_traffic` in
+#   `list_scenarios` grade the check `not_applicable` — the contract is opt-in
+#   by adopter capability. Adopters who DO advertise it but return zero
+#   recorded calls during the assertion window grade as failed (the
+#   "controller present and observed nothing" path is the façade signal).
+#
+#   Runner attestation-mode selection (normative):
+#     Conforming runners MUST choose params.attestation_mode for
+#     comply_test_controller query_upstream_traffic calls in a way that
+#     preserves assertion coverage:
+#       1. If any upstream_traffic validation in the current step sets
+#          attestation_mode_required: "raw", request "raw".
+#       2. Else if any upstream_traffic validation in the current step declares
+#          payload_must_contain, request "raw"; arbitrary payload-path
+#          assertions cannot be evaluated from digest attestations.
+#       3. Else if every upstream_traffic validation in the current step can
+#          be evaluated from count/endpoint/purpose plus identifier_paths, and
+#          the runner can supply params.identifier_value_digests for all
+#          resolved identifier values, the runner SHOULD request "digest" when
+#          it knows the matching upstream calls are JSON-shaped
+#          (`application/json` or `*/*+json`). If the runner does not know the
+#          matching calls are JSON-shaped, request "raw"; digest-mode
+#          identifier_match_proofs are intentionally unavailable for non-JSON
+#          payloads, so an unconditional digest request can turn an otherwise
+#          evaluable identifier_paths assertion into not_applicable.
+#       4. Otherwise, request "raw".
+#     Controllers MAY still return digest-mode attestations for a raw request
+#     when local policy forbids raw payload disclosure; raw-required and
+#     payload_must_contain assertions then grade not_applicable as described
+#     below.
+#     Storyboard authors SHOULD NOT invent non-schema attestation-mode hints.
+#     Use attestation_mode_required: "raw" only for hard raw requirements;
+#     otherwise rely on this runner selection rule.
+#
+#   Fields:
+#     min_count: integer (default: 1)
+#                              # Minimum number of recorded calls that match
+#                              # the optional `endpoint_pattern` filter. Use 0
+#                              # only for negative assertions (e.g., a step
+#                              # MUST NOT cause upstream traffic).
+#     endpoint_pattern: string (optional)
+#                              # Glob pattern matched against
+#                              # recorded_calls[].endpoint. Examples:
+#                              #   "POST */audience/upload"
+#                              #   "POST *"           (any POST, useful when
+#                              #                      the storyboard stays
+#                              #                      platform-agnostic)
+#                              # When omitted, all recorded calls in the window
+#                              # match. Storyboards in the shared corpus SHOULD
+#                              # stay platform-agnostic — host-specific patterns
+#                              # like "POST api.tiktok.com/*" couple the spec
+#                              # to specific upstream platforms and break for
+#                              # adopters whose upstream is something else.
+#                              # Reserve host-specific patterns for
+#                              # adapter-specific extensions, not shared
+#                              # storyboards.
+#     payload_must_contain: array of MatchSpec (optional)
+#                              # Each entry asserts a path/value pair that MUST
+#                              # appear in at least one matching call's payload.
+#                              # Path syntax: JSONPath-lite positional path
+#                              # with `[*]` array wildcards. Shared storyboards
+#                              # SHOULD use the bare dotted form. Examples:
+#                              #   "users[*].hashed_email"
+#                              #   "data.audience.members[*].id"
+#                              # The literal `[*]` wildcard flattens over arrays
+#                              # and passes if any element matches. The runner
+#                              # also tolerates optional `$` / `$.` prefixes and
+#                              # standalone numeric index tokens for SDK
+#                              # compatibility (for example `items.[0].id`);
+#                              # shared storyboards should avoid indexes unless
+#                              # a position-specific assertion is intentional.
+#                              # RFC 9535 JSONPath descendant syntax (`$..foo`) is NOT
+#                              # supported — the runner uses
+#                              # JSONPath-lite matching from @adcp/sdk. Path matching only
+#                              # applies when recorded_calls[].content_type is
+#                              # JSON-shaped — defined as `application/json`
+#                              # or any structured-syntax-suffix `*/*+json`
+#                              # per RFC 6839 §3.1 (e.g., `application/vnd.api+json`,
+#                              # `application/scim+json`). All other content
+#                              # types (form-urlencoded, multipart, plain text,
+#                              # newline-delimited JSON like `application/json-seq`,
+#                              # JSON Lines `application/jsonl`) take the non-
+#                              # JSON path. Against non-JSON ALL match modes
+#                              # (`present` / `equals` / `contains_any`) grade
+#                              # `not_applicable`. Earlier sketches downgraded
+#                              # `match: present` to a terminal-key substring
+#                              # search of the raw payload string; that
+#                              # heuristic creates false positives (a payload
+#                              # mentioning the key name in any context — URL
+#                              # fragment, comment, unrelated metadata field —
+#                              # would pass), which is exactly the anti-façade
+#                              # contract this surface exists to protect.
+#                              # Storyboards that need a "the upstream call
+#                              # carried this value" signal against non-JSON
+#                              # payloads use `identifier_paths` instead — that
+#                              # surface substring-searches storyboard-supplied
+#                              # VALUES (not path-derived strings), which is
+#                              # encoding-agnostic and doesn't suffer the
+#                              # false-positive surface.
+#                              # Shape:
+#                              #   - path: <dotted-with-[*] path>
+#                              #     match: present | equals | contains_any
+#                              #     value: <value>            # for equals
+#                              #     allowed_values: [...]     # for contains_any
+#                              # Example:
+#                              #   - path: "users[*].hashed_email"
+#                              #     match: present
+#     identifier_paths: array of string (optional)
+#                              # Paths into the storyboard's sample_request that
+#                              # name the load-bearing identifiers the adapter
+#                              # MUST forward upstream. The runner extracts the
+#                              # values at these paths from the storyboard's
+#                              # request and asserts each value appears in at
+#                              # least one matching recorded_call's payload at
+#                              # any depth. This is the strongest single
+#                              # anti-façade signal — adapters cannot satisfy
+#                              # it by fabricating a constant placeholder hash,
+#                              # because the placeholder will not match the
+#                              # storyboard's supplied value.
+#                              # Portable path grammar (normative):
+#                              # request-payload-relative dotted paths where
+#                              # each segment is an identifier key optionally
+#                              # followed by one `[*]` array wildcard. Example:
+#                              #   "audiences[*].add[*].hashed_email"
+#                              # Supported segment form:
+#                              #   [A-Za-z_][A-Za-z0-9_-]* or
+#                              #   [A-Za-z_][A-Za-z0-9_-]*[*]
+#                              # Unsupported forms MUST be rejected at
+#                              # storyboard load/lint time rather than treated
+#                              # as paths that resolve to zero values:
+#                              # bracket-quoted keys (`foo["bar"]`), numeric
+#                              # indexes (`items[0]`), recursive descent
+#                              # (`$..foo`), absolute roots (`$.foo`), empty
+#                              # segments, and reserved roots (`request.*`,
+#                              # `response.*`, `context.*`). The path is
+#                              # already relative to the storyboard step's
+#                              # sample_request; adding an explicit root makes
+#                              # the same storyboard ambiguous across runners.
+#                              # Each path MAY resolve to a single value or an
+#                              # array; all resolved values MUST be present in
+#                              # the recorded payload. Replaces the earlier
+#                              # `buyer_identifier_echo: boolean` shorthand
+#                              # (which left "what counts as an identifier" to
+#                              # convention) with an explicit enumeration.
+#                              # Example:
+#                              #   identifier_paths:
+#                              #     - "audiences[*].add[*].hashed_email"
+#                              #     - "audiences[*].add[*].external_id"
+#     purpose_filter: array of string (optional)
+#                              # Restrict the assertion to recorded_calls whose
+#                              # `purpose` field matches one of the listed
+#                              # values. Allowed values:
+#                              #   "platform_primary" | "measurement" |
+#                              #   "attribution" | "creative_serving" |
+#                              #   "identity" | "other"
+#                              # Use this when an adapter legitimately makes
+#                              # multiple kinds of upstream calls during a step
+#                              # — e.g., a sales-non-guaranteed buyer agent
+#                              # creating a campaign AND calling a measurement
+#                              # vendor AND posting an attribution event during
+#                              # the same create_media_buy. The storyboard
+#                              # scopes to `purpose_filter: [platform_primary]`
+#                              # so the other calls don't muddy the
+#                              # campaign-creation assertion.
+#                              # When omitted, the assertion considers all
+#                              # recorded_calls regardless of purpose. Calls
+#                              # without a `purpose` field (adopter didn't
+#                              # classify) are treated as `purpose: other` for
+#                              # filter matching — they're included by filters
+#                              # listing `other` and excluded by filters that
+#                              # don't. Runners MUST include a count of
+#                              # unclassified calls in the validation_result's
+#                              # `actual` when `purpose_filter` is set and
+#                              # zero recorded_calls match the filter, so a
+#                              # façade that misclassifies-into-`other` to
+#                              # dodge filtering produces a noisy zero rather
+#                              # than a silent one.
+#     since: prior_step_id (optional, default: this step's request timestamp)
+#                              # Bound the lookup window. By default, the
+#                              # runner queries traffic recorded since the
+#                              # storyboard step issued its AdCP request, so
+#                              # only side-effects of THIS step are counted.
+#                              # Set to a prior step's id to include traffic
+#                              # caused by an earlier step (e.g., for a
+#                              # cumulative-effect assertion). The runner uses
+#                              # its own wall-clock timestamp from when it
+#                              # issued the AdCP request as the lower bound —
+#                              # the controller's `since_timestamp` echo is
+#                              # informational, not authoritative. See
+#                              # runner-output-contract.yaml > validation_result
+#                              # for the timestamp boundary semantics
+#                              # (inclusive lower bound, clock-skew tolerance).
+#     attestation_mode_required: string (optional)
+#                              # When set to "raw", the storyboard requires
+#                              # raw payload introspection. Recorded calls
+#                              # returned in `attestation_mode: digest` grade
+#                              # the upstream_traffic check `not_applicable`
+#                              # rather than failing — the storyboard is
+#                              # asserting against payload contents the
+#                              # adopter's privacy policy doesn't permit
+#                              # disclosing.
+#                              # When omitted, the storyboard accepts either
+#                              # mode: assertions degrade gracefully in digest
+#                              # mode (`payload_must_contain` grades
+#                              # not_applicable per call; `identifier_paths`
+#                              # works via the digest-of-identifier echo
+#                              # mechanism, see "Digest-mode behavior" below).
+#                              # Use "raw" sparingly — the digest mode exists
+#                              # so EU/privacy-conscious adopters can support
+#                              # upstream_traffic conformance without returning
+#                              # hashed-PII payloads. Forcing raw excludes
+#                              # those adopters; only use it when the
+#                              # storyboard's assertion genuinely cannot be
+#                              # expressed in identifier-digest form.
+#
+#   Digest-mode behavior:
+#     `attestation_mode: digest` per recorded_call swaps payload introspection
+#     for digest-based echo verification (see comply-test-controller-response.json
+#     > UpstreamTrafficSuccess > recorded_calls > DigestAttestation):
+#       - `payload_must_contain` arbitrary path assertions: NOT supported in
+#         digest mode. Each entry grades `not_applicable` per affected call,
+#         with a note explaining the mode constraint.
+#       - JCS non-finite numbers: when a digest-mode upstream_traffic response
+#         cannot be produced because the parsed JSON-like value tree contains
+#         a non-finite numeric value (`NaN`, `+Infinity`, or `-Infinity`), the
+#         controller MUST NOT coerce that value to `null`, a string, or any
+#         other placeholder for digest computation; it returns ControllerError
+#         with error code `JCS_NON_FINITE_NUMBER`; the runner grades the
+#         affected validation `not_applicable`, not failed, with a note citing
+#         the RFC 8785/JCS constraint.
+#       - `payload_length`: MUST equal the exact byte length covered by
+#         `payload_digest_sha256` — RFC 8785 (JCS) canonical bytes for
+#         JSON-shaped content after redaction, or post-redaction raw body bytes
+#         for non-JSON content. It is not the original outbound body length
+#         before JSON parsing, redaction, or canonicalization.
+#       - `identifier_paths` echo verification: supported via the controller's
+#         `identifier_match_proofs[]`. The runner computes SHA-256 of each
+#         resolved identifier value, sends the digests in
+#         `params.identifier_value_digests`, and reads the per-digest
+#         `found: true | false` proof from the response. Plaintext identifiers
+#         never reach the controller; the controller never returns plaintext
+#         payloads. The privacy boundary holds in both directions.
+#       - `min_count` and `endpoint_pattern`: unchanged — count and URL
+#         matching work identically in both modes.
+#
+#   Grading output (see runner-output-contract.yaml > validation_result):
+#     `expected` echoes the declared assertion fields. `actual` is a
+#     diagnostic summary: { matched_count, total_calls, missing_payload_paths,
+#     missing_identifier_values }. The full recorded_calls array lives in
+#     `response.payload` so an implementor can see the exact upstream
+#     payloads the runner observed.
+#
+#   Adopter integration:
+#     The adopter implements `query_upstream_traffic` in its
+#     comply_test_controller. The controller records outbound HTTP calls in a
+#     caller-scoped buffer (sandbox-only — production builds MUST 404 the
+#     comply_test_controller tool entirely; multi-tenant sandboxes MUST key
+#     the buffer on the calling principal so cross-caller traffic does not
+#     leak). See comply-test-controller-request.json > query_upstream_traffic
+#     for the controller-side query contract.
+#
+#   Example — sync_audiences must POST hashed members upstream:
+#     - check: upstream_traffic
+#       description: "sync_audiences caused upstream traffic with hashed members"
+#       endpoint_pattern: "POST *"
+#       min_count: 1
+#       payload_must_contain:
+#         - path: "users[*].hashed_email"
+#           match: present
+#       identifier_paths:
+#         - "audiences[*].add[*].hashed_email"
+#         - "audiences[*].add[*].external_id"
+#
+# Cross-response assertions — `check: cross_response_field_equal` / `cross_response_count_distinct`:
+#   These check kinds operate on the RESPONSE SET of a step that dispatches
+#   multiple parallel requests under the `parallel_dispatch_runner` contract
+#   (see test-kits/parallel-dispatch-runner.yaml). They are NOT applicable to
+#   single-dispatch steps: a single response trivially has cardinality 1 on any
+#   field. Runners MUST gate these checks on the step declaring
+#   `requires_contract: parallel_dispatch_runner` and a `parallel_dispatch`
+#   block; outside that gate the check MUST grade `not_applicable`. The
+#   canonical use case is rule-9 testing (concurrent retries / first-insert-wins
+#   from L1/security.mdx#idempotency) — fire N parallel calls with the same
+#   `idempotency_key`, assert exactly one resource was created.
+#
+#   `parallel_dispatch` is a step-level block that travels alongside `requires_contract`:
+#
+#     - id: concurrent_create
+#       task: create_media_buy
+#       requires_contract: parallel_dispatch_runner
+#       parallel_dispatch:
+#         count: 2                    # required, integer 2..10. The runner fires
+#                                     # `count` logically-simultaneous requests
+#                                     # using the SDK's batch primitive.
+#         same_idempotency_key: true  # required for rule-9 tests; future
+#                                     # non-idempotency parallel tests MAY set
+#                                     # this false.
+#         barrier_timeout_ms: 5000    # optional; default 5000 — how long the
+#                                     # runner waits at the dispatch barrier
+#                                     # before firing requests anyway.
+#       validations:
+#         - check: cross_response_count_distinct
+#           path: "media_buy_id"
+#           allowed_values: [1]
+#           description: "Exactly one media_buy_id across all dispatches"
+#         - check: cross_response_field_equal
+#           path: "media_buy_id"
+#           description: "All dispatches converge on the same media_buy_id"
+#
+#   `cross_response_field_equal`:
+#     - path (string, required): JSON path checked on every resolved response.
+#       All non-null observed values at `path` MUST be deeply equal. A missing
+#       value on a response is treated as a non-equal observation and fails the
+#       check.
+#
+#   `cross_response_count_distinct`:
+#     - path (string, required): JSON path scanned on every resolved response.
+#     - allowed_values (array of integers, required): permitted cardinalities of
+#       distinct values observed at `path` across the resolved response set.
+#       For rule-9 tests, `allowed_values: [1]` — exactly one resource was
+#       created across all dispatches.
+#
+#   Resolution semantics. A seller adopting the reject-and-redirect policy under
+#   rule 9 may return `IDEMPOTENCY_IN_FLIGHT` on the second (or later) dispatch
+#   rather than block. The runner MUST resolve `IDEMPOTENCY_IN_FLIGHT` responses
+#   to their eventual cached response before applying these checks: wait
+#   top-level `error.retry_after` seconds and retry with the SAME idempotency_key,
+#   exactly as the buyer-side normative behavior specifies. The cross-response
+#   check operates on the resolved response set, not the raw initial set. If
+#   resolution exceeds the runner's step-timeout budget, the step grades
+#   `step_timeout` (a runner-synthesized code), not a rule-9 violation.
+#
+#   Grading output. `expected` carries the declared assertion fields
+#   (`path`, plus `allowed_values` for `count_distinct`). `actual` summarizes
+#   what was observed across the resolved response set:
+#     cross_response_field_equal → { observed_values: [...] }
+#     cross_response_count_distinct → { observed_cardinality, observed_values: [...] }
+#   See runner-output-contract.yaml > validation_result > expected/actual for
+#   the canonical shapes.
+#
+# --- Runner output ---
+#
+# How a runner MUST report validation results — including which fields a
+# failing validation carries (exact request, response, JSON Pointer, expected
+# vs. actual, schema $id, schema URL) and how skipped storyboards MUST
+# distinguish not_applicable vs. no_phases vs. missing_tool — is defined in
+# runner-output-contract.yaml. Storyboard authors SHOULD assume failures will
+# be rendered with that detail and write descriptions accordingly.
+#
+# Runner grading codes introduced by the context accumulator and fixture
+# seeding:
+#
+#   capture_path_not_resolvable
+#       Emitted on the capturing step when a `context_outputs:` entry's
+#       `path` does not resolve to a usable value in the step's response body.
+#       The step grades as failed — the capture declared a contract that the
+#       response did not meet. Distinct from unresolved_substitution so tooling
+#       can discriminate "producer contract breach" from "consumer missing
+#       state."
+#
+#       "Does not resolve to a usable value" covers all three cases:
+#         1. Path structurally absent (e.g. `signals[0]` when array is empty)
+#         2. Path present but value is null
+#         3. Path present but value is "" (empty string)
+#       Capturing null or "" produces fabricated downstream state and is as
+#       incorrect as a missing path; runners MUST treat all three as failures.
+#
+#       Output shape: runner-output-contract.yaml > validation_result, with
+#       check: capture_path_not_resolvable, expected: the declared path string,
+#       actual: the resolved value (or null if absent), json_pointer: RFC 6901
+#       form of the path.
+#
+#   unresolved_substitution
+#       Emitted on a consumer step when a referenced $context.<name> or
+#       {{prior_step.<id>.<field>}} value is not populated at step-execution
+#       time (either because the producer step failed, was skipped, or never
+#       ran). The step grades as failed. Also used at preflight when a
+#       substitution is statically unresolvable (e.g., {{runner.webhook_url:*}}
+#       on a run with no webhook receiver) to grade the entire storyboard
+#       not_applicable before execution. A `$build_assets_from_format`
+#       directive whose valid captured format cannot be satisfied by the
+#       runner's test-kit assets is not an unresolved substitution: it uses the
+#       storyboard-level not_applicable disposition with `skip_result.reason:
+#       fixture_unavailable` at the earliest preflight after the referenced
+#       format is captured and before any intervening side-effecting step. The
+#       runner terminates the remaining storyboard without prerequisite
+#       cascades. This runner-owned coverage gap MUST NOT count as a failed step.
+#
+#       Output shape: runner-output-contract.yaml > validation_result, with
+#       check: unresolved_substitution, expected: the unresolved token string,
+#       actual: null, json_pointer: null (pre-wire failure — no response
+#       payload available).
+#
+#   fixture_seed_unsupported
+#       Emitted when a storyboard's `prerequisites.controller_seeding` requires
+#       a seed_* scenario the seller does not implement (returned as
+#       UNKNOWN_SCENARIO). The storyboard grades not_applicable (coverage gap),
+#       not failed.
+#
+#   unresolved_scenario_reference
+#       A detailed sub-reason under the canonical skip `reason:
+#       not_applicable`. Emitted when a parent storyboard's
+#       `requires_scenarios:` entry cannot be resolved against the source
+#       tree (no file declares that `id:`). Per the
+#       `detailed_reason_mapping` convention in runner-output-contract.yaml,
+#       runners MUST populate the canonical `reason: not_applicable` and
+#       encode `unresolved_scenario_reference` in `detail`. detail MUST
+#       follow the shape
+#       'requires_scenarios reference "<scenario_id>" did not resolve
+#       against the source tree' — and when multiple references are
+#       unresolved, detail MUST enumerate every unresolved id with the
+#       same shape (comma-separated or newline-delimited) so downstream
+#       tooling can parse each. Runners MUST NOT silently pass the parent
+#       as if the missing scenario had contributed.
+#
+#       Distinct from `fixture_seed_unsupported`: that reason indicates an
+#       agent-coverage gap (seller doesn't implement a seed scenario);
+#       `unresolved_scenario_reference` indicates a source-tree authoring
+#       bug that the build-time lint SHOULD have caught. Dashboards and
+#       summary reports SHOULD NOT conflate the two when aggregating
+#       `not_applicable` counts — they are categorically different
+#       signals.
+#
+#       The `lint:storyboard-branch-sets` script surfaces this as a
+#       build-time error via the `unresolved_scenario_reference` rule so
+#       unresolved references in a well-linted corpus never reach a
+#       runner. The runner-side grading is defensive for corpora that
+#       bypass the lint or come from pre-lint CI states.
+#
+# --- Webhook receiver (optional, for outbound-webhook conformance) ---
+#
+# Storyboards that verify outbound webhook conformance (signing, idempotency_key
+# presence, idempotency_key stability across retries) require the runner to host
+# a webhook receiver during test execution. The receiver URL is injected into
+# push_notification_config on steps that trigger webhooks, and a subsequent
+# expect_webhook* step asserts on what arrived.
+#
+# Receiver behavior is specified by a test-kit contract referenced from the
+# storyboard's prerequisites.test_kit field — see test-kits/webhook-receiver-runner.yaml
+# for the reference contract (endpoint modes, retry-replay shape, client-primitive
+# hooks). Storyboard authors MUST declare the contract in prerequisites; the
+# webhook-emission universal (universal/webhook-emission.yaml) is the reference
+# consumer, including idempotency replay side-effect assertions that need
+# webhook observation. Keep those assertions on webhook-specific storyboards so
+# core idempotency requests do not depend on runner-only URL templating.
+#
+# --- Substitution variables (webhook-receiver enabled) ---
+#
+# When a storyboard's test-kit contract declares a webhook_receiver, the runner
+# exposes these substitutions for use in sample_request payloads and expectations:
+#
+#   {{runner.webhook_base}}            → HTTPS base of the runner's receiver for this run
+#   {{runner.webhook_url:<step_id>}}   → per-step URL; unique per step
+#                                      The URL may contain a runner-internal
+#                                      capture id for receiver routing. That id
+#                                      is not the protocol operation_id. When a
+#                                      request includes
+#                                      push_notification_config.operation_id,
+#                                      the runner MUST make any URL capture id
+#                                      distinct so sellers that parse the URL
+#                                      fail a filter.body.operation_id payload
+#                                      assertion.
+#
+# Per-step receivers are the only supported shape in v1 so each step's
+# expect_webhook matches only its own deliveries. Fan-in scenarios are deferred
+# (see "shared_receiver" note below).
+#
+# --- Webhook-assertion steps ---
+#
+# The following step `task` values are only valid when the storyboard's test-kit
+# declares a webhook_receiver. They are scheduled alongside regular task steps
+# and reference earlier steps by id.
+#
+# task: expect_webhook
+#   Wait up to `timeout_seconds` (default 30) for a webhook delivered to the
+#   triggered receiver, validate the payload, then apply the filter. Runner
+#   delegates to the `@adcp/client` AsyncHandler; the webhook is observed via
+#   onActivity with type webhook_received, and WebhookMetadata (including
+#   idempotency_key) is provided by the client.
+#
+#   `triggered_by` scopes the assertion to the per-step receiver URL. Payload
+#   fields are matched through `filter.body` using dotted paths. Authors MUST
+#   use `filter.body.operation_id` to assert the protocol payload operation_id;
+#   `filter.operation_id` is reserved for runner receiver routing and is not the
+#   protocol correlation field.
+#
+#   Fields:
+#     triggered_by: <step_id>     # earlier step whose task triggered this webhook
+#     filter:
+#       body:                     # match by webhook payload fields, never URL path
+#         operation_id: "op_explicit_from_push_notification_config"
+#         status: "completed"     # optional; match any status if omitted
+#     timeout_seconds: 30         # optional; fail if no match within window
+#     expect_idempotency_key: true  # default true; assert the key is present and pattern-valid
+#     webhook_payload_schema_ref: <path>
+#                                 # optional — validate payload against the webhook schema
+#                                 # (distinct from step-level `schema_ref` and
+#                                 # `response_schema_ref`, which apply to the caller→agent
+#                                 # request/response. Webhook steps have no caller request,
+#                                 # so the payload schema is named explicitly to avoid
+#                                 # overload of schema_ref.)
+#     expect_max_deliveries_per_logical_event: 1
+#                                 # optional — if set, runner asserts at most N distinct
+#                                 # logical webhook events (grouped by idempotency_key)
+#                                 # arrive within the window. Used by idempotency storyboards
+#                                 # to catch duplicate-side-effect bugs where a seller
+#                                 # re-executes on replay with a fresh idempotency_key.
+#     requires_contract: <contract_id>
+#                                 # optional — step grades as not_applicable (not fail)
+#                                 # when the named test-kit contract is not in scope.
+#                                 # Used on cross-specialism assertions that depend on
+#                                 # webhook_receiver_runner being active.
+#
+#   Error modes: no_webhook_received, schema_violation, missing_idempotency_key,
+#   invalid_idempotency_key_format, signature_invalid (when signature assertion
+#   on), duplicate_webhook_on_replay (when
+#   expect_max_deliveries_per_logical_event exceeded).
+#
+# task: expect_no_webhook
+#   Observe the receiver associated with `triggered_by` for the full bounded
+#   window and pass only when zero deliveries match `filter`. The default
+#   `timeout_seconds` is 5. The assertion fails immediately on the first match
+#   with `unexpected_webhook_received`; it does not validate a payload schema or
+#   require an idempotency_key on the passing path because no payload exists.
+#
+#   Fields:
+#     triggered_by: <step_id>
+#     filter: { ... }             # same shape as expect_webhook
+#     timeout_seconds: 5          # optional; default 5
+#
+#   Error modes: unexpected_webhook_received.
+#
+# task: expect_webhook_retry_keys_stable
+#   The runner's webhook receiver deterministically returns 5xx for the first
+#   `retry_trigger.count` deliveries of the matching webhook, then 2xx. The
+#   runner asserts that every delivery within the retry window carries the
+#   byte-identical idempotency_key (the sender regenerating a key on retry is a
+#   conformance failure).
+#
+#   Signature verification on every delivery in the retry loop: when 9421 is in
+#   effect (i.e., push_notification_config was registered without `authentication`),
+#   the runner MUST verify the 9421 signature on EVERY delivery including retries
+#   and MUST reuse the run-scoped (keyid, nonce) replay store (see test-kit
+#   client_primitives.signature_replay_store). A publisher that stably reuses
+#   idempotency_key (correct) but also reuses the 9421 `nonce` sig-param
+#   (incorrect — nonce MUST be fresh per delivery) would pass retry-stability
+#   but MUST fail signature-replay dedup on the second delivery. Running the two
+#   checks together catches this class of bug; a retry-stability step that skips
+#   signature verification hides it.
+#
+#   Fields:
+#     triggered_by: <step_id>
+#     filter: { ... }             # same shape as expect_webhook
+#     retry_trigger:
+#       count: 3                  # runner returns 5xx for the first N deliveries.
+#                                 # MUST be 1 ≤ count ≤ 10 (test-kit
+#                                 # retry_replay_contract.max_count). Counts outside
+#                                 # the range fail storyboard validation; sellers
+#                                 # legitimately exercise N=3 for back-off sanity,
+#                                 # higher counts risk becoming DoS amplifiers in
+#                                 # proxy_url mode.
+#       http_status: 503          # optional; default 503. MUST be in
+#                                 # {429, 500, 502, 503, 504} (test-kit
+#                                 # retry_replay_contract.allowed_statuses). Other
+#                                 # statuses are not retryable signals for
+#                                 # at-least-once senders and fail validation.
+#     timeout_seconds: 90         # longer default — must cover sender's retry back-off
+#     expect_min_deliveries: 2    # fail if fewer deliveries observed within window
+#     verify_signature_on_every_delivery: true
+#                                 # default true when 9421 is in effect. Set false
+#                                 # ONLY when the storyboard is explicitly registered
+#                                 # for legacy HMAC mode; in 9421 mode, silently
+#                                 # disabling retry-delivery signature verification
+#                                 # hides nonce-replay bugs.
+#
+#   Error modes: insufficient_retries, idempotency_key_rotated,
+#   idempotency_key_format_changed, signature_replayed (nonce reused across
+#   retries), signature_invalid (any retry delivery fails 9421).
+#
+# task: expect_webhook_signature_valid
+#   Assert that the inbound webhook passed the 9421 webhook-signing verifier
+#   checklist (see docs/building/implementation/security.mdx#verifier-checklist-for-webhooks).
+#   Runner delegates verification to `verifyWebhookSignature` from
+#   `@adcp/sdk/signing` (see test-kit contract). Negative conformance —
+#   signatures that MUST be rejected with a specific webhook_signature_* code —
+#   is exercised through the published static conformance vectors under
+#   /compliance/{version}/test-vectors/webhook-signing/, not by this step type.
+#
+#   Cross-step replay dedup: the runner MUST share a single (keyid, nonce) replay
+#   store across ALL expect_webhook_signature_valid invocations (and all retry
+#   deliveries under expect_webhook_retry_keys_stable) in a storyboard run — see
+#   test-kit client_primitives.signature_replay_store. A per-step store would let
+#   a publisher reuse the same (keyid, nonce) bytes at two expect_webhook_signature_valid
+#   steps and pass both in isolation, silently missing cross-step replay.
+#
+#   Fields:
+#     triggered_by: <step_id>
+#     filter: { ... }             # same shape as expect_webhook
+#     timeout_seconds: 30
+#     require_tag: "adcp/webhook-signing/v1"  # default; sanity check
+#
+#   Error modes: signature_invalid, signature_expired, signature_key_unknown,
+#   signature_key_purpose_invalid, signature_digest_mismatch,
+#   signature_tag_invalid, signature_replayed.
+#
+# task: expect_substitution_safe
+#   Assert that a creative preview (preview_html or preview_url) contains
+#   substituted catalog-item macro values that are percent-encoded per RFC 3986
+#   (unreserved-whitelist), per docs/creative/universal-macros#substitution-safety-catalog-item-macros.
+#   Runner delegates HTML parsing, URL extraction, and the encoding check to
+#   the `@adcp/client` SubstitutionObserver primitive (see test-kit
+#   substitution-observer-runner.yaml). Catches implementations that pass raw
+#   attacker bytes through substitution, including CRLF injection, bidi-override
+#   spoofing, reserved-character URL breakout, and javascript:-scheme injection.
+#
+#   Fields:
+#     source: html_inline | url_fetch  # default html_inline
+#     source_path: "<JSON pointer into previous step's response>"
+#     macro_template: "<URL template with {MACRO} placeholders>"
+#     catalog_bindings:
+#       - macro: "{SKU}"
+#         catalog_item_id: "<item_id in the synced catalog>"
+#         vector_name: "<fixture entry name, e.g. reserved-character-breakout>"
+#         # Optional overrides for custom (non-canonical-fixture) vectors:
+#         # raw_value: "<attacker-shaped input>"
+#         # expected_encoded: "<RFC 3986 unreserved-encoded form>"
+#     require_every_binding_observed: true  # default true. Opt to false only
+#                                           # when preview shows a partial
+#                                           # catalog and partial observation
+#                                           # is legitimate (document why in
+#                                           # the narrative). The old field
+#                                           # name require_all_bindings_observed
+#                                           # is deprecated.
+#     requires_contract: substitution_observer_runner
+#
+#   Error modes: substitution_encoding_violation, nested_macro_re_expansion,
+#   substitution_scheme_injection, substitution_binding_missing,
+#   preview_source_unavailable, preview_url_unusable (with sub-reason:
+#   http_status | content_type | size_exceeded | redirect_returned |
+#   ssrf_blocked | fetch_timeout).
+#
+# task: expect_rate_limit_not_replayed
+#   Assert that a `RATE_LIMITED` response on idempotency-cache insert MUST NOT
+#   be cached as the canonical replay for that idempotency_key, per
+#   L1/security.mdx#idempotency rule 3 ("Only successful responses are cached")
+#   and bullet 8 (insert-rate ceiling). The runner drives a sequential fresh-
+#   key burst against the named mutating-op target until a RATE_LIMITED
+#   response appears, then re-submits the same idempotency_key after waiting
+#   the RATE_LIMITED response's top-level `error.retry_after` seconds. The cross-step assertion fails if the
+#   replay returned the cached RATE_LIMITED response shape.
+#
+#   When `max_attempts` requests complete without producing a RATE_LIMITED
+#   response, the step grades with `skip_result.reason: not_applicable` and
+#   `skip_result.detail: rate_limit_not_triggered` — a seller's sandbox limiter
+#   may legitimately sit above the contract's burst ceiling. Runner delegates the burst,
+#   wait, and replay mechanics to the `@adcp/client` RateLimitTripObserver
+#   primitive (see test-kit rate-limit-trip-runner.yaml).
+#
+#   Fields:
+#     rate_limit_trip:
+#       trip_target_task: "<task name, e.g. create_media_buy>"
+#                                       # the mutating-op the burst targets
+#       trip_target_sample_request: { ... }
+#                                       # full canonical request body; the runner
+#                                       # rewrites `idempotency_key` per attempt
+#                                       # and on the replay reuses the captured key
+#       max_attempts: 200               # 50–500; see test-kit max_attempts bounds
+#       replay_max_wait_seconds: 30     # optional; caps the post-trip wait. Sellers
+#                                       # returning retry_after > this fail with
+#                                       # `replay_wait_exceeded` rather than the
+#                                       # invariant itself.
+#     requires_contract: rate_limit_trip_runner
+#
+#   Failure outcomes: rate_limit_response_cached_as_replay,
+#   missing_retry_after, replay_wait_exceeded.
+#
+#   Not-applicable outcome: rate_limit_not_triggered. This is emitted when
+#   max_attempts exhausts without observing RATE_LIMITED and MUST NOT be
+#   counted as a seller conformance failure.
+#
+# --- shared_receiver (deferred / out of scope v1) ---
+#
+# Storyboards with multiple webhook-triggering steps MUST use per-step receiver
+# URLs; the default substitution `{{runner.webhook_url:<step_id>}}` gives each
+# step its own receiver so expect_webhook scopes cleanly. A `shared_receiver: true`
+# flag on an emitting step would route all emissions from that step to a shared
+# bucket (for fan-in dedup tests), but the mechanics — filter matching across
+# multiple emitters, retry-replay policy precedence, shared replay store scoping —
+# are underspecified. shared_receiver is OUT OF SCOPE for the first
+# webhook-emission universal release; storyboard authors MUST NOT set it. Future
+# fan-in storyboards will either flesh out the semantics or adopt a different
+# idiom (e.g., per-step receivers with cross-step assertions on aggregate
+# observations).
+#
+# --- Context accumulator and substitution ---
+#
+# The runner maintains a storyboard-run-scoped context accumulator populated from
+# three sources (evaluated in this precedence, high to low):
+#
+#   1. Step `context_outputs:` captures (populated after a step's validations pass)
+#   2. Storyboard-root `context:` block (literal values fixed at run start)
+#   3. Test-kit substitutions where explicitly scoped into context
+#
+# Two substitution syntaxes are supported in `sample_request` payloads:
+#
+#   $context.<name>               — single-token substitution. The runner replaces
+#                                    the literal string "$context.<name>" with the
+#                                    captured value. Example:
+#                                      account_id: '$context.account_id'
+#                                    resolves to the account_id captured from an
+#                                    earlier step. Works inside scalar values and
+#                                    inside array elements.
+#
+#   {{prior_step.<id>.<field>}}   — Mustache-style interpolation for webhook
+#                                    filters, operation_id matching, and any other
+#                                    cross-step reference. Resolves against the
+#                                    named prior step's captured response fields.
+#                                    Used by expect_webhook and expect_*_webhook_*
+#                                    assertion steps (see Webhook-assertion steps
+#                                    above) and by storyboards correlating an
+#                                    async operation_id across steps.
+#
+# Runner requirements:
+#   - Substitution runs at step-execution time, after resolving the test-kit
+#     overlay and before applying auth. The agent under test receives the
+#     substituted value — it MUST NOT see the literal `$context.foo` or `{{...}}`
+#     token on the wire (see Unresolved substitution behavior below for the one
+#     legitimate exception: preflight not_applicable grading when a substitution
+#     is statically unresolvable).
+#   - $context.<name> referencing an unpopulated name (the capturing step failed
+#     or did not run) grades the referencing step as failed with
+#     `unresolved_substitution`. The storyboard MUST NOT proceed with a fabricated
+#     default — test outcomes on synthesized state are not signal.
+#   - Captured values are typed as the JSON path resolved against the response;
+#     numeric and object captures are preserved (they do not round-trip through
+#     string coercion).
+#   - When the per-task request builder discards body fields and
+#     `$context.<name>` inside `sample_request` never reaches the wire (the
+#     observable symptom is the agent answering with an unfiltered result
+#     and the round-trip assertion grading on fabricated state), use
+#     `context_inputs:` as the post-builder injection path. See the
+#     `context_inputs:` field definition above for the full semantics and
+#     when it is acceptable to reach for.
+#
+# --- Context echo contract (agent-facing) ---
+#
+# Agents under test MUST obey the following when a storyboard sends a `context:`
+# block on its sample_request:
+#
+#   1. Successful responses MUST echo the `context` object verbatim. The agent
+#      MUST NOT mutate, filter, or reorder fields. The echo applies whether the
+#      response is synchronous (`completed`) or asynchronous (`submitted` /
+#      `working`).
+#   2. Error responses MUST also echo the `context` object verbatim. Context is
+#      used as a correlation surface; dropping it on failure prevents buyers from
+#      correlating errors with the originating call.
+#   3. Agents MUST NOT synthesize a `context` object when the caller did not
+#      send one. A storyboard whose `sample_request` lacks `context:` MUST see
+#      the response lack `context:` — fabricating `{ correlation_id: "..." }` on
+#      the agent side is a conformance failure because the storyboard validator
+#      is asserting on what-the-caller-sent, not what-the-agent-invented.
+#   4. Agents MUST NOT parse or act on `context` contents — see
+#      docs/building/integration/context-sessions.mdx for the normative
+#      agent-facing rules. The storyboard runner's validators rely on verbatim
+#      echo and will fail any mutation.
+#
+# Runners MUST NOT auto-inject an implicit `context:` block on sample_request
+# payloads the storyboard did not declare. Storyboards that want to assert on
+# `context.correlation_id` MUST declare the context block explicitly on each
+# sample_request. This keeps the sample_request an honest shape of what the
+# agent will receive, and prevents an agent from passing only because the
+# runner compensated for its missing contract.
+#
+# --- Unresolved substitution behavior ---
+#
+# When a storyboard references a substitution variable the runner cannot
+# resolve — whether a `$context.*` token or a `{{...}}` token, including tokens
+# nested in object properties or array elements — the runner MUST NOT ship the
+# literal token to the agent under test. Shipping unresolved substitutions is a
+# conformance bug on the runner side, not a grading signal about the agent.
+#
+# Runners MUST resolve substitutions at one of two preflight points:
+#   1. Before the storyboard run starts — if any step references an unresolvable
+#      substitution, grade the storyboard as not_applicable with
+#      `unresolved_substitution` as the reason, and skip execution entirely.
+#      This is the preferred behavior for absent webhook receivers.
+#   2. Before individual step execution — if a later-computed substitution
+#      (e.g., {{prior_step.<id>.task_id}}) fails to resolve, grade the
+#      step as failed with `unresolved_substitution`.
+#
+# Exception for runner-owned creative-asset coverage: a
+# `$build_assets_from_format` directive can only be resolved after a prior step
+# captures the seller's valid format contract. If the format context exists but
+# the runner's test kit cannot synthesize every required slot, the runner MUST
+# apply a storyboard-level `not_applicable` disposition with
+# `skip_result.reason: fixture_unavailable` at the earliest preflight after the
+# referenced format is captured, not wait until the directive-bearing step.
+# This preflight MUST occur before any intervening side-effecting step. The
+# runner then terminates the remaining storyboard without executing later
+# phases or emitting `prerequisite_failed` cascades. If the format context itself
+# is absent because its producer failed or was skipped, point 2 and the normal
+# prerequisite cascade still apply.
+#
+# Neither path may ship a literal `$context.*` or `{{...}}` token on the wire.