fix: allow owned signals discovery without activation

[bokelley]

Summary

• make signal-owned discovery-only by requiring and advertising get_signals without activate_signal
• keep signal-marketplace on the existing get_signals + activate_signal surface
• add OwnedSignalsPlatform as the public protocol for seller-owned signal catalogs
• add regressions for required-method validation and tools/list advertisement

Root Cause

signal-owned reused the marketplace/provisioned signal method and advertised-tool set, forcing seller-owned signal adopters to expose a fictional activation step even when returned first-party signals are already usable on seller inventory.

Validation

• uv run --extra dev python -m pytest tests/test_decisioning_specialisms.py tests/test_decisioning_advertised_per_specialism.py tests/test_decisioning_capabilities_projection.py tests/test_platform_router.py tests/test_decisioning_handler_shims.py
• commit pre-commit hooks: black, ruff, mypy, bandit, whitespace, YAML/JSON, large file, merge conflict, case conflict, private key checks

[aao-ipr-bot]

LGTM. Clean split — the protocol shape now matches the spec, where activate-signal-request requires destinations with minItems: 1, so activating an owned signal onto the seller's own inventory was always a structurally nonsense call.

Things I checked

• SignalsPlatform(OwnedSignalsPlatform[TMeta], Protocol, Generic[TMeta]) with @runtime_checkable produces the right isinstance semantics on both arms — verified by test_owned_signals_platform_runtime_checkable_with_get_signals_only at tests/test_decisioning_specialisms.py:117-124.
• _KNOWN_SPECIALISM_PROTOCOLS lists both OwnedSignalsPlatform and SignalsPlatform, so _all_specialism_methods() still unions to {get_signals, activate_signal} even though vars(SignalsPlatform) now only contains activate_signal (platform_router.py:138-150).
• Boot vs runtime gates stay consistent: REQUIRED_METHODS_PER_SPECIALISM["signal-marketplace"] (dispatch.py:209-215) still requires both methods; _require_platform_method("activate_signal") at handler.py:2274 plus the _OPTIONAL_PLATFORM_METHODS listing at handler.py:347 gives buyers a clean UNSUPPORTED_FEATURE rejection instead of an AttributeError if they call activation against an owned-only seller.
• Conventional-commits framing as fix: is defensible — no existing signal-owned adopter breaks (relaxed validation passes both shapes), and tools/list dropping activate_signal on signal-owned is the right direction per spec. Borderline feat!: if a buyer hard-coded the advertised set, but tolerable.
• specialism enum slugs unchanged (schemas/cache/3.0/enums/specialism.json); no schema touch needed.

Follow-ups (non-blocking — file as issues)

• get_signals response schema still requires deployments (schemas/cache/3.0/signals/get-signals-response.json:90-100). Owned-signal adopters need to emit a self-referential type:"agent" deployment entry. Worth a one-line contract note in the OwnedSignalsPlatform.get_signals docstring at src/adcp/decisioning/specialisms/signals.py:69-86 so first adopters don't ship a schema-invalid response.
• No router-synthesis test for a signal-owned-only child platform in tests/test_platform_router.py. The handler-level advertisement test in test_decisioning_advertised_per_specialism.py covers the wire surface; the router fan-out path is still implicitly covered via the union but not explicitly locked.
• Release notes / CHANGELOG should flag the tools/list advertisement change for signal-owned sellers as a buyer-observable delta.

Minor nits (non-blocking)

1. Stale docstring assertion in _protocol_method_names. src/adcp/decisioning/platform_router.py:162-166 still claims "there's no Protocol inheritance to chase, so a single vars() pass gets every declared method." That's no longer true now that SignalsPlatform inherits from OwnedSignalsPlatform. The code is fine because both Protocols are independently enumerated in _KNOWN_SPECIALISM_PROTOCOLS, but the next maintainer who adds a sub-Protocol without listing it independently will be misled.
2. Comment cross-reference in _OPTIONAL_PLATFORM_METHODS. src/adcp/decisioning/handler.py:345-347 reads as if this set is what keeps activate_signal required for marketplace. The actual gate is REQUIRED_METHODS_PER_SPECIALISM["signal-marketplace"]; cross-reference it explicitly so the defense-in-depth role is obvious.
3. examples/hello_seller_signals.py is now marketplace-only in spirit but still implements both methods. Docstring update is accurate; a follow-up hello_owned_signals.py would be the cleaner pairing.

Approving.

[aao-ipr-bot]

LGTM. Follow-ups noted below. The dual-Protocol split with SignalsPlatform(OwnedSignalsPlatform[TMeta], ...) is the right shape — preserves static narrowing via isinstance, matches the existing CreativeBuilderPlatform / CreativeAdServerPlatform precedent, and the validation/advertised-tools/manifest tables all line up.

Things I checked

• Inheritance preserves runtime_checkable semantics. vars(SignalsPlatform) carries only activate_signal post-split; vars(OwnedSignalsPlatform) carries get_signals. Both classes are in _KNOWN_SPECIALISM_PROTOCOLS (src/adcp/decisioning/platform_router.py:138-150), so _all_specialism_methods() (line 184) still unions to {get_signals, activate_signal} — synthesized delegation is unaffected. isinstance(get_signals_only_impl, SignalsPlatform) returns False as asserted at tests/test_decisioning_specialisms.py:119-128.
• Defense-in-depth on marketplace adopters is intact. REQUIRED_METHODS_PER_SPECIALISM["signal-marketplace"] at src/adcp/decisioning/dispatch.py:212-216 still requires both methods at boot; the new _require_platform_method("activate_signal") gate at src/adcp/decisioning/handler.py:2274 is runtime defense-in-depth, not a weakened gate.
• Existing adopter compatibility. A signal-owned adopter that previously implemented both methods (forced) still validates after this change — get_signals required, activate_signal simply ignored by the gate. Pure additive relaxation.
• Capabilities projection. SPECIALISM_TO_PROTOCOLS[\"signal-owned\"] still maps to {\"signals\"}, so a discovery-only owned-signals platform emits supported_protocols: [\"signals\"] correctly (tests/test_decisioning_capabilities_projection.py:146-157).
• OwnedSignalsPlatform is added to adcp.decisioning.__all__ at src/adcp/decisioning/__init__.py:362 — clean public-surface addition, no removals.

Follow-ups (non-blocking — file as issues)

• Manifest hand-edit is fragile. schemas/cache/3.0/manifest.json is a cached upstream artifact; scripts/sync_schemas.py (replace_cache_from_bundle) rmtrees and rewrites schemas/cache/3.0/ wholesale on every sync. The CI schema-drift check at .github/workflows/ci.yml:347-353 logs but does NOT fail on diff, so this PR passes CI either way. If upstream 3.0.7 (src/adcp/ADCP_VERSION) still lists signal_owned in tools.activate_signal.specialisms, the next clean python scripts/sync_schemas.py silently reverts the hand-edit and test_signal_owned_manifest_exercises_discovery_only (tests/test_decisioning_specialisms.py:212-231) starts failing. Confirm upstream adcontextprotocol/adcp has shipped the matching manifest update on the 3.0.x train, or add a post-sync patch step so the SDK contract doesn't depend on a manifest line nobody else is updating.
• Conventional-commit prefix is a judgment call. fix: is defensible (relaxing a fictional activation requirement IS the bug fix), but the diff also adds a new public export OwnedSignalsPlatform. feat: → 5.7.0 would be the more honest semver signal against the 5.6.0 baseline at .release-please-manifest.json. Non-blocking — maintainer call.
• JS-side parity unverified. Docstring at src/adcp/decisioning/specialisms/signals.py:25-26 claims this mirrors src/lib/server/decisioning/specialisms/signals.ts in the JS SDK. If the JS side still ships a single SignalsPlatform interface, the cross-SDK conformance storyboard will see an advertised-tools mismatch for owned-signal adopters.

Minor nits (non-blocking)

1. get_signals docstring drift. src/adcp/decisioning/handler.py:2253 still reads "Catalog discovery for signal-marketplace / signal-owned" — accurate, but every other comment in this diff was updated to call out the discovery-only owned variant. Cosmetic.

Approving on the strength of the inheritance shape plus the manifest test catching the regression if it happens.