fix(decisioning): gate async-completion webhook on SPEC_WEBHOOK_TASK_TYPES

[bokelley]

Problem

After #930, _project_handoff (src/adcp/decisioning/dispatch.py) calls emit_terminal_completion_webhook (src/adcp/decisioning/webhook_emit.py) for every async task. SDK-internal, non-spec task types — notably finalize_proposal, an interception of get_products in proposal_dispatch.py that is not a spec wire op and not in SPEC_WEBHOOK_TASK_TYPES — flow through with no webhook target wired. The emitter then hit its target is None branch and logged a spurious WARNING:

neither webhook_sender nor webhook_supervisor is wired — terminal webhook silently dropped

on every async finalize, even on a correctly-configured server.

The SDK's own documented rule (the docstring above SPEC_WEBHOOK_TASK_TYPES) is that task types not in the spec enum skip webhook delivery and rely on tasks/get polling / publishStatusChange.

Fix

Hoist the SPEC_WEBHOOK_TASK_TYPES check to the top of emit_terminal_completion_webhook — after the enabled gate, before the target is None warning — and return silently for non-spec task types. This mirrors how the sync emitter (maybe_emit_sync_completion) gates. The previous spec-enum check lower in the function (which logged a WARNING on skip) is removed since the new gate short-circuits earlier and silently.

Behavior:

• non-spec task type (e.g. finalize_proposal) → skip silently, no warning, no emission (the bug);
• spec task type with a real target → emits (unchanged);
• spec task type with target=None (genuine misconfig) → the existing target-None warning still fires.

Tests

Three new tests in tests/test_decisioning_webhook_emit.py exercising the emitter directly:

• non-spec task type (finalize_proposal) with push config and target=None → no emission AND asserts via caplog that no "silently dropped" warning is logged;
• spec task type (create_media_buy) with a target → still emits with correct payload;
• spec task type with target=None → the misconfig warning still fires.

Verification: pytest tests/test_decisioning_webhook_emit.py (29 passed), make lint, make typecheck all green.

Closes #931

🤖 Generated with Claude Code

[aao-ipr-bot]

Clean fix. Hoisting the SPEC_WEBHOOK_TASK_TYPES gate above the target is None branch is the right shape — non-spec SDK-internal task types like finalize_proposal were never webhook-eligible, so they should never reach the misconfig warning meant for genuinely-dropped spec notifications.

Things I checked

• Gate ordering loses no required side effect. The new if method_name not in SPEC_WEBHOOK_TASK_TYPES: return at webhook_emit.py:429 runs before strip_credentials_from_wire_result (L457) and the extractors — all pure/read-only, and only ever consumed by the send_mcp call non-spec types never reach. No registry write lives in this function; _project_handoff already persisted terminal state upstream.
• No credential-strip bypass. Every path that still reaches target.send_mcp (L470) passes through the stripper at L457 unchanged — the reorder only removed a return that sat after the strip, it moved nothing between strip and send. CREDENTIAL_BEARING_METHODS ∩ SPEC_WEBHOOK_TASK_TYPES (create_media_buy, update_media_buy, sync_accounts, sync_creatives) all pass the moved-up gate and still hit the stripper. Credential-bearing non-spec methods now return at the top gate and never send — a strict reduction in delivery surface. security-reviewer: clean.
• Old lower gate fully removed, no orphan. The L459-468 WARNING is deleted, not stranded. Behavior change is intentional and documented: non-spec types are now silent in both the target-None and target-present cases (previously a non-spec type with target=None short-circuited at the earlier warning, so the lower gate only fired when target was non-None).
• Tests exercise the real gate. Membership matches webhook_emit.py:72-92 — create_media_buy is in the enum, finalize_proposal is not. The firing test's result == {\"media_buy_id\": \"mb_1\"} assertion survives the strip (media_buy_id isn't a scrubbed key). The third test confirms the genuine-misconfig warning still fires for a spec type with target=None. code-reviewer: clean.
• Docstring updated to match. The diff removes the "non-spec type logs a WARNING" bullet from the "Logs a WARNING when" section and narrows the target-None bullet to SPEC-eligible types — consistent with the new control flow.
• fix(decisioning): is the correct semver signal; no public-API surface changed, no breaking diff.

Minor nits (non-blocking)

1. Two doc sites now carry the same rule. The hoisted-gate rationale is spelled out both in the docstring (L384-397) and in the inline comment at L418-427. Fine as-is — load-bearing reorders earn the redundancy.

Closes #931 with a regression test that asserts the WARNING is absent via caplog. Right way to lock a log-noise bug.

LGTM.