feat!: AI wrappers always send through capture_ai

[eli-r-ph]

💡 Motivation and Context

• PostHog AI wrapper events go to the analytics endpoint unless the client sets enable_full_ai_capture. The analytics endpoint caps an event near 1 MB, so large generations can be rejected there.
• The AI endpoint accepts events up to 8 MiB, and its lane batches and retries AI traffic away from analytics events (feat!: send capture through v1 by default; capture_ai to /i/v1/ai/events #1017, feat: AI lane config and size-aware batching #1022).
• One flag controls both the routing and how much content the wrappers keep. A user who wants the larger limit must also turn off truncation and media redaction.

This PR sends every wrapper event through capture_ai. enable_full_ai_capture now controls only content: truncation and media redaction.

Changes

• The OpenAI, Anthropic, Gemini, LangChain, Claude Agent SDK and OpenAI Agents integrations always send through capture_ai, with or without enable_full_ai_capture. They all use one helper, _capture_ai_event.
• A client object without a capture_ai method, such as a custom stand-in, still receives capture calls.
• The enable_full_ai_capture docstrings no longer say the flag routes events.
• _use_ai_lane and _enable_multimodal_capture stay deprecated aliases of enable_full_ai_capture. Their names are now misleading, but callers that set them get the same result as before.
• MCP events stay on capture, because they are not $ai_* events.

Warning

Two breaking effects for users:

• A test suite that passes a Mock client to a wrapper and asserts on mock.capture now sees the call on mock.capture_ai. The old helper tests kept Mock clients on capture on purpose; those tests are replaced.
• A deployment that does not serve /i/v1/ai/events now drops wrapper events. Before this PR, wrapper events without the opt-in used /i/v1/batch/.

Divergence from sdk-specs

The capture-ai spec in PostHog/sdk-specs ties wrapper routing to the flag. This PR diverges from it. Proposed spec change:

-accepts much larger payloads than analytics capture. PostHog's own AI wrapper
-libraries route through it when the client opts in; users call it directly for
-manual AI capture.
+accepts much larger payloads than analytics capture. PostHog's own AI wrapper
+libraries always route through it; users call it directly for manual AI capture.
@@ ## Configuration
-Node — client option), default false. When true, PostHog's AI wrapper
-libraries:
-
-1. route their events through the dedicated AI path instead of analytics
-   `capture`,
-2. skip string truncation (a cut through base64 corrupts media; large text
+Node — client option), default false. It governs content only, not routing.
+When true, PostHog's AI wrapper libraries:
+
+1. skip string truncation (a cut through base64 corrupts media; large text
    contexts stay intact), and
-3. pass media (raw bytes, base64, data URIs) through unredacted instead of
+2. pass media (raw bytes, base64, data URIs) through unredacted instead of
    replacing it with type-aware placeholders.

Plus a new requirement: "AI wrappers send through capture_ai", with one scenario per flag value that asserts the wrapper event arrives on the AI endpoint only. The spec PR is not opened yet.

💚 How did you test it?

• The wrapper suites now assert on capture_ai (about 465 assertions, renamed by a script). Every wrapper test therefore checks the default routing, with the flag off.
• test_routes_only_through_ai_lane sends through a real client, with the flag off and on. It asserts the AI path is the only path posted, so it also catches an event sent on both lanes.
• test_client_without_capture_ai_falls_back_to_capture covers a client object with no capture_ai, with the flag off and on.
• The per-wrapper "opt-in routes through capture_ai" tests, the Mock client tests and one duplicate lane test are removed. The tests above cover them.
• Break-on-purpose, each with failing tests: the old flag gate restored (311 failures), events sent on both lanes (12), the fallback removed (16).
• The public API snapshot is unchanged.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran sampo add to generate a changeset file

Covered by the existing capture-v1-major changeset; the migration guide lands later in this stack.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Written with Cursor (Claude Opus) under the direction of the assignee. Skills used: writing-tests, writing-pr-descriptions.

Agreed before implementation: wrappers always use capture_ai, the flag controls only content, and the PR calls out the spec divergence.

Agent calls worth review:

• The duck-typed fallback to capture stays. Removing it would break custom client objects that have no capture_ai.
• The deprecated aliases stay, because the posthog monorepo still sets them.