Search summaries, block-revision assist and workflow runners always use the classic GPT endpoint, which V2 cannot configure

Nobody has claimed this yet.

## Summary

Four subsystems read the classic single-endpoint GPT configuration **unconditionally**, with no multi-endpoint branch. They therefore always use `azure_openai_gpt_*` / `enable_gpt_apim` / `azure_apim_gpt_*`, regardless of whether `enable_multi_model_endpoints` is on.

Because the V2 admin surface deliberately does not carry the classic chat form, **these four subsystems cannot be configured from V2 at all.**

## Which consumers have a connections branch, and which do not

Counting references on `paullizer-react-v2-ui`:

| File | `enable_gpt_apim` | `enable_multi_model_endpoints` | |
|---|---|---|---|
| `functions_search_service.py` | 1 | **0** | no connections branch |
| `functions_block_revision_assist.py` | 1 | **0** | no connections branch |
| `functions_workflow_runner.py` | 1 | **0** | no connections branch |
| `functions_personal_workflows.py` | 1 | **0** | no connections branch |
| `route_backend_chats.py` | 8 | 5 | resolves connections first |
| `functions_documents.py` | 5 | 1 | resolves connections first |
| `route_backend_agents.py` | 4 | 4 | resolves connections first |
| `route_backend_conversation_export.py` | 3 | 1 | resolves connections first |

`_create_summary_client` in `functions_search_service.py` is representative — it builds its client straight from the classic settings with no connections lookup anywhere in the function.

## Why this matters

The AI Models rebuild (#1415, #1419, #1426) made V2's AI Models group **look** complete: connections, chat, embeddings and image generation all have real controls. But an administrator working only in V2 has no way to configure the endpoint that search summarisation, block-revision assist, the workflow runner and personal workflows actually use.

Worse, the failure is quiet. Those subsystems do not fall back to connections — they simply use whatever the classic config holds, which on a deployment that has only ever been configured through V2 may be empty or stale. Nothing in V2 indicates this.

This was partly obscured by wording: it is natural to describe `enable_gpt_apim` as "routes chat through API Management". It does considerably more than that, and the flag's blast radius is wider than the Chat section that now hosts its toggle.

## Suggested directions

Two coherent options, plus a stopgap:

1. **Give these four call sites a connections branch**, so they resolve a model the same way chat, documents, agents and export already do. This is the direction that lets V1 be retired.
2. **Bring the classic single-endpoint chat form into V2** so at least the configuration is reachable, accepting the duplication.
3. **Stopgap:** state it in the docs and in the Chat section's notice — that APIM and the classic endpoint apply to more than chat, and that these subsystems always use them. #1419 has already added a sentence to `docs/admin/ai-models.md` along these lines.

Option 1 is the real fix. Until one of these happens, **V1 cannot be retired**, because there is configuration that exists only there.

## Provenance

Found during the three-phase AI Models rebuild. The #1419 session traced `enable_gpt_apim`'s consumers before writing its help text and correctly declined the narrower "routes chat through API Management" wording; that trace is what surfaced this. Verified independently against `origin/paullizer-react-v2-ui`.

## Related

- #1415 — global model connections (phase 1, merged)
- #1419 — Chat section (phase 2), which declares `enable_gpt_apim`
- #1426 — Embeddings and Image generation (phase 3)
- #1420 — a separate capability-scan defect found in the same work

Start with _create_summary_client in functions_search_service.py, then compare its classic-settings lookup with the connections resolution in route_backend_chats.py and functions_documents.py. Trace the corresponding paths in functions_block_revision_assist.py, functions_workflow_runner.py, and functions_personal_workflows.py. Done means all four consumers can use the model configured through V2 multi-model endpoints rather than unconditionally using the classic GPT settings.

Written by the indexing model from the issue text.