usage estimation: wire the non-streaming ensemble sub-call events (panel + judge)

[jarvis9443]

Follow-up to #794 (AISIX-Cloud#1074). The non-streaming ensemble path emits one usage event per panel member plus the judge (emit_subcalls in chat.rs), but none of those events go through the token-estimation fallback: with a no-usage backend they still record silent zeros.

The streaming ensemble judge IS wired (#794 arms the SSE Drop guard with the judge's own request). Wiring the non-streaming side needs the judge request and the per-member answer texts at the emit site, which ensemble.rs currently keeps internal — that refactor was out of scope for the #794 review fixes (audit finding L2, deferred explicitly).

Scope:

• panel member events: prompt from the client request, completion from the member's buffered answer text
• judge event: prompt from judge_request(...) output, completion from outcome.response text
• both flagged usage_estimated per the feat(usage): estimate tokens locally when the upstream reports no usage #794 or-semantics helper
• test: wiremock ensemble with a usage-less backend asserting non-zero flagged sub-call events

[coderabbitai]

🔗 Related PRs

#783 - fix(proxy): enforce each routing target's own model rate limit [merged]
#784 - feat(obs): add model label to aisix_llm_tokens_by_client_total and record it on /v1/responses [merged]
#785 - feat(metrics): extend client_type coverage and add operator UA mapping rules [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[jarvis9443]

Done — merged in #806 (f8b5e0fc).

Every ensemble sub-call usage event now runs through the #1074 estimation fallback when the backend omits usage: the non-streaming panel members + judge, and the streaming-path panel members (the streamed judge was already wired by #794). Each is estimated (prompt from the sub-call's own request, completion from its answer text) and flagged usage_estimated, via one shared helper keyed off the sub-call's resolved upstream model. Telemetry-only — no client body/stream rewrite, and no CP/schema change (the sub-call events already carry the field).

Covered by a new DP E2E (a real ensemble over usage-less upstreams, non-streaming + streaming) asserting per-record that every sub-call ships its own flagged event — verified failing before the fix, passing after — plus unit tests for the estimator or-semantics.