gen_ai.agent.name never reflects the active custom agent for top-level (non-subagent) sessions

Copilot Chat's OTel export (`docs/monitoring/agent_monitoring.md`) documents `gen_ai.agent.name` as required on `invoke_agent` spans and the `copilot_chat.agent.*` metrics. In practice, for a session where the user has directly selected a custom `.agent.md` agent (not one dispatched via `runSubagent`), this field is always the generic `"GitHub Copilot Chat"` — including for `copilot_chat.session.start`, `copilot_chat.agent.turn.count`, and `copilot_chat.agent.invocation.duration` — never the actual persona's name.

Confirmed against a real 22,950-line/242MB capture: zero occurrences of a directly-selected custom agent's real name anywhere in the file, across every `session.start` event. Subagents dispatched via `runSubagent` *do* get real names in the same file (11k+ occurrences), confirming the identity-threading mechanism exists — it's just not wired up for the foreground/top-level case.

Root cause, from source (as it stood in the former `microsoft/vscode-copilot-chat` repo, prior to archival — same logic should now live under this repo's Copilot Chat extension code): `provideCustomAgents()` (e.g. `githubOrgCustomAgentProvider.ts`, the local-file equivalent) returns bare file URIs, not distinct `vscode.ChatParticipant` registrations — VS Code loads the `.agent.md` content as a persona layered onto one fixed shared participant (`editsAgent`, see `chatAgents.ts`). `emitSessionStartEvent()` (`genAiEvents.ts`) derives `gen_ai.agent.name` from that participant, so every directly-selected custom agent collapses to the same generic identity regardless of which persona is actually active.

This is also user-visible in a confusing way: the agent will correctly self-report its real persona name in chat if asked directly ("Which agent are you?" → "JARVIS"), since that's a prompt-layer fact the model has — but that identity never makes it into the telemetry layer.

**Ask:** thread the active custom agent's real name into `gen_ai.agent.name` for top-level/foreground sessions too, not just `runSubagent`-dispatched ones — the persona is already known internally (it's driving the system prompt and tool list), it's just not being surfaced to the OTel identity fields.