Salvages #29722 onto current main. Authorship preserved (Bryan Bednarski's commit), with a small follow-up cleanup commit from me.

What this is

Phase 1 of the NeMo telemetry stack, restructured per the #29722 review into an observer-only PR with an immediate in-repo consumer:

• Backend-neutral observer hooks for plugins — session, turn, API request, tool, approval, and subagent lifecycle events — with stable correlation IDs (session_id, task_id, turn_id, api_request_id, tool_call_id, parent/child subagent ids). Extends VALID_HOOKS with api_request_error and subagent_start.
• Bundles the optional NeMo-Relay observability plugin (plugins/observability/nemo_relay) as a real consumer of the new hooks, peer to the existing langfuse plugin. Fails open when the optional nemo-relay package isn't installed.
• The speculative middleware surface from the original PR was removed (deferred to a follow-up paired with an adaptive consumer).

Review concerns from #29722 — all addressed

• Per-call overhead, zero plugins registered: payload construction is now gated behind has_hook() presence checks; request payloads return by reference when no middleware rewrites; the sanitized response payload no longer embeds raw response objects; redundant deepcopies removed. Contributor benchmark: no-listener path dropped from ~20 ms → ~0.0004 ms at 5 MB context.
• Middleware with no consumer: middleware module/registry removed entirely; PR is observer-only with nemo_relay as the in-repo consumer.
• Schema-version threading: centralized in invoke_hook (setdefault), not ~20 hand-edited call sites.
• post_tool_call dual-emit: hardcoded set replaced with shared AGENT_RUNTIME_POST_HOOK_TOOL_NAMES frozenset + agent_runtime_owns_post_tool_hook() helper.
• api_request_id format: docs now state it's opaque — "do not parse its string format."

Follow-up commit (mine)

• Restored two unrelated trailing blank lines (test_file_tools_cwd_resolution.py, test_tool_search.py) that the original branch incidentally stripped, to keep the salvage scoped to the feature.

Testing

• Unit: 179 targeted tests pass (plugins, model_tools, session-boundary hooks, nemo_relay plugin suite, the two reverted test files) + full tests/run_agent/ green.
• E2E (real imports, isolated HERMES_HOME): verified (1) has_hook() returns False with no plugin → hot path skipped; (2) a registered observer fires with turn_id/api_request_id and auto-injected schema version; (3) nemo_relay register() succeeds and all hooks no-op (don't crash) when the nemo-relay package is absent → _get_runtime() returns None.
• ruff clean on all changed source files.

Attribution

Cherry-picked from #29722 with Bryan Bednarski's authorship preserved in git log. Requires the AUTHOR_MAP chore PR to merge first so contributor_audit.py passes.

Closes #29722