Summary

4-deliverable batched bucket: 2 code follow-ups to #189 + 2 research-only docs (in companion kora-docs PR linked below).

Deliverable A — escalation_reason structured telemetry field

Threaded through the existing telemetry layers:

• `agent/cost_state_holder.py` — `record_inference(..., escalation_reason: Optional[str] = None)` plumbs to `record_call`
• `agent/cost_ladder_wire.py` — `record_inference_from_response` forwards the kwarg
• `kora_cli/telemetry/cost_telemetry.py` — `_RouteCounters.escalation_reason_breakdown: Dict[str, int]` lights up only when both `escalated_to_opus=True` AND reason is non-empty string
• `kora_hermes_plugin/haiku_router/plugin.py` — handler now returns `{"reissue_with": ..., "escalation_reason": }`
• `agent/conversation_loop.py` — reissue site reads `escalation_reason` from plugin result + passes to `record_inference_from_response`

After this lands, cost-telemetry snapshot exposes per-reason escalation breakdown — cockpit panels can show "X% low_confidence_marker, Y% short_response_for_long_input".

Deliverable B — api_call_count accounting confirmation

Per #189 PR body the chosen semantic was "transparent upgrade" — a re-issue is part of the same iteration; `api_call_count` is NOT incremented and `iteration_budget` is NOT consumed. Documented explicitly in:

• `kora_hermes_plugin/haiku_router/plugin.py` module docstring (new "# api_call_count accounting (feat(kora): KR-HERMES-LOCAL-EXT-REISSUE + KR-HAIKU-ROUTER-PLUGIN — completes Lock R3-2 Phase C #189 follow-up B — confirmed)" section)
• `agent/conversation_loop.py` inline comment at the reissue site

Pinned by two structural tests that slice the loop source between the hook firing and the `post_api_request` observer and assert there's NO `api_call_count += 1` and NO `iteration_budget.consume(` in that slice.

Deliverables C + D — research only (separate docs PR)

Land in kora-docs at `kora_docs/14_research/`. Both available for review at:

• Companion PR: https://github.com/rafe-walker/kora-docs/pull/new/docs/kora-KR-CC3-CLEANUP-AND-DAEMON-PREP-MEGABUCKET
• C: `daemon_listeners_via_gateway_2026-05-24/` — per-listener audit (16 listeners; 4 categories) + 6-phase migration plan + DaemonCoordinator dissolution analysis. Zero new Hermes extensions required for the migration.
• D: `upstream_pr_packaging_2026-05-24/` — REPORT + 5 ready-to-submit upstream-Hermes PR drafts (one per local extension). Recommended submission order: KR-1 ST1: Kora runtime fork + baseline recon #1-3 now (3+ days soak met), KR-1 ST4: Operational verification + console-script finalize (closes KR-1) #4-5 wait for 7-day stability.

Test plan

• Deliverable A tests: 7 new test cases covering field plumbing (escalation_reason omitted / non-escalated / empty / non-string fallback / snapshot shape stability)
• Deliverable B tests: 2 structural pins (no `api_call_count += 1` and no `iteration_budget.consume(` between hook firing and post_api_request observer)
• Regression: 178/178 focused-scope tests pass (haiku_router + kora_hermes_plugin + cost_ladder + cost_telemetry + cost_telemetry_listener + post_llm_can_reissue)
• Burn-in verification post-merge: confirm cockpit cost-telemetry panel reads the new `escalation_reason_breakdown` field (CC#2 follow-on if desired; not blocking)

🤖 Generated with Claude Code