Operators with the OpenSearch / ES UBI plugin installed (today; Solr's first-party `solr.UBIComponent` lights up with the sibling `infra_adapter_solr` MVP2 release) can derive judgments from real clic
Operators with the OpenSearch / ES UBI plugin installed (today; Solr's first-party `solr.UBIComponent` lights up with the sibling `infra_adapter_solr` MVP2 release) can derive judgments from real click + dwell behavior via three converters
Operators with the OpenSearch / ES UBI plugin installed (today; Solr's first-party `solr.UBIComponent` lights up with the sibling `infra_adapter_solr` MVP2 release) can derive judgments from real clic
-
-
-
+
Implementing 0
@@ -742,8 +730,6 @@
Dependency graph (feat_ + infra_)
classDef plan fill:#fef9c3,stroke:#854d0e,color:#854d0e;
classDef spec fill:#dbeafe,stroke:#1e40af,color:#1e40af;
classDef idea fill:#f1f5f9,stroke:#334155,color:#334155;
- feat_ubi_judgments["ubi judgments"]
- class feat_ubi_judgments implement;
feat_contextual_help_mvp2["contextual help mvp2"]
class feat_contextual_help_mvp2 done;
diff --git a/docs/00_overview/planned_features/02_mvp2/feat_ubi_judgments/pipeline_status.md b/docs/00_overview/planned_features/02_mvp2/feat_ubi_judgments/pipeline_status.md
deleted file mode 100644
index 570ce1d7..00000000
--- a/docs/00_overview/planned_features/02_mvp2/feat_ubi_judgments/pipeline_status.md
+++ /dev/null
@@ -1,64 +0,0 @@
-# Pipeline Status — UBI Judgments (engine-neutral User Behavior Insights)
-
-## Idea
-- Status: Complete
-- File: [`idea.md`](idea.md)
-- Origin: external review 2026-05-22; 2026-05-27 reframe (Solr bundle); 2026-05-29 `feat_ubi_onramp` merged back in; 2026-05-29 preflight refresh
-
-## Spec
-- Status: Approved
-- Date: 2026-05-29
-- File: [`feature_spec.md`](feature_spec.md)
-- Cross-model review: GPT-5.5 passed (3 cycles; cap hit; 10 findings — 1 H + 1 M cycle 1/2 + 2 H + 4 M + 2 L cycle 3 — all accepted and applied in place; see spec D-10)
-- Phases: 1 default (single-phase delivery); contingency Phase 2 split decided at impl-plan-gen time if bundled diff exceeds ~1500 LOC
-- Scope: ~700 LOC backend + ~350 LOC frontend + ~300 LOC tests + 1 additive Alembic migration (`0021_judgment_lists_generation_params.py`)
-
-## Plan
-- Status: Approved
-- Date: 2026-05-29
-- File: [`implementation_plan.md`](implementation_plan.md)
-- Cross-model review: GPT-5.5 passed (3 cycles; cap hit; all 3 findings accepted — see plan footer)
-- Stories: 14 across 5 epics
-- Phases covered: Phase 1 (all 11 FRs, single-phase delivery per spec D-6)
-
-## Implementation
-- Status: **In progress — Epic 1 of 5 complete** (branch `feat/ubi-judgments`, parent `main` at `68fa357c`)
-- Pacing: Epic by epic, pause between each (operator confirmation 2026-05-29; multi-session execution to respect single-conversation context limits)
-
-### Done
-
-| Epic | Story | Description | Commit |
-|---|---|---|---|
-| 1 | 1.1 | Migration `0021_judgment_lists_generation_params` (JSONB column for UBI worker resume) | `5acdee15` |
-| 1 | 1.2 | `domain/ubi/` pure-domain library (features, async converter Protocol + 3 impls, position-bias prior, 58 unit tests) | `6036586a` |
-| — | — | Planning bundle (idea refresh + spec + plan + dashboards) | `84c810aa` |
-
-### Next session — resume at Epic 2
-
-Run: `/impl-execute docs/00_overview/planned_features/02_mvp2/feat_ubi_judgments/implementation_plan.md 2.1`
-
-Epic 2 stories (the next 3 to ship):
-
-| Story | Description | Estimated scope |
-|---|---|---|
-| 2.1 | `UbiReader` service — engine-agnostic two-index scan + client-side join | ~200 LOC backend + 2 integration tests (the canned-features + the no-cluster-writes guard) |
-| 2.2 | Readiness service + `start_ubi_judgment_generation` dispatcher refactor — shared helpers extracted from `start_judgment_generation` (parity-preserving) | ~250 LOC backend + 1 integration test for the dispatcher matrix |
-| 2.3 | `_SourceBreakdown` evolution + `JudgmentSourceFilterWire` widening + 4 new wire Literals (`UbiConverterKind`, `JudgmentGenerationMethodWire`, `UbiReadinessRungWire`, `UbiMappingStrategyWire`) | ~100 LOC backend + 2 contract tests + 1 unit test |
-
-Sequencing: 2.1 first (used by 2.2), then 2.2 + 2.3 in parallel (different files, no shared state). Branch tip remains `feat/ubi-judgments`; no rebase needed.
-
-### Remaining work (Epics 3–5)
-
-| Epic | Stories | Theme |
-|---|---|---|
-| 3 | 3.1, 3.2, 3.3, 3.4 | API endpoints + worker + agent tool |
-| 4 | 4.1, 4.2, 4.3 | Frontend (enums + hook + badge, dialog method picker + nudge + sparse card, value-delta + recovery cards) |
-| 5 | 5.1, 5.2 | Docs (runbook + glossary + FAQ + tutorial + umbrella spec patches) + E2E suite (4 specs + `seed_ubi.ts` helper) |
-
-After Epic 5 completes: post-implementation ceremony (test coverage audit, deferred work extraction, tangential sweep, guide impact, push, CI watch, Gemini adjudication, final GPT-5.5 review, finalization to `implemented_features/2026_MM_DD_feat_ubi_judgments/`).
-
-## Notes for `/impl-execute` resume
-- **Epic 1 is durable** — the migration is round-trip-clean against the local Postgres, the domain library has 58 passing unit tests, and the branch is push-ready (will push at end of this session per the operator's "push after Epic 1" instruction).
-- **No code in subsequent stories depends on un-pushed Epic 1 state** — Story 2.1 imports `backend.app.domain.ubi.features.UbiEvent` and `aggregate_features`; Story 2.2 imports `backend.app.services.ubi_reader` (Story 2.1's output) + the existing dispatcher helpers; etc.
-- **The plan's §11 consistency review remains green** — none of the cycle-3 fixes contradicted Epic 1 stories (the `generation_kind: 'ubi'` discriminator goes in the JSONB at INSERT time in Story 2.2's dispatcher; the column itself was added in Story 1.1).
-- **Resume command:** `/impl-execute docs/00_overview/planned_features/02_mvp2/feat_ubi_judgments/implementation_plan.md 2.1` (the second argument is the next story ID; `--all` would batch-run Epics 2-5, which is the right call if the next session has token budget for it).
diff --git a/docs/08_guides/README.md b/docs/08_guides/README.md
index 059c71ca..0bf62d21 100644
--- a/docs/08_guides/README.md
+++ b/docs/08_guides/README.md
@@ -60,6 +60,7 @@ Shipped guides:
**Workflows still NOT covered by a guide:**
- **D4 (full) — Open PR end-to-end**: the trigger is captured in guide 02; the worker's GitHub interaction requires a registered config-repo + real PAT, which is operator-credential territory.
+- **UBI judgment generation**: guide 09 covers the LLM-as-judge path; the UBI (click/dwell) alternative needs a cluster with captured behavior traffic, which is operator-data territory. See the [UBI judgment-generation runbook](../03_runbooks/ubi-judgment-generation.md) and the optional UBI upgrade step in [`tutorial-first-study.md`](tutorial-first-study.md).
### Adding or regenerating a guide
diff --git a/docs/08_guides/llm-endpoint-setup.md b/docs/08_guides/llm-endpoint-setup.md
index a15f1caa..28a94b0c 100644
--- a/docs/08_guides/llm-endpoint-setup.md
+++ b/docs/08_guides/llm-endpoint-setup.md
@@ -7,6 +7,8 @@ RelyLoop talks to **any OpenAI-compatible endpoint** via a single environment va
This guide walks through the most common configurations side by side.
+> **Not the only judgment source.** This guide configures the LLM endpoint that powers LLM-as-judge generation, the digest narrative, and the chat agent. If your cluster has captured click/dwell traffic, User Behavior Insights (UBI) generates judgments from real user behavior with no external LLM — see the [UBI judgment-generation runbook](../03_runbooks/ubi-judgment-generation.md).
+
## The pattern in one paragraph
RelyLoop uses the `openai` Python SDK pointed at whatever URL you set in `OPENAI_BASE_URL`. The SDK speaks the OpenAI Chat Completions wire protocol (`POST /v1/chat/completions`). Anything that serves that protocol — OpenAI itself, Azure OpenAI's OpenAI-compatible mode, Ollama, LM Studio, vLLM, HuggingFace TGI, OpenRouter, LiteLLM proxy, or any other compatible server — works without changing a line of RelyLoop code. The API key is mounted at `./secrets/openai_key` (or whatever path you point `OPENAI_API_KEY_FILE` at) per [`CLAUDE.md`](../../CLAUDE.md) Absolute Rule #2 on secrets handling.
diff --git a/docs/08_guides/quick-tour.md b/docs/08_guides/quick-tour.md
index f5c46e51..138d8459 100644
--- a/docs/08_guides/quick-tour.md
+++ b/docs/08_guides/quick-tour.md
@@ -359,10 +359,12 @@ Three points to land, in order:
optimization engine doesn't care which backend implements the search;
it tunes whatever knobs the adapter exposes. Your existing Fusion
query pipelines map directly.
-2. **Real signals next release.** MVP1.5 ("Real Signals") replaces
- LLM-as-judge with UBI click/dwell data as a first-class judgment
- source. Your existing Fusion Signals capture maps directly — you
- grade studies against your users' real behavior, not an LLM's guess.
+2. **Real signals, available now.** User Behavior Insights (UBI) is a
+ first-class judgment source alongside LLM-as-judge: when your cluster
+ has captured click/dwell traffic, you grade studies against your
+ users' real behavior instead of an LLM's guess. UBI is engine-neutral
+ (Elasticsearch, OpenSearch, Solr) and complements the LLM path — the
+ demo above uses LLM because it needs no traffic to run.
3. **Open source, self-hosted.** Apache 2.0. Run it on a laptop, run it
on your own infra. The PR-based ship workflow keeps your existing CI
and existing reviewers in charge of production.
diff --git a/docs/08_guides/workflows-overview.md b/docs/08_guides/workflows-overview.md
index b4c29dbf..0e3fa9b9 100644
--- a/docs/08_guides/workflows-overview.md
+++ b/docs/08_guides/workflows-overview.md
@@ -101,6 +101,14 @@ These define *what you're tuning* (the query template knobs) and *what good look
- **Side effects:** Creates a `judgment_lists` row (status=`generating`), enqueues the worker. Cost-gated by daily OpenAI budget. ~$0.01–$0.05 with `gpt-4o-mini` on the 48-query tutorial set.
- **Auto-recovery:** If the worker crashes mid-list, the boot-time sweep + the every-15-minute `resume_stuck_judgment_lists` cron re-enqueue stuck lists (capped at 24 attempts/day to prevent infinite loops). See [`feat_judgments_periodic_resume_sweep`](../00_overview/implemented_features/2026_05_14_feat_judgments_periodic_resume_sweep/).
+#### B4a. Generate judgments via UBI (User Behavior Insights)
+- **Solves:** Grading study trials against *real user behavior* (clicks, dwell time) instead of an LLM's prediction. Complementary to B4 — UBI needs a cluster that has captured click/dwell traffic; LLM works with zero traffic. Engine-neutral: works on Elasticsearch, OpenSearch, and Solr from the same `ubi_queries` + `ubi_events` schema.
+- **Readiness probe:** `GET /api/v1/clusters/{id}/ubi-readiness` reports a rung (`rung_0`..`rung_3`) from the captured event volume. The "Generate judgments" modal surfaces an on-ramp nudge below `rung_3` and a sparse-data card when coverage is thin.
+- **How (UI):** [`/query-sets/[id]`](../../ui/src/app/query-sets/[id]/page.tsx) → "Generate judgments" modal → **Method** picker: `UBI (click-through)`, `UBI (dwell-time)`, or `Hybrid UBI + LLM` (UBI for high-signal pairs, LLM fills the long tail). The detail page then shows a value-delta card comparing the UBI ratings against any prior LLM list.
+- **How (API):** `POST /api/v1/judgments/generate-from-ubi` returns `202 ACCEPTED` with a `judgment_list_id`. The `generate_judgments_from_ubi` worker runs the converter; the chat agent exposes the same path via the `generate_judgments_from_ubi` tool.
+- **Side effects:** Creates a `judgment_lists` row (status=`generating`, `source='click'`). Read-only against the cluster — RelyLoop never writes to UBI indices. Pure-UBI converters cost no LLM tokens; hybrid mode costs match B4 for the filled pairs only.
+- **Runbook:** [`../03_runbooks/ubi-judgment-generation.md`](../03_runbooks/ubi-judgment-generation.md) for per-engine UBI capture setup + converter selection.
+
#### B5. Import pre-curated judgments (tutorial / sideload path)
- **Solves:** Bypassing LLM generation when you already have human-labeled judgments (e.g., from Amazon ESCI, a previous tool, or hand-curation).
- **How (API):** `POST /api/v1/judgment-lists/import` — bulk-insert with strict validation (every `query_id` must exist in the query set, duplicate `(query_id, doc_id)` rejected). Sets status=`complete` immediately, no LLM call.
@@ -108,7 +116,7 @@ These define *what you're tuning* (the query template knobs) and *what good look
#### B6. Review and override individual judgments
- **Solves:** The LLM gets things wrong — sometimes spectacularly. Engineers need to inspect the (query, doc, rating, LLM-reasoning) tuples and override the bad ones, *without* re-running generation.
-- **How (UI):** [`/judgments/[id]`](../../ui/src/app/judgments/[id]/page.tsx) — table with a source filter (`llm` / `human`), inline override.
+- **How (UI):** [`/judgments/[id]`](../../ui/src/app/judgments/[id]/page.tsx) — table with a source filter (`llm` / `human` / `click`), inline override.
- **How (API):** `PATCH /api/v1/judgment-lists/{id}/judgments/{judgment_id}` — UPSERT semantics; the human override coexists with the LLM judgment and supersedes it for scoring.
#### B7. Calibrate LLM judgments against human ground truth
diff --git a/state.md b/state.md
index 864949f1..b123ce3f 100644
--- a/state.md
+++ b/state.md
@@ -14,8 +14,8 @@ MVP1 (v0.1) **shipped** — all six differentiators live (Bayesian/TPE optimizer
## Current branch / execution context
-- **Branch:** `feat/ubi-judgments`. `feat_ubi_judgments` shipped end-to-end (11 of 13 stories: migration + domain + reader + readiness service + dispatcher + 2 endpoints + worker + agent tool + frontend dialog + value-delta cards + runbook/FAQ/data-model docs). Stories 5.2 (E2E suite) + the 7-of-10 Story 5.1 doc artifacts deferred to `chore_ubi_e2e_suite` + `chore_ubi_docs_followup`; DB-backed integration tests deferred to `chore_ubi_integration_tests`. All three follow-up idea files committed.
-- **Active feature:** `feat_ubi_judgments` PR pending push.
+- **Branch:** none active — `feat_ubi_judgments` (PR #317) squash-merged to `main` 2026-05-29; folder finalized to `implemented_features/2026_05_29_feat_ubi_judgments/` via the `docs/finalize-ubi-judgments` docs PR. See `/pipeline status` for the next MVP2 item.
+- **Active feature:** none in flight.
- **Alembic head:** `0021_judgment_lists_generation_params` (added by `feat_ubi_judgments` Story 1.1 — JSONB column for UBI worker resume payload).
- **Python:** 3.13. **Frontend stack:** Next 16 (App Router + Turbopack), React 19, Tailwind 4 (CSS-first), Vitest 4, ESLint 9 (flat), TypeScript 6, Playwright (chromium, single worker) for E2E.
- **Coverage gates:** backend 80% (`fail_under` in pyproject), UI vitest + tsc + ESLint + Next build, plus a full-stack smoke E2E job. Live pass counts: see the latest `pr.yml` run (the historical per-feature counts moved to `state_history.md`).
@@ -24,24 +24,24 @@ MVP1 (v0.1) **shipped** — all six differentiators live (Bayesian/TPE optimizer
Detail + reasoning for each is in [`state_history.md`](state_history.md).
-- **2026-05-29** — `feat_ubi_judgments` (PR pending push). MVP2 second feature: engine-neutral User Behavior Insights judgment generation. Ships 11 of 13 planned stories: migration 0021 (judgment_lists.generation_params JSONB) → domain/ubi/ pure-domain library (FeatureVec, async SignalsConverter Protocol + 3 impls, position-bias prior) → UbiReader (engine-neutral two-index scan + client-side join, no new adapter method) → ubi_readiness classifier (rung_0..rung_3, 60s Redis cache) → start_ubi_judgment_generation dispatcher (refactor extracts 5 shared helpers; LLM dispatcher parity preserved by all 12 existing tests) → 5 new wire Literals + _SourceBreakdown three-term evolution (FR-10) → GET /clusters/{id}/ubi-readiness + POST /judgments/generate-from-ubi endpoints → generate_judgments_from_ubi Arq worker with mapping_strategy + hybrid LLM-fill callback → 21st agent tool + orchestrator prompt update → frontend method picker dialog + on-ramp nudge + sparse-data card + value-delta + ambiguous-skip recovery cards → operator runbook + 3 FAQ entries + data-model patches. Test count: +110 new tests across unit (~95) + contract (~17) layers; all 1715 backend unit + 931 UI vitest green; mypy --strict clean across 501 files. Stories 5.2 (E2E) + remaining Story 5.1 docs + DB-backed integration tests deferred via 3 follow-up idea files (chore_ubi_e2e_suite, chore_ubi_docs_followup, chore_ubi_integration_tests).
+- **2026-05-29** — `feat_ubi_judgments` (PR #317, squash-merged). MVP2 second feature: engine-neutral User Behavior Insights judgment generation, shipped end-to-end (all 13 stories incl. E2E + DB-backed integration tests + operator docs — none deferred). Migration 0021 (judgment_lists.generation_params JSONB) → domain/ubi/ pure-domain library (FeatureVec, async SignalsConverter Protocol + 3 impls, position-bias prior) → UbiReader (engine-neutral two-index scan + client-side join, no new adapter method) → ubi_readiness classifier (rung_0..rung_3, 60s Redis cache) → start_ubi_judgment_generation dispatcher (refactor extracts 5 shared helpers; LLM dispatcher parity preserved by all 12 existing tests) → 5 new wire Literals + _SourceBreakdown three-term evolution (FR-10) → GET /clusters/{id}/ubi-readiness + POST /judgments/generate-from-ubi endpoints → generate_judgments_from_ubi Arq worker with mapping_strategy + hybrid LLM-fill callback → 21st agent tool + orchestrator prompt update → frontend method picker dialog + on-ramp nudge + sparse-data card + value-delta + ambiguous-skip recovery cards → operator runbook + 3 FAQ entries + data-model patches → 4 Playwright E2E specs (rung_0/rung_3/hybrid/source-filter) green against the live ES-backed stack. **Real-engine E2E caught a production bug**: UbiReader requested `size=50000 > ES index.max_result_window (10000)` → "all shards failed" swallowed by the adapter → spurious `UBI_INSUFFICIENT_DATA` on dense clusters; fixed by clamping both index scans to `ES_MAX_RESULT_WINDOW=10000` + regression guard. Cross-model review: 6 Gemini findings + 6 GPT-5.5 findings all adjudicated (fixed or documented as working-as-designed). Remaining follow-ups are pure deferrals, not gaps: `chore_ubi_reader_search_after_pagination` (P2, >10k-event clusters), `chore_ubi_hybrid_template_render` (P3, vestigial-template contract cleanup — current behavior correct per FR-2), `feat_demo_ubi_study_comparison` (P1, side-by-side UBI-vs-LLM demo study).
- **2026-05-29** — `feat_study_sub_warmup_guard` (PR #316). First MVP2 feature ships. Closes the Custom-mode sub-warmup gap left open by `chore_study_default_stop_conditions` (2026-05-23): adds a non-blocking inline amber warning to the create-study modal's Step 5 when operators enter `max_trials < 50` in Custom mode. Hoists the inline `50` at `optuna_runtime.py:154` to a module-level `STUDIES_TPE_WARMUP_FLOOR` constant; frontend `SUB_WARMUP_FLOOR` mirrors with the `// Values must match` discipline comment + value-lock unit test. 3 new backend pytest assertions (value lock + `floor-1=49` boundary + `floor=50` boundary using constant); 5 new vitest cases (AC-1..AC-4 + AC-6 submit-non-blocking). Single-phase per spec D-6; digest narrative routed to `feat_study_convergence_indicator`. Cross-model review: spec converged at cycle 3 (13/13 accepted), plan converged at cycle 3 (8 accepted + 1 rejected with counter-evidence), phase-gate converged at cycle 2 (clean), final GPT-5.5 + 1 Gemini finding accepted.
- **2026-05-29** — `docs: reclassify 2 deferred MVP1 items → 99_backlog/03_mvp3` (PR #310, docs-only). Empties `01_mvp1/` — MVP1 actionable backlog fully drained. `chore_demo_reseed_stale_recovery_atomic_cas` → `99_backlog/` (already Priority: Backlog); `infra_agent_sibling_worktree_isolation` → `99_backlog/` (phases 1+2 shipped, only phase3 remains, defer-until-incident). Dashboards regenerated.
- **2026-05-29** — `bug_smoke_studies_data_table_search_flake` (PR #308 + finalization #309). Hardened the flaky `studies-data-table.spec.ts:20` search-visibility assertion: scoped it to the `studies-table` element + 15s web-first timeout to ride out the debounce→refetch→render race on slow CI runners. e2e-only; no product change.
- **2026-05-29** — `ci(pr): SKIP_HEAVY_CI kill-switch` (PR #307, infra). Added an `if:` guard on the 5 `pr.yml` jobs over 1 min so a repo variable can skip them (temporary GitHub Actions budget measure). See the Active CI note above — variable currently set, auto-restores ~2026-06-01.
## In flight
-- `feat_ubi_judgments` (PR pending push). Ships the UBI substrate + API + worker + frontend + operator docs. Three follow-up idea files committed for the deferred sub-scope (E2E, doc completeness, DB-backed integration tests).
+- _None._ `feat_ubi_judgments` (PR #317) merged 2026-05-29; finalization docs PR (`docs/finalize-ubi-judgments`) in review.
## Queued (priority-ordered by dashboard / dep graph)
**Source of truth:** [`docs/00_overview/DASHBOARD.md`](docs/00_overview/DASHBOARD.md) + [`docs/00_overview/MVP1_DASHBOARD.md`](docs/00_overview/MVP1_DASHBOARD.md) (regenerated by the `mvp1-dashboard-regen` pre-commit hook). Run `/pipeline status` for the live view.
-**MVP1 backlog is fully drained** (`01_mvp1/` empty as of PR #310). The next stop is **MVP2 / v0.2 — "Three-Engine + Real Signals"**. The `02_mvp2/` bucket currently holds 11 folders (run `ls docs/00_overview/planned_features/02_mvp2/` for the live list):
+**MVP1 backlog is fully drained** (`01_mvp1/` empty as of PR #310). The next stop is **MVP2 / v0.2 — "Three-Engine + Real Signals"**. With `feat_ubi_judgments` merged, the `02_mvp2/` bucket holds 18 folders (run `ls docs/00_overview/planned_features/02_mvp2/` for the live list):
-- **Headliners:** `infra_adapter_solr` (Apache Solr adapter), `feat_ubi_judgments` (UBI judgment source), `feat_chat_last_message_preview`, `feat_fts_rank_ordering`.
-- **Bugs held for MVP2:** `bug_chat_long_conversation_truncation` (investigation `bug_fix.md` exists; pullable forward but deferred for scope discipline — latency-of-impact is zero today), `bug_webhook_concurrent_merge_race_timing_sensitive`.
-- **Chores:** `chore_auto_followup_parent_advisory_lock`, `chore_demo_seeding_integration_tests_rewrite`, `chore_studies_post_arq_spy_fixture`, `chore_template_library_expansion`, `infra_arq_subprocess_test`.
+- **Headliners:** `infra_adapter_solr` (Apache Solr adapter), `feat_fts_rank_ordering`, `feat_chat_last_message_preview`, `feat_overnight_autopilot`, `feat_query_normalization_tuning`, `feat_study_convergence_indicator`, `feat_demo_ubi_study_comparison` (side-by-side UBI-vs-LLM demo study, P1 — spun out of `feat_ubi_judgments`).
+- **Bugs held for MVP2:** `bug_chat_long_conversation_truncation` (investigation `bug_fix.md` exists; pullable forward but deferred for scope discipline — latency-of-impact is zero today), `bug_webhook_concurrent_merge_race_timing_sensitive`, `bug_seed_meaningful_demos_silent_bulk_errors`.
+- **Chores/infra:** `chore_auto_followup_parent_advisory_lock`, `chore_demo_seeding_integration_tests_rewrite`, `chore_studies_post_arq_spy_fixture`, `chore_template_library_expansion`, `chore_ubi_reader_search_after_pagination` (P2, search_after for >10k-event clusters — spun out of `feat_ubi_judgments`), `chore_ubi_hybrid_template_render` (P3, vestigial-template contract cleanup — spun out of `feat_ubi_judgments`), `infra_arq_subprocess_test`.
**Other buckets:** `03_mvp3/` (Observable — includes `infra_optuna_orphan_reaper`, deferred from MVP1 per spec §11 operational tolerance), `04_ga/`, `99_backlog/` (4 defer-until-incident items), `00_unsure/` (`bug_seed_meaningful_demos_silent_bulk_errors`).