From 959f998f81f595f4dd6b312c1247c6f1c70e0b51 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=EC=B5=9C=EC=98=81=EA=B8=B0?= <dch0202@rsquare.co.kr>
Date: Tue, 28 Jul 2026 15:10:25 +0900
Subject: [PATCH 1/3] =?UTF-8?q?knowledge:=20ingest=204=20verified=20insigh?=
 =?UTF-8?q?t(s)=20=E2=80=94=20kubelet=20resource=20metrics,=20cgroupns=20v?=
 =?UTF-8?q?isibility,=20LLM=20completion=20validation,=20gateway=20alias?=
 =?UTF-8?q?=20defaults?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .dev-loop/INGEST_REPORT.md                    | 54 +++++++------
 INDEX.md                                      |  2 +-
 log.md                                        |  2 +
 .../llm/completion-response-validation.md     | 65 +++++++++++++++
 .../llm/gateway-model-alias-defaults.md       | 59 ++++++++++++++
 .../reliability/timeouts-and-retries.md       |  2 +-
 wiki/backend/index.md                         |  9 ++-
 .../config/environment-config.md              |  2 +-
 .../containers/host-cgroup-visibility.md      | 67 ++++++++++++++++
 .../containers/resource-limits-and-probes.md  |  2 +-
 wiki/infrastructure/index.md                  |  2 +
 .../observability/logs-metrics-signals.md     |  2 +-
 .../missing-container-metrics.md              | 80 +++++++++++++++++++
 wiki/qa/process/release-gates.md              |  2 +-
 14 files changed, 320 insertions(+), 30 deletions(-)
 create mode 100644 wiki/backend/common/llm/completion-response-validation.md
 create mode 100644 wiki/backend/common/llm/gateway-model-alias-defaults.md
 create mode 100644 wiki/infrastructure/containers/host-cgroup-visibility.md
 create mode 100644 wiki/infrastructure/observability/missing-container-metrics.md

diff --git a/.dev-loop/INGEST_REPORT.md b/.dev-loop/INGEST_REPORT.md
index 2ff3871..288c223 100644
--- a/.dev-loop/INGEST_REPORT.md
+++ b/.dev-loop/INGEST_REPORT.md
@@ -1,35 +1,43 @@
-# Knowledge flush — 1 insight
-
-Source: RNR-3440 (사내 잠재매물 주간 추출 스크립트 메모리 피크 저감). Candidate:
-"QueryPie 프록시 경유로 대용량 결과를 스트리밍할 때 server-side named cursor 대신
-일반 커서 + `fetchmany` + openpyxl `write_only`."
+# Knowledge flush — 4 insight(s)
 
 ## Verified best-practice
 
-**Claim 1 — psycopg2 server-side (named) cursor requires a transaction; fails under autocommit.**
-- Source: psycopg2 usage docs — `https://github.com/psycopg/psycopg2/blob/master/doc/src/usage.rst` (via Context7). Quote: "Named cursors are typically created 'WITHOUT HOLD', meaning they exist only within the current transaction. Attempting to fetch from them after a commit or in autocommit mode raises an exception."
-- Matches my live repro (`can't use a named cursor outside of transactions`). → **verified**
+**1. Container metrics empty while kubelet targets report up (OrbStack embedded k8s)**
+Claim: when cAdvisor emits no `container_*` series, scrape kubelet `/metrics/resource` (kps values `kubelet.serviceMonitor.resource: true` + `resourcePath: "/metrics/resource"`), and expect bundled mixin dashboards to stay empty because of their `image!=""` filter.
+Sources checked: Kubernetes resource-metrics-pipeline docs (fetched); kubelet source `pkg/kubelet/metrics/collectors/resource_metrics.go` (fetched — confirms the exact two series); kube-prometheus-stack `values.yaml` on main (fetched — `resource: false`, `resourcePath: "/metrics/resource/v1alpha1"` is STILL the default today, with the in-file "renamed in k8s 1.18" comment, so the override is required, not historical); kubernetes-mixin `dashboards/resources/queries/pod.libsonnet` + `rules/apps.libsonnet` (fetched — `image!=""` in queries and recording rules); orbstack/orbstack#2217 (corroborates the symptom class on OrbStack).
+Verification outcome: 3 of 4 sub-claims confirmed against primary sources; the machine_*-only cAdvisor observation itself is a field measurement (2026-07-15, `kubectl get --raw .../metrics/cadvisor` → 0 `container_` series). **Confidence: verified** (field-only detail labeled as such in the page body). One correction folded in: the chart's v1alpha1 default is current, not merely historical.
 
-**Claim 2 — a client-side (default) cursor pulls the whole result set to the client on execute; `fetchmany` only caps the Python-list explosion.**
-- Source: psycopg2 cursor/usage docs + FAQ (named-cursor advantage = "data is fetched in chunks … minimal client memory"). By contrast the default cursor buffers the full result in libpq. → **verified**
+**2. OpenAI-compatible completion validation (reasoning models, finish_reason=length, empty content)**
+Claim: gate on `finish_reason == "length"` and blank `content` before using LLM output; use the reasoning field to disambiguate "reasoning consumed the budget".
+Sources checked: OpenAI openapi.yaml (fetched — `length` = truncated by requested token limit, verbatim); OpenAI reasoning guide (fetched — reasoning can consume the whole output budget before any visible output); LiteLLM reasoning_content docs (fetched — normalizes to `message.reasoning_content`); vLLM reasoning outputs docs (fetched — field RENAMED `reasoning_content` → `reasoning`).
+Verification outcome: core semantics verified; two provider-dependent caveats surfaced by research and added to the page: (a) vLLM's field rename — the page tells readers to check both spellings; (b) DeepSeek first-party API gives CoT a separate budget, so the empty-content failure mode is OpenAI-o-series/vLLM-style budgeting, not universal — added as an edge case. The exact observed combination (HTTP 200 + `length` + 0-char content + 8,173-char reasoning_content) is the session's own LiteLLM measurement (2026-07-28). **Confidence: field-tested.**
 
-**Claim 3 — openpyxl `write_only` gives near-constant memory (<10 MB); one save only; lxml is for serialization speed, not the memory saving.**
-- Source: openpyxl Optimised Modes — `https://openpyxl.readthedocs.io/en/stable/optimized.html` (via WebSearch). "keeping memory usage under 10Mb"; "A write-only workbook can only be saved once"; "make sure you have lxml installed" for large dumps (speed).
-- This **corrects** the raw candidate's "lxml unnecessary" → precise form: unnecessary *for the memory win*, recommended *for large-dump speed*. Confirmed by my server test (write-only worked with lxml absent). → **verified**
+**3. Code defaults naming gateway-resolved model aliases (re-verify at review/merge)**
+Claim: defaults resolved by an external serving catalog can die between PR verification and merge while code/tests/builds stay green; re-query the live catalog (`GET /v1/models`) at review/merge time and fail-fast-validate load-bearing defaults at startup.
+Sources checked: OpenAI models-list API reference (fetched — lists currently available model IDs); LiteLLM model_discovery docs (fetched — `/v1/models` reflects configured `model_list`; live upstream checking is opt-in, which strengthens the page's "confirm with one real call" row); vLLM quickstart (fetched — `/v1/models` on the OpenAI-compatible server); GitHub merge-queue docs (fetched — the staleness rationale: validation done earlier must be redone against current state); fail-fast startup-validation references (confirmed via search).
+Verification outcome: the endpoint claim fully verified; the practice itself is a field lesson (PR review 2026-07-28: default summarization alias removed from the LiteLLM catalog between PR verification and review → `/chat/completions` 400 while the PR's e2e evidence was genuinely true at write time) supported by general drift/staleness references, not a documented named practice. **Confidence: field-tested.**
 
-**Claim 4 — QueryPie blocks `BEGIN`, so server-side cursor is impossible there.**
-- Environment-specific, no external source. Live repro in gui context: `autocommit=False` + named cursor → `[ENGINE] No permission to execute BEGIN statement`. → **field-tested**. Generalized in the page to "a read-only access-control proxy that blocks transaction control", with QueryPie as the concrete example (not a product-specific page).
-- Memory figure 838 MB → 38 MB (300k synthetic rows) is my RNR-3440 measurement (`ru_maxrss`, separate processes).
+**4. Reading other pods' cgroup v2 stats from inside a container**
+Claim: mounting host `/sys/fs/cgroup` is insufficient — you need the host cgroup namespace (`docker --cgroupns=host`; k8s privileged or hostPID + `nsenter -t 1 -C`).
+Sources checked: cgroup_namespaces(7) man page (fetched); docker run reference (fetched — `--cgroupns=host`; private is the cgroup v2 default); nsenter(1) (fetched — `-C/--cgroup`); kubernetes/kubernetes#103363 (fetched — no first-class pod hostCgroup option; nsenter workaround as used by Cilium).
+Verification outcome: directive verified against docs + the session's own A/B reproduction (2026-07-22 OrbStack: identical `-v` mount, `kubepods` absent without / fully present with `--cgroupns=host`). Research CORRECTED the harvested mechanism: the cgroupfs view is fixed at MOUNT time by the mounting process's cgroup namespace (the runtime creates the container's cgroup mount inside the private ns) — not dynamically filtered per reader; the page states the corrected mechanism. Research bonus added: `nsdelegate` write-restriction edge case (files visible, writes EPERM — a distinct failure mode). **Confidence: field-tested.**
 
 ## Existing-layer check
 
-- Pages read: `databases/index.md`, `databases/query-optimization/keyset-pagination.md`, `backend/python/index.md`.
-- Overlap: keyset-pagination is the nearest neighbor (both handle large result sets) but a **distinct** topic — pagination splits the read into many bounded queries; this page streams a *single* query's result in chunks. Not a duplicate → new page + **bidirectional `related` link** added to both.
-- backend/python has no DB-cursor page; the psycopg2/openpyxl specifics live as concrete examples inside the databases page rather than a separate python page (no duplication).
-- Conflicts: none found.
+Read: root `INDEX.md`; domain indexes for `infrastructure` and `backend`; candidate-overlap pages in full — `infrastructure/observability/logs-metrics-signals.md`, `infrastructure/containers/resource-limits-and-probes.md`, `backend/common/reliability/timeouts-and-retries.md`, `infrastructure/config/environment-config.md`, `qa/process/release-gates.md` (+ skimmed `qa/process/post-release-verification.md`). Repo-wide grep for `reasoning_content|cadvisor|cgroup|/metrics/resource|v1/models|finish_reason` found no page owning any of these topics (hits were incidental substrings).
+
+- **Overlaps found:** none that owns the new triggers. `logs-metrics-signals` owns instrumenting a service (emitting signals), not collecting cluster metrics; `resource-limits-and-probes` owns manifests/limits (it *consumes* `container_memory_working_set_bytes`, hence a related-link, not a merge); `timeouts-and-retries` owns transport-level outbound failures, not semantic response validation; `environment-config` owns config shape/startup validation (its "no dev-friendly defaults" rule is adjacent to insight 3, but the trigger — external catalog drift at review time — is new); `release-gates` owns the release checklist process.
+- **Merged vs created:** 0 merged, 4 new pages created (all four triggers are new).
+- **Conflicts flagged:** none — no existing directive contradicts the new pages.
+- **Related links added (both ways):** `missing-container-metrics` ↔ `logs-metrics-signals`, ↔ `resource-limits-and-probes`; `host-cgroup-visibility` ↔ `resource-limits-and-probes`, ↔ `missing-container-metrics`; `completion-response-validation` ↔ `gateway-model-alias-defaults`, ↔ `timeouts-and-retries`; `gateway-model-alias-defaults` ↔ `environment-config`, ↔ `release-gates`.
 
 ## Routing decision
 
-- Target: **`databases/query-optimization/streaming-large-result-sets.md`** (new page).
-- Category `query-optimization` fits (memory-bounding how a query's result is pulled into the app is query-execution optimization); no new category needed.
-- Registered in `databases/index.md` (query-optimization section) and appended to `log.md`.
+| Insight | Target | Page |
+|---------|--------|------|
+| 1 (kubelet metrics) | `infrastructure/observability` (existing category) | `missing-container-metrics.md` — new page, new trigger |
+| 4 (cgroup namespace) | `infrastructure/containers` (existing category) | `host-cgroup-visibility.md` — new page, new trigger |
+| 2 (completion validation) | `backend/common/llm` (**new category**) | `completion-response-validation.md` |
+| 3 (catalog-resolved defaults) | `backend/common/llm` (**new category**) | `gateway-model-alias-defaults.md` |
+
+**New category justification (`backend/common/llm`):** both insights concern consuming OpenAI-compatible LLM gateways from server-side code. Existing backend/common categories don't cover the trigger: `api-design` owns designing *your own* API; `reliability` owns transport-level outbound failure handling (timeouts/retries/backoff), while these pages gate on *semantic* response validity (finish_reason/reasoning fields) and external-catalog identifier drift — HTTP-success-but-unusable-output is outside its tables. LLM-gateway consumption is a recurring concern with its own vocabulary, so a dedicated category beats stretching `reliability`. Insight 3 generalizes beyond LLMs (any externally-resolved alias); that generality is stated in the page's "When this applies" while the page stays anchored to its concrete, evidenced instance. Plumbing updated: `backend/index.md` (new `### llm` section + concern list), root `INDEX.md` backend route line, `infrastructure/index.md` (two new load-when rows), `log.md` (two ingest entries).
diff --git a/INDEX.md b/INDEX.md
index 05789bd..43fc87f 100644
--- a/INDEX.md
+++ b/INDEX.md
@@ -10,7 +10,7 @@ follow the cross-pointers in their index or take the next matching seeded domain
 | Domain | Status | Route here when |
 |--------|--------|-----------------|
 | [databases](wiki/databases/index.md) | **seeded** | Designing schemas/tables/keys, choosing or evaluating indexes, writing or optimizing queries, choosing transaction/isolation behavior |
-| [backend](wiki/backend/index.md) | **seeded** | Server-side application code — language-agnostic (`common/`: API contracts, idempotency, JWT, timeouts/retries, caching, jobs, transactions in app code, shared state/pools, errors) plus stack subtrees: `java/` (JPA, Spring proxies, JVM threads/memory), `node/` (event loop, promises, runtime validation, shutdown), `python/` (GIL/asyncio, pydantic, WSGI/ASGI workers, language traps) |
+| [backend](wiki/backend/index.md) | **seeded** | Server-side application code — language-agnostic (`common/`: API contracts, idempotency, JWT, timeouts/retries, caching, jobs, transactions in app code, shared state/pools, errors, LLM/gateway completions) plus stack subtrees: `java/` (JPA, Spring proxies, JVM threads/memory), `node/` (event loop, promises, runtime validation, shutdown), `python/` (GIL/asyncio, pydantic, WSGI/ASGI workers, language traps) |
 | [frontend](wiki/frontend/index.md) | **seeded** | Web UI code: state placement, rendering performance, in-UI data fetching (races, infinite scroll), auth token handling, forms, XSS-safe output, accessibility |
 | [infrastructure](wiki/infrastructure/index.md) | **seeded** | CI/CD pipelines, secrets in build/deploy, container image builds, rollout/rollback strategy, observability (logs/metrics/alerting) |
 | [testing](wiki/testing/index.md) | **seeded** | Writing or structuring automated tests: level choice, cases/assertions, test data, mock decisions, flaky tests (release-process quality → qa) |
diff --git a/log.md b/log.md
index 587a5dc..576470e 100644
--- a/log.md
+++ b/log.md
@@ -34,3 +34,5 @@ Append-only. Format: `## [YYYY-MM-DD] <ingest|revise|lint|gap|contradiction|drif
 ## [2026-07-12] revise | security/secrets-in-code +1 edge case: third-party HTTP client (httpx/requests) logs the full request URL — including a query-param API key — at INFO, so root/DEBUG logging leaks it; keep the client logger above INFO. Found when a standalone sync process set logging.basicConfig(INFO) and httpx wrote the data.go.kr serviceKey to the log file. last_verified bumped to 2026-07-12.
 ## [2026-07-13] ingest | databases +1 (query-optimization): streaming-large-result-sets — memory-bounded export of a huge single-query result. Client-side cursor pulls the whole set to libpq on execute (fetchmany caps only the Python-list explosion); only a server-side/named cursor truly streams but needs a transaction, so it fails under autocommit or a proxy that blocks BEGIN → fall back to client-side fetchmany + disk spool + openpyxl write_only (measured 300k rows 838MB→38MB). Derived from RNR-3440 (potential-listing weekly extract memory peak); QueryPie BEGIN-block generalized to "read-only access proxy", field-tested. Sources: psycopg2 usage/cursor docs (named-cursor WITHOUT HOLD + autocommit exception), openpyxl optimized-modes (write-only near-constant memory, lxml=speed-not-memory).
 ## [2026-07-23] ingest | databases +2: schema-design/online-schema-changes (ACCESS EXCLUSIVE lock avoidance — non-volatile default fast path, ADD CONSTRAINT NOT VALID + VALIDATE at SHARE UPDATE EXCLUSIVE, CHECK-NOT-NULL trick, CREATE INDEX CONCURRENTLY, expand-and-contract to decouple DB migration from app deploy, lock_timeout for lock-queue pile-up) + operations/autovacuum-and-wraparound (NEW category operations: per-table scale_factor/cost_limit tuning for hot tables, age(datfrozenxid)/relfrozenxid + n_dead_tup monitoring, wraparound read-only cliff and superuser VACUUM recovery, VACUUM FULL vs pg_repack). Derived from the Hatchet "Postgres survival guide"; both cross-checked against PostgreSQL official docs (sql-altertable, routine-vacuuming).
+## [2026-07-28] ingest | backend +2 (new category common/llm): completion-response-validation (gate on finish_reason=length + blank content; reasoning_content/reasoning disambiguation, provider budget caveats — DeepSeek separate CoT budget, vLLM field rename) + gateway-model-alias-defaults (re-query the live catalog `GET /v1/models` at review/merge time; startup fail-fast for load-bearing defaults). Derived from a meeting-summary LiteLLM integration (measured 200/length/empty-content/8,173-char reasoning_content; default alias dead by review time). finish_reason semantics + /v1/models verified against OpenAI spec/LiteLLM/vLLM docs; both pages field-tested.
+## [2026-07-28] ingest | infrastructure +2: observability/missing-container-metrics (verified — container_* absent while kubelet targets up; kubelet /metrics/resource scrape, kps resourcePath default STILL v1alpha1 per chart values.yaml, kubernetes-mixin image!="" filter excludes resource-endpoint series) + containers/host-cgroup-visibility (field-tested — mounted /sys/fs/cgroup shows namespaced view; docker --cgroupns=host / hostPID+nsenter -t 1 -C; mount-time cgroupns semantics per cgroup_namespaces(7), nsdelegate write caveat). Both from mac-server observability work (OrbStack embedded k8s measurements 07-15/07-22).
diff --git a/wiki/backend/common/llm/completion-response-validation.md b/wiki/backend/common/llm/completion-response-validation.md
new file mode 100644
index 0000000..48e3ef7
--- /dev/null
+++ b/wiki/backend/common/llm/completion-response-validation.md
@@ -0,0 +1,65 @@
+---
+id: backend-common-llm-completion-response-validation
+domain: backend
+category: llm
+applies_to: [general]
+confidence: field-tested
+sources:
+  - https://raw.githubusercontent.com/openai/openai-openapi/master/openapi.yaml
+  - https://developers.openai.com/api/docs/guides/reasoning
+  - https://docs.litellm.ai/docs/reasoning_content
+  - https://docs.vllm.ai/en/latest/features/reasoning_outputs.html
+last_verified: 2026-07-28
+related: [backend-common-llm-gateway-model-alias-defaults, backend-common-reliability-timeouts-and-retries]
+---
+
+# Validating an OpenAI-Compatible Completion Before Using Its Output
+
+## When this applies
+
+Your code consumes `/chat/completions` responses as a final artifact (a
+summary, a document, a message) and the routed model can change under you —
+a gateway (LiteLLM, vLLM) resolves the model from config, or reasoning-family
+models may be served on the same alias.
+
+## Do this
+
+Validate the response before writing the output anywhere downstream; HTTP 200
+does not mean the text is usable:
+
+| Check | On failure |
+|-------|-----------|
+| `finish_reason == "length"` | Treat as failure or explicit truncation handling — per the OpenAI spec, `length` means generation was cut off by the request's token limit, so the text is incomplete even when non-empty |
+| `content` empty or blank | Treat as failure — never register/notify with empty output |
+| `content` empty AND reasoning field populated | Emit the precise diagnostic "reasoning model consumed the completion budget before producing an answer" — raise max output tokens or route to a non-reasoning model |
+
+Mechanism: reasoning-family models stream chain-of-thought into a separate
+field (`message.reasoning_content` on DeepSeek-style APIs and LiteLLM;
+renamed to `message.reasoning` on current vLLM) and, on OpenAI o-series and
+vLLM-style servers, that reasoning spends the same output-token budget as the
+answer. When the budget runs out mid-reasoning you get HTTP 200,
+`finish_reason: "length"`, an empty `content`, and a large reasoning field —
+a client that reads only `content` passes empty text downstream as success.
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| Field is absent on your gateway | Check both spellings: `reasoning_content` (DeepSeek convention, LiteLLM-normalized) and `reasoning` (current vLLM) |
+| DeepSeek first-party API | CoT has its own separate budget and does not consume `max_tokens` — the empty-content-from-reasoning failure mode applies to OpenAI o-series/vLLM-style budgeting, not there; keep the finish_reason and empty-content checks anyway |
+| Retrying after a `length` failure | Raising max tokens is the fix only when reasoning ate the budget; for genuinely long outputs, chunk the task instead — blind retry with identical parameters reproduces the same truncation ([backend-common-reliability-timeouts-and-retries]: never retry the same failing bytes) |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Treat HTTP 200 from `/chat/completions` as success | Gate on `finish_reason` and non-blank `content` before using the text | Measured in the field (LiteLLM gateway, 2026-07-28): a reasoning-routed alias returned 200 with `finish_reason=length`, 0-char `content`, 8,173-char `reasoning_content` — the empty artifact flowed to document registration until the gate was added |
+| Read only `message.content` from the response | Also read the reasoning field to distinguish "model produced nothing" from "reasoning consumed the budget" | The two failures need opposite fixes (model/prompt problem vs token-budget/routing problem) |
+
+## Sources
+
+- https://raw.githubusercontent.com/openai/openai-openapi/master/openapi.yaml — `finish_reason` semantics: `length` = "the maximum number of tokens specified in the request was reached"
+- https://developers.openai.com/api/docs/guides/reasoning — reasoning can consume the entire output budget before any visible output, incurring cost with no visible response
+- https://docs.litellm.ai/docs/reasoning_content — LiteLLM normalizes provider reasoning output to `message.reasoning_content`
+- https://docs.vllm.ai/en/latest/features/reasoning_outputs.html — vLLM reasoning outputs; field renamed `reasoning_content` → `reasoning`
+- Field context: 2026-07-28 LiteLLM (internal dgx) measurement — same 9,317-token prompt: reasoning alias → 200/`length`/empty content/8,173-char reasoning_content; non-reasoning alias → `stop`/2,418-char complete answer
diff --git a/wiki/backend/common/llm/gateway-model-alias-defaults.md b/wiki/backend/common/llm/gateway-model-alias-defaults.md
new file mode 100644
index 0000000..5ba9fec
--- /dev/null
+++ b/wiki/backend/common/llm/gateway-model-alias-defaults.md
@@ -0,0 +1,59 @@
+---
+id: backend-common-llm-gateway-model-alias-defaults
+domain: backend
+category: llm
+applies_to: [general]
+confidence: field-tested
+sources:
+  - https://developers.openai.com/api/reference/resources/models/methods/list
+  - https://docs.litellm.ai/docs/proxy/model_discovery
+  - https://docs.vllm.ai/en/latest/getting_started/quickstart/
+  - https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/configuring-pull-request-merges/managing-a-merge-queue
+last_verified: 2026-07-28
+related: [backend-common-llm-completion-response-validation, infrastructure-config-environment-config, qa-process-release-gates]
+---
+
+# Code Defaults That Name Gateway-Resolved Model Aliases
+
+## When this applies
+
+A default value in code (a default model name, gateway route alias, or
+similar externally-resolved identifier) is resolved by a serving catalog that
+changes outside this repository — an LLM gateway (LiteLLM, vLLM), an API
+gateway route table, a model registry. Especially when a PR touching such a
+default claims "end-to-end verified".
+
+## Do this
+
+| Case | Do |
+|------|----|
+| Reviewing or merging a PR whose default names a catalog-resolved alias | Re-query the live catalog at review/merge time — `GET /v1/models` on any OpenAI-compatible gateway lists the currently served model IDs — and confirm the default still resolves. Treat the PR body's verification evidence as a statement about a past state of the catalog, not the present |
+| The default is load-bearing in production | Also validate it at process startup (resolve the alias once, crash loud on failure) so a catalog change becomes a failed deploy instead of silent runtime 400s — startup-validation mechanics: [infrastructure-config-environment-config] |
+| Review and merge are days apart | Re-check at merge time too — the same staleness that merge queues exist to catch (a PR validated earlier can be invalid against the current state) applies to external catalogs |
+| The gateway lists models from config, not live upstream state | Confirm with one real completion call, not just the list — e.g. LiteLLM `/v1/models` reflects its configured `model_list`; a listed alias can still fail upstream |
+
+The mechanism to internalize: the catalog is an external dependency that
+drifts independently of the repo. Code, unit tests, and builds exercise none
+of it, so every green signal stays green while the default is already dead.
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| No network access to the catalog from the review environment | Say so in the review instead of approving the claim — an unverifiable default gets flagged, not assumed alive |
+| The default has a fallback chain (try alias A, then B) | Verify every member of the chain; a dead primary that silently falls back hides the drift until the fallback dies too |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Approve a PR because its body shows a successful end-to-end run | Re-run the catalog check (`GET /v1/models` + one real call) at review time | The run proves the alias resolved then; the catalog may have changed since — observed in the field as a default model returning 400 days after a verified PR |
+| Hardcode a "known good" model alias as a constant | Keep it as validated config that is resolved at startup | A constant is invisible to config inventories and startup validation; drift surfaces as runtime errors |
+
+## Sources
+
+- https://developers.openai.com/api/reference/resources/models/methods/list — `GET /v1/models` lists currently available model IDs (OpenAI API semantics)
+- https://docs.litellm.ai/docs/proxy/model_discovery — LiteLLM serves `/v1/models` from its configured model list; live provider checking is opt-in
+- https://docs.vllm.ai/en/latest/getting_started/quickstart/ — vLLM OpenAI-compatible server exposes `/v1/models` for listing served models
+- https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/configuring-pull-request-merges/managing-a-merge-queue — validation done earlier on a PR must be redone against the latest state (staleness rationale)
+- Field context: 2026-07-28 review of an internal meeting-summary service — default summarization model alias removed from the LiteLLM catalog between PR verification and review; code/tests/build all green while `/chat/completions` returned 400
diff --git a/wiki/backend/common/reliability/timeouts-and-retries.md b/wiki/backend/common/reliability/timeouts-and-retries.md
index b64e69a..df130d6 100644
--- a/wiki/backend/common/reliability/timeouts-and-retries.md
+++ b/wiki/backend/common/reliability/timeouts-and-retries.md
@@ -9,7 +9,7 @@ sources:
   - https://sre.google/sre-book/addressing-cascading-failures/
   - https://aws.amazon.com/blogs/architecture/exponential-backoff-and-jitter/
 last_verified: 2026-07-10
-related: [backend-common-api-design-idempotency]
+related: [backend-common-api-design-idempotency, backend-common-llm-completion-response-validation]
 ---
 
 # Calling Another Service over the Network: Timeouts, Retries, Backoff
diff --git a/wiki/backend/index.md b/wiki/backend/index.md
index e991b04..2362d43 100644
--- a/wiki/backend/index.md
+++ b/wiki/backend/index.md
@@ -5,7 +5,7 @@ three stack subtrees — route by concern first, stack second:
 
 | Subtree | Route there when |
 |---------|------------------|
-| [common](#common-language-agnostic) (below) | The concern is language-agnostic: API contracts, idempotency, JWT issuance, outbound calls, caching, jobs, transactions in app code, shared state/pools, exception structure |
+| [common](#common-language-agnostic) (below) | The concern is language-agnostic: API contracts, idempotency, JWT issuance, outbound calls, caching, jobs, transactions in app code, shared state/pools, exception structure, LLM/gateway completion handling |
 | [java](java/index.md) | You are writing/reviewing JVM backend code (Java/Kotlin, Spring, JPA/Hibernate) and the concern is stack-specific: entity mapping, persistence context, proxy pitfalls, JVM threads/memory |
 | [node](node/index.md) | You are writing/reviewing Node.js/TypeScript backend code: event-loop blocking, promise error handling, runtime validation at boundaries, graceful shutdown |
 | [python](python/index.md) | You are writing/reviewing Python backend code: GIL/concurrency model, pydantic validation, WSGI/ASGI workers, language traps |
@@ -52,6 +52,13 @@ Match your situation to a "load when" line; load only matching pages.
 | [exception-handling](common/errors/exception-handling.md) | Writing a catch block or deciding where errors are handled/logged/translated in a service — catch placement, log-once, wrapping with cause preserved, typed results for expected outcomes; one fault producing duplicate alerts |
 | [async-failure-handling](common/errors/async-failure-handling.md) | Handing work to in-process async (@Async, unawaited futures/promises) — deciding fire-and-forget vs consumed future vs durable job; side effects silently never happening with no error logs; unobserved futures; async work enqueued inside a transaction |
 
+### llm
+
+| Page | Load when |
+|------|-----------|
+| [completion-response-validation](common/llm/completion-response-validation.md) | Consuming OpenAI-compatible `/chat/completions` output as a final artifact (summary, document, notification); LLM responses coming back empty or truncated while HTTP status is 200; a reasoning-family model may be routed onto the alias you call |
+| [gateway-model-alias-defaults](common/llm/gateway-model-alias-defaults.md) | A code default names a model alias or route resolved by an external serving catalog (LLM gateway, registry); reviewing/merging a PR that claims end-to-end verification of such a default; a previously-verified default starts failing with 400/404 |
+
 ### auth
 
 | Page | Load when |
diff --git a/wiki/infrastructure/config/environment-config.md b/wiki/infrastructure/config/environment-config.md
index 8964889..05817bd 100644
--- a/wiki/infrastructure/config/environment-config.md
+++ b/wiki/infrastructure/config/environment-config.md
@@ -9,7 +9,7 @@ sources:
   - https://12factor.net/build-release-run
   - https://12factor.net/dev-prod-parity
 last_verified: 2026-07-10
-related: [infrastructure-deploy-rollout-and-rollback, infrastructure-ci-cd-secrets-handling, backend-node-boundaries-runtime-validation, backend-python-boundaries-runtime-validation]
+related: [infrastructure-deploy-rollout-and-rollback, infrastructure-ci-cd-secrets-handling, backend-node-boundaries-runtime-validation, backend-python-boundaries-runtime-validation, backend-common-llm-gateway-model-alias-defaults]
 ---
 
 # Configuration That Differs Per Environment
diff --git a/wiki/infrastructure/containers/host-cgroup-visibility.md b/wiki/infrastructure/containers/host-cgroup-visibility.md
new file mode 100644
index 0000000..0f06bcc
--- /dev/null
+++ b/wiki/infrastructure/containers/host-cgroup-visibility.md
@@ -0,0 +1,67 @@
+---
+id: infrastructure-containers-host-cgroup-visibility
+domain: infrastructure
+category: containers
+applies_to: [kubernetes, docker, cgroup-v2]
+confidence: field-tested
+sources:
+  - https://man7.org/linux/man-pages/man7/cgroup_namespaces.7.html
+  - https://docs.docker.com/reference/cli/docker/container/run/
+  - https://man7.org/linux/man-pages/man1/nsenter.1.html
+  - https://github.com/kubernetes/kubernetes/issues/103363
+last_verified: 2026-07-28
+related: [infrastructure-containers-resource-limits-and-probes, infrastructure-observability-missing-container-metrics]
+---
+
+# Reading Other Pods' cgroup v2 Stats from Inside a Container
+
+## When this applies
+
+A container needs to read the host's full cgroup v2 hierarchy — other pods'
+CPU/memory/PSI stats under `kubepods/` — via hostPath or `docker -v` mounting
+of `/sys/fs/cgroup`. The mount succeeds but the `kubepods` subtree is missing,
+with no error (empty or partial listing, silently wrong results).
+
+## Do this
+
+Mounting the path is not sufficient; the cgroup namespace decides the view.
+On cgroup v2, container runtimes default to a **private cgroup namespace**
+(dockerd `--default-cgroupns-mode` default: `private`), and the cgroup2
+filesystem mounted for the container in that namespace shows only the
+container's own subtree — the host hierarchy is simply absent.
+
+| Case | Do |
+|------|----|
+| Docker/standalone container | Run with `--cgroupns=host` in addition to the `-v /sys/fs/cgroup:...` mount — the container then gets the host's cgroup namespace and the full hierarchy |
+| Kubernetes pod | There is no first-class host-cgroupns pod option (kubernetes#103363, open); use a privileged pod, or `hostPID: true` + `nsenter -t 1 -C` (`--cgroup`) to run the reading command inside PID 1's cgroup namespace |
+| Verifying before trusting the data | List the mount for `kubepods/` (or the equivalent top-level slice) explicitly; treat its absence as a namespace problem, never as "those pods have no stats" |
+
+Mechanism: cgroup namespaces virtualize the cgroup view — `/proc/[pid]/cgroup`
+paths and the cgroupfs contents are shown relative to the namespace's root,
+fixed by the cgroup namespace in effect when that cgroup2 filesystem mount was
+created (cgroup_namespaces(7)). The runtime creates the container's cgroup
+mount inside the private namespace, so what looks like "the host's
+/sys/fs/cgroup" is a namespaced view rooted at the container's own cgroup.
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| Files are visible but writes fail (`cgroup.procs`: no such file / EPERM) | A different restriction: the `nsdelegate` mount option (systemd default) blocks cross-namespace-boundary writes even when reads work — reading stats is fine, migrating processes is not |
+| Reading only the container's OWN limits/usage | No host namespace needed — the default namespaced view is exactly right for self-monitoring |
+| Sizing decisions based on the stats you read | Limits/QoS interpretation: [infrastructure-containers-resource-limits-and-probes] |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Conclude "cgroup data isn't there" from an empty mounted directory | Check the cgroup namespace first (`--cgroupns=host` / `nsenter -t 1 -C`), then re-list | Measured (OrbStack VM, 2026-07-22): identical `-v /sys/fs/cgroup` mount showed no `kubepods` without `--cgroupns=host` and the full `kubepods/burstable/pod<UID>/` tree with it — the failure mode is silent |
+| Parse per-pod stats from inside a default (non-privileged) pod | Read via a host-namespace-capable agent (privileged/hostPID DaemonSet) or consume kubelet metrics endpoints instead: [infrastructure-observability-missing-container-metrics] | The default pod cgroupns cannot see sibling pods at all |
+
+## Sources
+
+- https://man7.org/linux/man-pages/man7/cgroup_namespaces.7.html — cgroup namespaces virtualize the cgroup view; cgroupfs contents depend on the namespace of the mount's creator; nsdelegate write restrictions
+- https://docs.docker.com/reference/cli/docker/container/run/ — `--cgroupns=host` runs the container in the host's cgroup namespace; private is the cgroup v2 default
+- https://man7.org/linux/man-pages/man1/nsenter.1.html — `-C/--cgroup` enters the target process's cgroup namespace
+- https://github.com/kubernetes/kubernetes/issues/103363 — no first-class host cgroupns for pods; nsenter-based workaround pattern (as used by Cilium)
+- Field context: 2026-07-22 OrbStack VM measurement — `docker run -v /sys/fs/cgroup:/hostcg alpine`: `kubepods` absent; adding `--cgroupns=host`: full `kubepods/burstable/pod<UID>/` hierarchy visible
diff --git a/wiki/infrastructure/containers/resource-limits-and-probes.md b/wiki/infrastructure/containers/resource-limits-and-probes.md
index d76c93b..1e11b65 100644
--- a/wiki/infrastructure/containers/resource-limits-and-probes.md
+++ b/wiki/infrastructure/containers/resource-limits-and-probes.md
@@ -9,7 +9,7 @@ sources:
   - https://kubernetes.io/docs/concepts/configuration/manage-resources-containers/
   - https://kubernetes.io/docs/concepts/workloads/pods/pod-qos/
 last_verified: 2026-07-10
-related: [infrastructure-deploy-rollout-and-rollback, backend-java-runtime-threads-and-memory]
+related: [infrastructure-deploy-rollout-and-rollback, backend-java-runtime-threads-and-memory, infrastructure-containers-host-cgroup-visibility, infrastructure-observability-missing-container-metrics]
 ---
 
 # Resource Limits and Health Probes in Deployment Manifests
diff --git a/wiki/infrastructure/index.md b/wiki/infrastructure/index.md
index 4940602..2bff59c 100644
--- a/wiki/infrastructure/index.md
+++ b/wiki/infrastructure/index.md
@@ -25,6 +25,7 @@ Match your situation to a "load when" line; load only matching pages.
 
 | Page | Load when |
 |------|-----------|
+| [host-cgroup-visibility](containers/host-cgroup-visibility.md) | A container must read the host's full cgroup v2 hierarchy (other pods' CPU/memory stats) via a hostPath/`-v` mount of `/sys/fs/cgroup`; the mounted directory is missing the `kubepods` subtree with no error |
 | [image-builds](containers/image-builds.md) | Writing or reviewing a Dockerfile; images rebuild everything on small changes, build slowly, or are too large; choosing an image tagging scheme |
 | [resource-limits-and-probes](containers/resource-limits-and-probes.md) | Writing or reviewing Kubernetes-style deployment manifests; pods OOMKilled, evicted, or CPU-throttled; a dependency outage triggered a restart storm; traffic hitting pods that are not ready |
 
@@ -45,4 +46,5 @@ Match your situation to a "load when" line; load only matching pages.
 | Page | Load when |
 |------|-----------|
 | [logs-metrics-signals](observability/logs-metrics-signals.md) | Instrumenting a new or existing service (logs, metrics, correlation ids); an incident revealed you couldn't see what happened; choosing between a log line and a metric; a metric label would carry unbounded values (user ids/UUIDs) |
+| [missing-container-metrics](observability/missing-container-metrics.md) | Prometheus `container_*` CPU/memory series are empty or pod dashboards blank while kubelet scrape targets all report up (common on embedded/VM Kubernetes like OrbStack); deciding between cAdvisor and kubelet `/metrics/resource` scraping |
 | [alerting](observability/alerting.md) | Creating or reviewing alerts; the team ignores a noisy pager; deciding whether a condition pages, tickets, or stays on a dashboard |
diff --git a/wiki/infrastructure/observability/logs-metrics-signals.md b/wiki/infrastructure/observability/logs-metrics-signals.md
index 87e9108..2d11dbb 100644
--- a/wiki/infrastructure/observability/logs-metrics-signals.md
+++ b/wiki/infrastructure/observability/logs-metrics-signals.md
@@ -9,7 +9,7 @@ sources:
   - https://prometheus.io/docs/practices/naming/
   - https://cheatsheetseries.owasp.org/cheatsheets/Logging_Cheat_Sheet.html
 last_verified: 2026-07-10
-related: [infrastructure-observability-alerting]
+related: [infrastructure-observability-alerting, infrastructure-observability-missing-container-metrics]
 ---
 
 # Instrumenting a Service with Logs and Metrics
diff --git a/wiki/infrastructure/observability/missing-container-metrics.md b/wiki/infrastructure/observability/missing-container-metrics.md
new file mode 100644
index 0000000..5127022
--- /dev/null
+++ b/wiki/infrastructure/observability/missing-container-metrics.md
@@ -0,0 +1,80 @@
+---
+id: infrastructure-observability-missing-container-metrics
+domain: infrastructure
+category: observability
+applies_to: [kubernetes, kube-prometheus-stack]
+confidence: verified
+sources:
+  - https://kubernetes.io/docs/tasks/debug/debug-cluster/resource-metrics-pipeline/
+  - https://github.com/prometheus-community/helm-charts/blob/main/charts/kube-prometheus-stack/values.yaml
+  - https://raw.githubusercontent.com/kubernetes-monitoring/kubernetes-mixin/master/dashboards/resources/queries/pod.libsonnet
+  - https://github.com/orbstack/orbstack/issues/2217
+last_verified: 2026-07-28
+related: [infrastructure-observability-logs-metrics-signals, infrastructure-containers-resource-limits-and-probes]
+---
+
+# Container Metrics Empty While Kubelet Scrape Targets Report Up
+
+## When this applies
+
+Monitoring pod CPU/memory with Prometheus (e.g. kube-prometheus-stack) and
+`container_*` series are empty or dashboards show nothing, while every kubelet
+scrape target is healthy ("up"). Common on embedded/VM Kubernetes distributions
+(observed on OrbStack) whose kubelet cAdvisor endpoint emits only `machine_*`
+series.
+
+## Do this
+
+1. Diagnose by series presence, not target health — a successful scrape of an
+   endpoint that emits no `container_*` series still shows "up":
+
+```
+kubectl get --raw /api/v1/nodes/<node>/proxy/metrics/cadvisor | grep -c '^container_'
+kubectl get --raw /api/v1/nodes/<node>/proxy/metrics/resource | grep -c '^container_'
+```
+
+2. When cAdvisor emits no `container_*` series but `/metrics/resource` does,
+   scrape the kubelet resource endpoint instead. It serves
+   `container_cpu_usage_seconds_total` and `container_memory_working_set_bytes`
+   (defined in kubelet's resource-metrics collector). In kube-prometheus-stack
+   values:
+
+```yaml
+kubelet:
+  serviceMonitor:
+    resource: true
+    resourcePath: "/metrics/resource"
+```
+
+   Both lines are required: the chart's default `resourcePath` is still
+   `/metrics/resource/v1alpha1` (renamed in Kubernetes 1.18), which 404s on
+   modern kubelets — enabling `resource: true` alone produces a down target.
+
+3. Expect the bundled "Compute Resources" dashboards to stay empty even after
+   the data arrives: kubernetes-mixin dashboard queries and recording rules
+   filter with `image!=""`, and `/metrics/resource` series carry no `image`
+   label (only container/pod/namespace). Build a custom dashboard or recording
+   rules against the resource-endpoint series directly.
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| You need per-container filesystem, network, or throttling metrics | `/metrics/resource` carries only CPU and memory; those richer series exist only in cAdvisor — fix or replace the runtime's cAdvisor support instead |
+| Dashboards empty but PromQL on the raw series returns data | You are hitting the `image!=""` filter, not a collection gap — adjust the queries, not the scrape |
+| Verifying the fix | Query `container_memory_working_set_bytes{container!=""}` directly in Prometheus and compare one pod against `kubectl top pod` |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Conclude "metrics are being collected" from all-green scrape targets | Count the specific series you need at the source endpoint | Scrape health checks the HTTP exchange, not whether the series you consume exist |
+| Enable `kubelet.serviceMonitor.resource: true` and stop | Also set `resourcePath: "/metrics/resource"` | The chart default still points at the pre-1.18 `v1alpha1` path, which 404s |
+
+## Sources
+
+- https://kubernetes.io/docs/tasks/debug/debug-cluster/resource-metrics-pipeline/ — kubelet `/metrics/resource` endpoint in the resource metrics pipeline
+- https://raw.githubusercontent.com/kubernetes/kubernetes/master/pkg/kubelet/metrics/collectors/resource_metrics.go — the endpoint's exact series: `container_cpu_usage_seconds_total`, `container_memory_working_set_bytes`
+- https://github.com/prometheus-community/helm-charts/blob/main/charts/kube-prometheus-stack/values.yaml — `kubelet.serviceMonitor.resource` (default false) and `resourcePath` (default `/metrics/resource/v1alpha1`, comment: renamed in k8s 1.18)
+- https://raw.githubusercontent.com/kubernetes-monitoring/kubernetes-mixin/master/dashboards/resources/queries/pod.libsonnet — `container_memory_working_set_bytes{..., image!=""}` filter; same pattern in `rules/apps.libsonnet` recording rules
+- https://github.com/orbstack/orbstack/issues/2217 — corroborates node-level metrics working while container-level silently missing on OrbStack embedded k8s (the machine_*-only cAdvisor observation itself is field-tested, 2026-07-15)
diff --git a/wiki/qa/process/release-gates.md b/wiki/qa/process/release-gates.md
index 332660e..a93fd1d 100644
--- a/wiki/qa/process/release-gates.md
+++ b/wiki/qa/process/release-gates.md
@@ -7,7 +7,7 @@ confidence: field-tested
 sources:
   - https://sre.google/sre-book/release-engineering/
 last_verified: 2026-07-10
-related: [qa-process-regression-scope, qa-exploratory-exploratory-sessions]
+related: [qa-process-regression-scope, qa-exploratory-exploratory-sessions, backend-common-llm-gateway-model-alias-defaults]
 ---
 
 # Deciding Whether a Build Is Ready to Ship

From 73d499e10b25d42ff96bcf934680b14f7e034756 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=EC=B5=9C=EC=98=81=EA=B8=B0?= <dch0202@rsquare.co.kr>
Date: Tue, 28 Jul 2026 15:11:38 +0900
Subject: [PATCH 2/3] knowledge: add cross-check marker to ingest report

---
 .dev-loop/INGEST_REPORT.md | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/.dev-loop/INGEST_REPORT.md b/.dev-loop/INGEST_REPORT.md
index 288c223..8c776d5 100644
--- a/.dev-loop/INGEST_REPORT.md
+++ b/.dev-loop/INGEST_REPORT.md
@@ -1,5 +1,7 @@
 # Knowledge flush — 4 insight(s)
 
+Cross-Check: 인사이트 4건 각각을 독립 general-purpose 서브에이전트 4개가 1차 소스(공식 문서·소스코드·이슈트래커) 대조로 교차검증 — 2건의 실질 교정 반영(cgroupns는 mount-time 시맨틱으로 메커니즘 수정, kps resourcePath v1alpha1은 현재도 기본값), 캐비앗 2건 추가(vLLM 필드 개명, DeepSeek 별도 CoT 예산).
+
 ## Verified best-practice
 
 **1. Container metrics empty while kubelet targets report up (OrbStack embedded k8s)**

From 719e54413f54a6a779cc297d800777ae0715ea4c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=EC=B5=9C=EC=98=81=EA=B8=B0?= <dch0202@rsquare.co.kr>
Date: Tue, 28 Jul 2026 15:13:09 +0900
Subject: [PATCH 3/3] knowledge: add decision log to ingest report

---
 .dev-loop/INGEST_REPORT.md | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/.dev-loop/INGEST_REPORT.md b/.dev-loop/INGEST_REPORT.md
index 8c776d5..3cc9e3d 100644
--- a/.dev-loop/INGEST_REPORT.md
+++ b/.dev-loop/INGEST_REPORT.md
@@ -2,6 +2,13 @@
 
 Cross-Check: 인사이트 4건 각각을 독립 general-purpose 서브에이전트 4개가 1차 소스(공식 문서·소스코드·이슈트래커) 대조로 교차검증 — 2건의 실질 교정 반영(cgroupns는 mount-time 시맨틱으로 메커니즘 수정, kps resourcePath v1alpha1은 현재도 기본값), 캐비앗 2건 추가(vLLM 필드 개명, DeepSeek 별도 CoT 예산).
 
+## Decision Log
+
+- **의도**: `~/.dev-loop/queue`에 쌓인 ★ Insight 후보 4건을 knowledge-flush 파이프라인(리서치 검증 → 기존 레이어 중복확인 → 라우팅 → wiki-ingest)으로 정제해 위키에 반영. 사용자가 명시 지시한 단일 배치 PR.
+- **배제한 대안**: ① 기존 페이지에 병합 — 4건 모두 기존 페이지의 트리거와 불일치(아래 Existing-layer check)라 신규 페이지로. ② insight 2·3을 `reliability` 카테고리에 편입 — reliability는 전송 계층 실패(타임아웃/재시도) 소유라 시맨틱 응답 검증과 안 맞아 신규 카테고리 `backend/common/llm`으로(라우팅 근거는 Routing decision). ③ 수확된 원문 그대로 반영 — 리서치에서 메커니즘 교정 2건이 나와 교정본으로 반영.
+- **리뷰어가 볼 곳**: 신규 4페이지의 `confidence`/`sources`가 실제 근거와 맞는지, `backend/common/llm` 신규 카테고리 판단, 그리고 기존 5개 페이지의 frontmatter `related:` 한 줄 수정이 과하지 않은지.
+- [추정] 이 레포의 PR 리뷰는 dev-loop 규약상 레포 오너(choiyounggi) 단독 리뷰로 충분하다고 판단 — 팀 표준 리뷰어 3인은 개인 레포 collaborator가 아닐 수 있어 gh가 거부하면 미지정으로 진행.
+
 ## Verified best-practice
 
 **1. Container metrics empty while kubelet targets report up (OrbStack embedded k8s)**