From 1ff1aaf51484f9a61fbc0611eb725176ebc6dd2a 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: Thu, 30 Jul 2026 17:46:14 +0900
Subject: [PATCH 1/2] knowledge: ingest 4 verified insight(s)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- platforms/processes/non-interactive-cli-invocation (new): a non-interactive
  flag is not a closed stdin — pair it with </dev/null, a fail-fast switch and a
  timeout; split a zero-output hang by whether the far side logged the request.
- platforms/shells/command-text-inspected-before-execution (new): a gate reads
  the command as pre-expansion text, so a quoted argument defeats its extractor
  while an unexpanded $VAR resolves to a nonexistent path — two distinct
  refusals; write literal paths and create gate-read files in a prior command.
- qa/document-verification/editing-a-gated-document (new category page): the
  author-side counterpart to spec-document gates — inventory anchors before
  editing, describe upstream as observed shape, scope a check outside the region
  quoting it, record scoped conditions instead of global counts.

Sources live-verified (nohup, ssh(1)/ssh_config(5), timeout(1), pgrep(1),
Claude Code hooks, POSIX 2.6, Vale scopes/existence, markdownlint MD013);
regex failure modes reproduced against hooks/pre-flush-pr-gate.sh.
---
 .dev-loop/INGEST_REPORT.md                    | 172 +++++++++++++++---
 log.md                                        |   1 +
 wiki/platforms/index.md                       |   2 +
 .../processes/background-services.md          |   2 +-
 .../non-interactive-cli-invocation.md         |  97 ++++++++++
 ...command-text-inspected-before-execution.md |  94 ++++++++++
 .../shells/portable-shell-scripts.md          |   2 +-
 .../editing-a-gated-document.md               | 107 +++++++++++
 wiki/qa/index.md                              |   6 +
 9 files changed, 459 insertions(+), 24 deletions(-)
 create mode 100644 wiki/platforms/processes/non-interactive-cli-invocation.md
 create mode 100644 wiki/platforms/shells/command-text-inspected-before-execution.md
 create mode 100644 wiki/qa/document-verification/editing-a-gated-document.md

diff --git a/.dev-loop/INGEST_REPORT.md b/.dev-loop/INGEST_REPORT.md
index 2ff3871..fd8052a 100644
--- a/.dev-loop/INGEST_REPORT.md
+++ b/.dev-loop/INGEST_REPORT.md
@@ -1,35 +1,163 @@
-# Knowledge flush — 1 insight
+# Knowledge flush — 4 insight(s)
 
-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`."
+4 queued candidates → **3 pages** (2 new platforms pages, 1 new qa page in a new
+category). Two qa candidates were merged into a single page because they are the same
+case seen twice; nothing was dropped.
+
+Cross-Check: primary-source verification of every directive (man pages, official tool
+docs) plus local reproduction of the two mechanism claims against this repo's own hook
+source. No independent adversarial agent review was run for this flush — the
+verification is documentary and reproducible, not a second opinion.
 
 ## 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. A non-interactive flag is not a closed stdin → `confidence: verified`
+**Claim.** When invoking a prompt-capable CLI unattended, pass the tool's
+non-interactive switch *and* redirect stdin from `/dev/null`; a hang with zero output
+is evidence about the client until the far side's log shows the request arrived.
+
+| Source | What it establishes |
+|---|---|
+| [nohup(1)](https://man7.org/linux/man-pages/man1/nohup.1.html) | "If standard input is a terminal, redirect it from an unreadable file" — detaching includes taking terminal stdin away |
+| [ssh(1)](https://man.openbsd.org/ssh.1) | `-n` "Redirects stdin from /dev/null (actually, prevents reading from stdin). This must be used when ssh is run in the background" |
+| [ssh_config(5)](https://man.openbsd.org/ssh_config.5) | `BatchMode=yes` disables "password prompts and host key confirmation requests", "useful in scripts and other batch jobs where no user is present" |
+| [timeout(1)](https://man7.org/linux/man-pages/man1/timeout.1.html) | "Start COMMAND, and kill it if still running after DURATION"; exit 124 on timeout — a distinguishable "blocked" signal |
+
+Two established tools implement *both* halves (close stdin **and** a fail-fast
+interaction switch) as separate mechanisms, which is exactly the candidate's directive.
+The client/server split step is field-derived (a gateway access log showed zero
+requests from the host during two hangs; `</dev/null` fixed it immediately) and is
+labelled as such in the page's Field context rather than presented as documented.
+
+### 2. Command-inspecting gates read pre-expansion text → `confidence: verified`
+**Claim.** Write values literally in an argument a gate inspects; create files a gate
+reads in a *prior* command.
+
+- [Claude Code hooks docs](https://code.claude.com/docs/en/hooks) — `PreToolUse` runs
+  "Before a tool call executes. Can block it"; the hook's stdin JSON carries
+  `tool_input.command`, the unexecuted command string. Exit 2 blocks and "stderr text
+  is fed back to Claude as an error message."
+- [POSIX shell 2.6](https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html)
+  — word expansion is performed by the shell as it processes the line, so an external
+  reader of the command text sees none of it applied.
+
+**Reproduced locally** against this repo's own extraction pattern
+(``--body-file[= ]+[^ '"`]+``, `hooks/pre-flush-pr-gate.sh:56`), 2026-07-30:
+
+| Command text | Extracted | Refusal |
+|---|---|---|
+| `--body-file "$REPO/…"` | *(empty)* | "no `--body-file` found" |
+| `--body-file $REPO/…` | literal `$REPO/…` | "body file does not exist yet" |
+| `--body-file "/abs/…"` | *(empty)* | "no `--body-file` found" |
+| `--body-file /abs/…` | `/abs/…` | passes |
+| `--body-file=/abs/…` | `/abs/…` | passes |
+
+This **corrects the candidate**, which attributed the failure to quoting alone. The run
+shows two distinct failure modes with two different error messages — quotes defeat
+*extraction*, an unexpanded variable defeats *resolution* — and that a quoted
+**literal** path fails too. The page states both; the sharper form is what makes the
+error message diagnostic.
+
+### 3+4. Editing a document that text gates check → `confidence: field-tested`
+**Claim.** Inventory a file's gate anchors before editing; describe an upstream
+contract as an observed shape rather than with definition verbs; scope a check outside
+the region that quotes it; record scoped conditions, not global counts.
 
-**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**
+The *hazards* are documented; the *directives* are field-derived, hence `field-tested`
+rather than `verified`:
 
-**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**
+| Source | What it establishes |
+|---|---|
+| [Vale `existence`](https://docs.vale.sh/checks/existence) | The check "looks for the 'existence' of particular tokens" as a word-bounded non-capturing group — a lexical gate matches patterns, not intent. This is why a purely descriptive sentence trips a "do not redefine" gate |
+| [Vale scopes](https://docs.vale.sh/topics/scopes.md) | Scopes restrict where a rule applies; "Any scope prefaced with `~` is negated" and scopes chain — keeping checks off regions such as code examples is first-class |
+| [markdownlint MD013](https://github.com/DavidAnson/markdownlint/blob/main/doc/md013.md) | Rules expose `code_blocks`/`tables`/`headings` booleans (default `true`) so quoted code can be excluded from a prose rule |
+| [pgrep(1)](https://man7.org/linux/man-pages/man1/pgrep.1.html) | "The running pgrep, pkill, or pidwait process will never report itself as a match" — self-exclusion is designed in because self-matching is the expected failure |
 
-**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).
+Field evidence retained in the page: a vague-word audit that matched its own quoted
+pattern (1 global hit → 3 after quoting the fix, 0 under an `awk`-scoped run; the same
+self-reference observed 3× across two documents), and a vocabulary gate
+(`(arena|pool)…(재정의|정의한다|규정한다)`) that failed a true descriptive sentence
+until it was rewritten as an observation, restoring the suite 60 → 61.
 
 ## 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 in full:** `INDEX.md`, `AGENTS.md`, `templates/page.md`,
+`wiki/platforms/index.md`, `wiki/qa/index.md`,
+`wiki/platforms/processes/background-services.md`,
+`wiki/platforms/shells/portable-shell-scripts.md`, `log.md`, plus the two in-flight
+pages `wiki/qa/document-verification/spec-document-gates.md` (PR #10) and
+`wiki/testing/docs-as-spec/document-conformance-checks.md` (PR #9), fetched from the
+fork. **Repo-wide greps** for `grep|self-referen|lexical|doc-as-spec`,
+`stdin|/dev/null|non-interactive|tty`, and `PreToolUse|pre-commit|hook`.
+**Queue dedup:** all 4 candidate hashes absent from `.processed.jsonl`.
+
+**Open-PR dedup was decisive here.** PRs #6–#10 are open and unmerged, and
+#7/#8/#9/#10 all sit in the doc-gate theme, so `main` alone understates coverage.
+Overlap verdicts:
+
+| Overlap candidate | Verdict |
+|---|---|
+| `qa-document-verification-spec-document-gates` (**PR #10, open**) | **Complement, not duplicate.** It owns the *gate author's* side (four axes, controls, fail-closed anchors). The new page owns the *document author's* side: what to do when your prose must survive gates that already exist, including gates whose pattern your text quotes. Neither self-reference scoping nor anchor inventory appears in #10; its nearest row ("Examples section satisfies the check") is a different cause |
+| `testing-docs-as-spec-document-conformance-checks` (**PR #9, open**) | No overlap. Positive/negative controls and GFM pipe parsing for checks under construction; says nothing about editing an already-gated document |
+| `testing-quality-tests-that-cannot-fail` | Adjacent, linked. Owns proving a *test* can fail. The new qa page routes gate-construction questions to it rather than restating them |
+| `qa-process-regression-scope` | Adjacent, linked. Supplies the "re-run the full set, compare the baseline" principle the new page applies to gate suites |
+| `platforms-processes-background-services` | Closest neighbour to insight 1 and **already covers `nohup … & disown`** for *lifetime*. It does not cover stdin as a blocking input or the client/server split. Kept separate (its "load when" is persistence), linked **both ways** |
+| `platforms-shells-portable-shell-scripts` | Closest neighbour to insight 2, owns quoting *for the shell* — and its rule is "quote every expansion". The new page narrows one argument read by an external gate, so an unqualified reader could see a contradiction; the new page's edge-case table states explicitly that it "narrows one argument, it does not license unquoted expansions elsewhere". Linked both ways |
+| `platforms-environment-path-resolution` | Linked only (literal-vs-resolved paths in non-interactive contexts) |
+
+**Conflicts flagged:** one *soft* directive tension (quote-everything vs. write-this-one-
+argument-literally), resolved inside the new page rather than by editing the old one. No
+factual contradiction found.
+
+**Reciprocal `related:` links added** to `background-services.md` and
+`portable-shell-scripts.md` (frontmatter only; `last_verified` deliberately not bumped,
+since nothing on those pages was re-verified).
+
+### Merge conflict to expect (please read before merging)
+This branch and **PR #10** both introduce the `## document-verification` section in
+`wiki/qa/index.md` at the same insertion point, each with its own page row. Whichever
+merges second will conflict there. **Resolution: keep one heading and both rows.** To
+keep the conflict to that single hunk, this branch deliberately does **not** touch
+`INDEX.md` — PR #10's qa route line ("automated verification of document deliverables
+(spec/RFC gates)") already covers this new page. If #10 is rejected instead, the qa
+route line in `wiki/qa/index.md` should gain a document-verification clause in a
+follow-up. No `related:` id in this branch points at #10's page, so nothing here breaks
+under either outcome; once both are merged the two pages are worth cross-linking.
 
 ## 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 | New category? |
+|---|---|---|
+| 1 — non-interactive CLI hang | `platforms/processes/non-interactive-cli-invocation.md` (new page) | No — `processes` already owns process/session lifetime |
+| 2 — gate reads command text | `platforms/shells/command-text-inspected-before-execution.md` (new page) | No — `shells`, justified below |
+| 3 + 4 — self-reference & lexical gates | `qa/document-verification/editing-a-gated-document.md` (new page) | **Yes** — `document-verification`, the same category PR #10 introduces |
+
+**Insight 1 → platforms, not debugging.** Half the insight is diagnostic, but the
+artifact the reader changes is the invocation command, and `AGENTS.md` routes by owned
+artifact. `debugging-methodology-reproduce-first` is linked for the isolation half.
+Merging into `background-services` was rejected: that page's case is *persistence*, and
+adding a blocking-stdin case would drift its "load when".
+
+**Insight 2 → `shells`, not a new category and not `tools`.** The mechanism is expansion
+timing — *when* the shell rewrites the line relative to other readers of it — which is
+shell semantics, so `shells` holds two coherent pages (portability; expansion timing vs.
+external inspectors). `tools` is BSD-vs-GNU userland differences, which this is not. A
+dedicated category (e.g. `policy-gates`) was considered and rejected as a one-page
+category with no second member in sight.
+
+**Insights 3+4 → one page, not two.** Both are the document author's side of the
+doc-gate loop: 3 is "the gate matched the pattern I quoted", 4 is "the gate matched the
+verb I used / the anchor I moved". `AGENTS.md` requires one case per page, and the shared
+case is "writing prose inside a document that lexical gates run over" — the directives
+interleave (scope the check, phrase as observation, record a scoped condition, re-run the
+suite), so splitting would have produced two pages that each need the other. **New
+category justified:** the existing qa categories are `process` (human release process),
+`environments`, `bug-reports`, `exploratory` — none covers automated checks over a
+written deliverable. PR #10 reached the same conclusion independently, which is
+corroboration rather than duplication.
+
+### Invariants checked on this branch
+Body lines 83 / 82 / 93 (limit 120) · all four required sections present on each page ·
+every `related:` id and inline `[page-id]` resolves (13 checks, 0 misses) · each new page
+listed in its domain index with a multi-use-case "load when" · every index relative link
+resolves · no banned vague qualifiers · `log.md` entry appended.
diff --git a/log.md b/log.md
index 587a5dc..a674085 100644
--- a/log.md
+++ b/log.md
@@ -34,3 +34,4 @@ 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).
+## [2026-07-30] ingest | 4 queued insights → 3 pages. platforms/processes/non-interactive-cli-invocation (a non-interactive flag is not a closed stdin: `</dev/null` + tool fail-fast switch + timeout; split a zero-output hang by whether the far side logged the request — nohup/ssh -n/BatchMode/timeout docs). platforms/shells/command-text-inspected-before-execution (NEW page in shells: a gate reads `tool_input.command` pre-expansion, so quoted args defeat its extractor while unexpanded `$VAR` resolves to a nonexistent path — two distinct refusals; write literal paths, create gate-read files in a prior command, pass dangerous-looking prose by file). qa/document-verification/editing-a-gated-document (NEW category page, author-side counterpart to spec-document gates: anchor inventory before editing, describe upstream as observed shape not with definition verbs, scope a check outside the region quoting it, record scoped conditions instead of global counts, re-run the full gate set against a baseline). Regex failure modes reproduced against hooks/pre-flush-pr-gate.sh; sources live-verified (nohup, ssh(1)/ssh_config(5), timeout(1), pgrep(1) self-exclusion, Claude Code hooks, Vale scopes/existence, markdownlint MD013).
diff --git a/wiki/platforms/index.md b/wiki/platforms/index.md
index dccb400..47065f5 100644
--- a/wiki/platforms/index.md
+++ b/wiki/platforms/index.md
@@ -15,6 +15,7 @@ Match your situation to a "load when" line; load only matching pages.
 | Page | Load when |
 |------|-----------|
 | [portable-shell-scripts](shells/portable-shell-scripts.md) | Writing a shell script that must run on more than one machine/OS/shell or in CI; a script that works locally fails elsewhere; choosing a shebang (bash vs sh); a bash script misbehaves in zsh or vice versa (unquoted vars, `=word`, array indexing); deciding how `set -euo pipefail` protects (and doesn't); building argument lists safely |
+| [command-text-inspected-before-execution](shells/command-text-inspected-before-execution.md) | A hook, policy gate, allow-list, or audit rule blocked a command that is correct as written; composing a command that must satisfy such a gate first try; deciding whether to write a path literally or as `"$VAR"` in an inspected argument; a gate reports an argument missing or a file nonexistent though both are right; a gate must read a file your command creates; prose containing a dangerous-looking command (release notes, docs, fixtures) trips a text scanner |
 
 ## tools
 
@@ -41,6 +42,7 @@ Match your situation to a "load when" line; load only matching pages.
 | Page | Load when |
 |------|-----------|
 | [background-services](processes/background-services.md) | Something must run persistently or on a schedule on a dev machine or server (daemon, watcher, cron-style job); a "started" process dies when the terminal/SSH/agent session ends; choosing nohup vs LaunchAgent vs systemd unit vs cron/timer; a job works in the terminal but fails under cron/launchd (minimal environment); wiring service logs and restart policy |
+| [non-interactive-cli-invocation](processes/non-interactive-cli-invocation.md) | Calling a tool that can prompt (agent CLI, ssh, git, package manager) from a script, CI step, hook, or agent session; such a call produced no output and never returned; deciding whether a hang belongs to the client, the network, or the far-side service; choosing the stdin/timeout/fail-fast switches for an unattended call; a TTY-detecting tool changes its output format under automation |
 
 ## toolchains
 
diff --git a/wiki/platforms/processes/background-services.md b/wiki/platforms/processes/background-services.md
index f6e9503..22b1e9a 100644
--- a/wiki/platforms/processes/background-services.md
+++ b/wiki/platforms/processes/background-services.md
@@ -11,7 +11,7 @@ sources:
   - https://man7.org/linux/man-pages/man1/nohup.1.html
   - https://man7.org/linux/man-pages/man1/loginctl.1.html
 last_verified: 2026-07-10
-related: [platforms-toolchains-version-management, platforms-shells-portable-shell-scripts]
+related: [platforms-toolchains-version-management, platforms-shells-portable-shell-scripts, platforms-processes-non-interactive-cli-invocation]
 ---
 
 # Keeping a Process Running Beyond the Terminal Session
diff --git a/wiki/platforms/processes/non-interactive-cli-invocation.md b/wiki/platforms/processes/non-interactive-cli-invocation.md
new file mode 100644
index 0000000..6fbfdc1
--- /dev/null
+++ b/wiki/platforms/processes/non-interactive-cli-invocation.md
@@ -0,0 +1,97 @@
+---
+id: platforms-processes-non-interactive-cli-invocation
+domain: platforms
+category: processes
+applies_to: [macos, linux]
+confidence: verified
+sources:
+  - https://man7.org/linux/man-pages/man1/nohup.1.html
+  - https://man.openbsd.org/ssh.1
+  - https://man.openbsd.org/ssh_config.5
+  - https://man7.org/linux/man-pages/man1/timeout.1.html
+last_verified: 2026-07-30
+related: [platforms-processes-background-services, platforms-shells-portable-shell-scripts, debugging-methodology-reproduce-first]
+---
+
+# Invoking an Interactive-Capable CLI From a Script, Hook, or Agent
+
+## When this applies
+
+You are calling a tool that can prompt (a coding-agent CLI, `ssh`, `git`, a package
+manager, an installer) from a script, CI step, hook, or agent session; or such a
+call produced no output and never returned, and you are deciding whether the tool,
+the network, or the far-side service is at fault.
+
+## Do this
+
+1. **Close stdin and pass the tool's own non-interactive switch — both, not either.**
+   A `-p`/`--print`/`--quiet` flag suppresses *rendering*; it does not promise the
+   tool never reads stdin. When stdin is still a terminal, any prompt the flag
+   didn't cover (a permission confirmation, a passphrase, a host-key question)
+   blocks forever with zero bytes written. This is why `ssh -n` exists — it
+   "Redirects stdin from /dev/null (actually, prevents reading from stdin)" and
+   "must be used when ssh is run in the background" — and why `nohup` does it for
+   you: "If standard input is a terminal, redirect it from an unreadable file."
+
+| Case | Do |
+|------|----|
+| One-shot call whose output you need in this step | `cmd … </dev/null` plus the tool's non-interactive flag |
+| Call that must outlive the session | `nohup cmd … </dev/null >"$LOG" 2>&1 & disown` ([platforms-processes-background-services]) |
+| Tool authenticates over the network | Add the tool's fail-fast switch so a missing credential errors instead of prompting: `ssh -o BatchMode=yes`, `GIT_TERMINAL_PROMPT=0`, `DEBIAN_FRONTEND=noninteractive` |
+| Any unattended call | Wrap in a timeout so a block fails loudly: `timeout <secs> cmd …` (on macOS, `gtimeout` — see [platforms-tools-bsd-vs-gnu-cli]) |
+
+2. **Give the call a deadline rather than watching it.** `timeout` will "Start
+   COMMAND, and kill it if still running after DURATION", exiting 124 when that
+   fires — a distinguishable signal that the call blocked, instead of an agent turn
+   that hangs until something else kills it.
+
+3. **When it still hangs, locate the boundary before blaming either side.** Zero
+   output is compatible with two very different faults: the request never left the
+   client, or the far side never answered. Split them with evidence from the far
+   side — the server/gateway access log for that source IP and time window, or a
+   packet/connection count:
+
+| Observation at the far side | Conclusion | Next action |
+|-----------------------------|------------|-------------|
+| No request recorded in the window | The call never got that far — it is blocked locally (stdin, auth prompt, DNS, proxy) | Re-run with `</dev/null` and a timeout; inspect the client's own debug/verbose log |
+| Request recorded, no response or a slow one | The far side owns it | Take the latency/error question to the service |
+
+4. **Re-run the identical command with stdin closed as the first diagnostic step.**
+   If a `</dev/null` run returns immediately, the hang was a prompt, and no
+   investigation of the far side is warranted.
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| Tool needs input you intend to supply | Feed it from a file or heredoc (`cmd <input.txt`), never leave it on the inherited terminal — an unfed prompt and a fed one are indistinguishable from the caller |
+| Tool detects a TTY and changes output format (color codes, progress bars, pagers) | Closing stdin is not enough; also disable the pager/color explicitly (`GIT_PAGER=cat`, `--no-color`) so parsers see stable text |
+| The call must run under a supervisor or scheduler | Services get no terminal at all, so the prompt fails differently there than in your shell — verify by log, not by launch exit code ([platforms-processes-background-services]) |
+| Wrapper CLI shells out to a second binary that prompts | Redirecting the wrapper's stdin covers the child only if the child inherits it; confirm with a timeout run before trusting the wrapper's own flags |
+| Hang reproduces even with stdin closed and a timeout | Treat it as a real far-side or network fault and continue from the boundary evidence in step 3 |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Trust a `--print`/`-p` flag to make a call safe for unattended use | Add `</dev/null` and a timeout alongside the flag | The flag governs output; a prompt the flag does not cover still reads the inherited terminal and blocks with no output |
+| Conclude "the model/gateway/service is slow or broken" from a call that produced nothing | Check the far side's log for the request first | A locally blocked call leaves no trace there, so a hang with no logged request is evidence about the client, not the service |
+| Watch a hung foreground call and cancel it by hand | Bound it with `timeout` and read exit 124 as "blocked" | Manual cancellation loses the distinction between slow and stuck, and burns the whole wait |
+| Debug the hang by re-running it the same way | Re-run it with stdin closed as the first variation | Closing stdin is a one-token change that either fixes it or eliminates the largest class of cause |
+
+## Sources
+
+- https://man7.org/linux/man-pages/man1/nohup.1.html — "If standard input is a terminal, redirect it from an unreadable file" — detaching a command includes taking its terminal stdin away
+- https://man.openbsd.org/ssh.1 — `-n` "Redirects stdin from /dev/null (actually, prevents reading from stdin). This must be used when ssh is run in the background"
+- https://man.openbsd.org/ssh_config.5 — `BatchMode=yes` disables "user interaction such as password prompts and host key confirmation requests", "useful in scripts and other batch jobs where no user is present"
+- https://man7.org/linux/man-pages/man1/timeout.1.html — "Start COMMAND, and kill it if still running after DURATION"; exit status 124 "if COMMAND times out, and --preserve-status is not specified"
+
+## Field context
+
+Distilled from a 2026-07 session that ran an agent CLI with its non-interactive
+`--tools` flag in the foreground: two runs hung (300 s and 150 s) with zero bytes
+of output. The gateway access log showed **no request from that host** in either
+window, which ruled out the model and the gateway; the same command launched with
+stdin taken from `/dev/null` completed its tool call immediately. The far-side-log
+check in step 3 is the field-derived part of this page; the stdin rules in step 1
+are the documented behavior the incident rediscovered.
diff --git a/wiki/platforms/shells/command-text-inspected-before-execution.md b/wiki/platforms/shells/command-text-inspected-before-execution.md
new file mode 100644
index 0000000..e69daff
--- /dev/null
+++ b/wiki/platforms/shells/command-text-inspected-before-execution.md
@@ -0,0 +1,94 @@
+---
+id: platforms-shells-command-text-inspected-before-execution
+domain: platforms
+category: shells
+applies_to: [bash, zsh, posix-sh]
+confidence: verified
+sources:
+  - https://code.claude.com/docs/en/hooks
+  - https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html
+last_verified: 2026-07-30
+related: [platforms-shells-portable-shell-scripts, platforms-environment-path-resolution]
+---
+
+# Commands Read as Text by a Gate Before the Shell Runs Them
+
+## When this applies
+
+A hook, policy gate, allow-list, or audit rule inspects your command line and
+decides whether it may run (agent PreToolUse hook, commit gate, sudo command
+pattern, CI policy check); the gate blocked or mis-parsed a command that is
+correct as written; or you are composing a command that must satisfy such a gate
+on the first attempt.
+
+## Do this
+
+1. **Write values the gate reads as literal text.** The gate is handed the command
+   as an unexecuted string — a Claude Code `PreToolUse` hook receives
+   `tool_input.command` (`{"tool_name":"Bash","tool_input":{"command":"npm test"}}`)
+   and runs "Before a tool call executes". Nothing has expanded yet: `"$VAR"` is
+   four literal characters plus a name, and the quote marks are part of the text.
+   Shell expansion is defined to happen when the shell processes the line, which is
+   after the gate has already decided.
+
+2. **Know which of the two failure modes you are in — the error message tells you.**
+   Both a quoted path and an unexpanded variable break a gate that extracts a file
+   argument, but for different reasons and with different symptoms:
+
+| What you wrote | What the gate extracts | How it fails |
+|----------------|------------------------|--------------|
+| `--body-file /abs/path/REPORT.md` | `/abs/path/REPORT.md` | Passes |
+| `--body-file="/abs/path/REPORT.md"` or `--body-file "$REPO/REPORT.md"` | nothing — the extraction pattern excludes the quote character, so the match fails outright | Gate reports the argument as **missing** ("no `--body-file` found"), which reads as a malformed command |
+| `--body-file $REPO/REPORT.md` (unquoted variable) | the literal string `$REPO/REPORT.md` | Gate reports the file as **nonexistent**, which reads as a missing deliverable |
+
+3. **Create the file a gate will read in an earlier, separate command.** The gate
+   runs before this command executes, so a file produced by a heredoc inside the
+   same command does not exist yet at inspection time and the gate fails closed.
+   Write the file in one call, reference it by literal path in the next.
+
+4. **When content must contain patterns the gate treats as dangerous, put the
+   content in a file with a non-shell tool and pass the path.** Release notes,
+   docs, or fixtures containing `curl … | sh` or `rm -rf` are data, but a
+   text-scanning gate cannot tell data from an invocation. A file written by an
+   editor/Write tool is never scanned as a command; `--notes-file` / `--body-file`
+   then carries it.
+
+5. **Read the gate's own extraction pattern when a correct command is refused.**
+   The pattern is the specification of what the gate can see. Reproduce it against
+   your exact command string before rewriting anything else — one run tells you
+   whether you are in the missing-argument or nonexistent-file mode above.
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| Blocking feedback appears without the gate's message | Exit code 2 sends the reason to **stderr**, not stdout; read stderr for the actual cause |
+| The gate matches an intended-as-prose mention of a dangerous command (in a commit message, doc, or test fixture) | Move the text into a file and pass it by path (step 4) rather than reshaping the sentence |
+| Path contains a space, so quoting is unavoidable | Relocate or symlink the target to a space-free path for gated commands; a gate that excludes quote characters cannot receive a quoted path at all |
+| The gate needs `~` expanded | Write the absolute path; a gate that resolves `~` itself is doing so on the literal tilde, which only works if it implements the expansion |
+| The same command must also be portable/robust as a script | Keep the gate-read argument literal and leave the rest of the script quoted normally ([platforms-shells-portable-shell-scripts]) — this page narrows one argument, it does not license unquoted expansions elsewhere |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Interpolate `"$VAR/file"` into an argument a gate inspects | Write the resolved absolute path literally | The gate sees pre-expansion text; the quote character can defeat its extractor entirely and the variable name never resolves for it |
+| Assume a blocked command means the deliverable is wrong | Reproduce the gate's extraction pattern against your literal command string first | A quoting-level extraction failure and a genuinely incomplete deliverable produce the same refusal, so fixing content wastes the round |
+| Build the file the gate checks with a heredoc in the same command | Write it in a prior command and reference the path | The gate is evaluated before execution, so the file is absent at decision time |
+| Reword prose to get a dangerous-looking string past a scanner | Put the prose in a file and pass `--notes-file`/`--body-file` | Editing meaning to satisfy a text scanner degrades the artifact; a file is not scanned as a command |
+
+## Sources
+
+- https://code.claude.com/docs/en/hooks — `PreToolUse` runs "Before a tool call executes. Can block it"; the hook's stdin JSON carries `tool_input.command` — the unexecuted Bash command string. Exit 2 blocks and "stderr text is fed back to Claude as an error message"
+- https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html — the shell's order of word expansion (tilde, parameter, command substitution, field splitting, quote removal) is performed by the shell as it processes the command, so an external reader of the command text sees none of it applied
+
+## Field context
+
+Reproduced against the extraction pattern of this repo's own flush gate
+(``--body-file[= ]+[^ '"`]+``, `hooks/pre-flush-pr-gate.sh`) on 2026-07-30: five
+variants run through that pattern gave `--body-file "$REPO/…"` → empty (blocked as
+missing), `--body-file $REPO/…` → literal `$REPO/…` (blocked as nonexistent),
+`--body-file "/abs/…"` → empty even with a literal path, while
+`--body-file /abs/…` and `--body-file=/abs/…` extracted correctly. A same-command
+heredoc body-file was separately blocked as not-yet-existing until moved to a
+preceding call.
diff --git a/wiki/platforms/shells/portable-shell-scripts.md b/wiki/platforms/shells/portable-shell-scripts.md
index dcab402..23eed9b 100644
--- a/wiki/platforms/shells/portable-shell-scripts.md
+++ b/wiki/platforms/shells/portable-shell-scripts.md
@@ -11,7 +11,7 @@ sources:
   - https://google.github.io/styleguide/shellguide.html
   - https://www.shellcheck.net/
 last_verified: 2026-07-10
-related: [platforms-tools-bsd-vs-gnu-cli, platforms-toolchains-version-management]
+related: [platforms-tools-bsd-vs-gnu-cli, platforms-toolchains-version-management, platforms-shells-command-text-inspected-before-execution]
 ---
 
 # Shell Scripts That Must Run on More Than One Machine or Shell
diff --git a/wiki/qa/document-verification/editing-a-gated-document.md b/wiki/qa/document-verification/editing-a-gated-document.md
new file mode 100644
index 0000000..a3e0387
--- /dev/null
+++ b/wiki/qa/document-verification/editing-a-gated-document.md
@@ -0,0 +1,107 @@
+---
+id: qa-document-verification-editing-a-gated-document
+domain: qa
+category: document-verification
+applies_to: [general]
+confidence: field-tested
+sources:
+  - https://man7.org/linux/man-pages/man1/pgrep.1.html
+  - https://docs.vale.sh/topics/scopes.md
+  - https://docs.vale.sh/checks/existence
+  - https://github.com/DavidAnson/markdownlint/blob/main/doc/md013.md
+last_verified: 2026-07-30
+related: [qa-process-acceptance-criteria, qa-process-regression-scope, testing-quality-tests-that-cannot-fail]
+---
+
+# Editing a Document That Automated Text Gates Check
+
+## When this applies
+
+You are adding to or rewording a document (RFC, spec, plan, audit record) that
+grep/regex gates or a lint config run over; you are recording an audit result
+*inside* the document the audit examined; or a check that passed before your edit
+now fails on wording whose meaning did not change.
+
+## Do this
+
+1. **Inventory the anchors before editing.** Grep the gate definitions for patterns
+   naming this file and treat every hit as a constraint on the edit. Anchors come in
+   kinds that break differently:
+
+| Anchor kind | Example | How an edit breaks it |
+|-------------|---------|-----------------------|
+| Phrase or vocabulary | a pattern requiring/forbidding certain verbs near a term | Reworded sentence stops matching, or starts matching a forbidden pattern |
+| Line or row count | "file has 444 lines", "table has 5 rows" | Any insertion shifts the count |
+| Substring | a section title or term quoted verbatim | Retitling or translating the term silently drops it |
+| Quoted original | a check that requires an upstream sentence to appear verbatim | Paraphrasing removes it while preserving meaning |
+
+2. **State facts about an upstream document as observations, not definitions.** A
+   lexical gate that means "do not redefine the upstream contract" can only
+   approximate that with vocabulary: Vale's `existence` check "looks for the
+   'existence' of particular tokens" and evaluates no meaning. So a true, purely
+   descriptive sentence trips it when it uses a defining verb. Write the observable
+   shape instead — "the contract table has two rows, `arena` and `pool`, and no
+   `heap` row" — which carries the same information and matches no definition
+   pattern.
+
+3. **Scope a check outside the region that quotes it.** When the document contains
+   the check command or its regex, the pattern matches its own quotation. Bound the
+   judgment to the normative region (`awk '/^### <audit heading>/{exit}'`, a line
+   range, or a section selector) rather than running it over the whole file. This is
+   standard practice in the tools themselves: Vale scopes are markup-aware and "Any
+   scope prefaced with `~` is negated", and markdownlint rules take
+   `code_blocks: false` to exclude quoted code from prose rules. `pgrep` hard-codes
+   the same defense — "The running pgrep, pkill, or pidwait process will never
+   report itself as a match."
+
+4. **Record verdicts as scoped conditions, never as a global count.** Write "0
+   matches outside §Audit" or "every match is inside a fenced block", not "1 match
+   found". A fixed number is invalidated by the next edit — including the edit that
+   quotes the finding — so a count-based verdict decays into a false statement while
+   looking precise.
+
+5. **Re-run the whole gate set after the edit and compare against the pre-edit
+   pass count.** Your edit can break a check owned by a different section or a
+   different task ([qa-process-regression-scope]). A baseline number (`60 → 61`) is
+   what distinguishes "my change fixed one" from "my change fixed one and broke two".
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| The quoted pattern must stay in the document (it is the audit record) | Keep the quote and narrow the check's scope; the document's job is to be readable, the check's job is to be scoped |
+| Fixing the report by quoting the offending match adds another match | Stop counting globally and switch to the scoped condition in step 4 — each quote-to-fix round otherwise raises the count and re-falsifies the statement |
+| Gate anchors on a line count you must change | Update the gate and the document in the same commit, and say so in the PR ([qa-process-acceptance-criteria]) |
+| The failure surfaces in another task's or another agent's check log | Attribute before repairing: identify which file the failing pattern targets, since a cross-file gate makes your edit look like their regression |
+| The gate is genuinely wrong (it forbids a correct sentence with no replacement available) | Change the gate, with a control proving it still catches the defect it owns — do not reword a correct document into a worse one |
+| You are authoring the gate rather than the document | This page covers the author's side; gate construction and its controls are a separate concern → [testing-quality-tests-that-cannot-fail] |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Edit a gated document and run only the check you were fixing | Grep the gates for anchors on that file first, then re-run the full set against a baseline count | Line-count and substring anchors break silently, and the breakage lands in a check you never looked at |
+| Write "X defines/mandates two variants" when describing an upstream contract | Write the observable shape: "the table has two rows and no Y row" | A vocabulary gate cannot separate describing from redefining, so the defining verb fails a sentence whose content is correct |
+| Run an audit pattern over the whole document that quotes it | Bound it to the normative region by heading range or line range | The quotation is a match, so the whole-file run measures the document's prose about itself |
+| Record "grep found 0 hits" as the audit result | Record the scoped condition that must hold ("no hits outside §Audit") | The bare count is true only for the file revision that produced it and silently becomes false on the next edit |
+
+## Sources
+
+- https://docs.vale.sh/checks/existence — the check "looks for the 'existence' of particular tokens", transformed into a word-bounded non-capturing group: a lexical gate matches patterns, not intent
+- https://docs.vale.sh/topics/scopes.md — scopes restrict where a rule applies via markup-aware selectors; "Any scope prefaced with `~` is negated" and scopes can be chained, so checks can be kept off regions like code examples
+- https://github.com/DavidAnson/markdownlint/blob/main/doc/md013.md — rules expose `code_blocks`, `tables`, `headings` booleans (default `true`) so quoted code can be excluded from a prose rule
+- https://man7.org/linux/man-pages/man1/pgrep.1.html — "The running pgrep, pkill, or pidwait process will never report itself as a match" — self-exclusion is designed in because self-matching is the expected failure
+
+## Field context
+
+Distilled from 2026-07 RFC/plan-authoring sessions in this repo. A vague-word audit
+of `docs/ROADMAP.md` reported 1 global hit — the audit command's own quoted pattern —
+and quoting the finding to fix it raised the count to 3, while an `awk`-scoped run
+over the normative region reported 0; the same self-reference appeared twice more
+(a negative-control example that always matched, and a `docs/*.md` glob matching its
+own document 45 times). Separately, adding the sentence "defines only the two
+variants arena and pool" to an RFC failed a sibling task's vocabulary gate
+(`(arena|pool)…(재정의|정의한다|규정한다)`); rephrasing to "the contract table has
+two rows and no heap row" restored the suite from 60 to 61 passing, and a pre-edit
+anchor survey of that file found three further anchors — a line count, a section
+substring, and a verbatim upstream sentence — all preserved by the same edit.
diff --git a/wiki/qa/index.md b/wiki/qa/index.md
index 92b7105..a87d73c 100644
--- a/wiki/qa/index.md
+++ b/wiki/qa/index.md
@@ -17,6 +17,12 @@ Match your situation to a "load when" line; load only matching pages.
 | [severity-and-priority](process/severity-and-priority.md) | Triaging a bug — deciding how bad it is and when it gets fixed; a triage stalled on a severity debate |
 | [post-release-verification](process/post-release-verification.md) | A release just deployed to production; defining what "released safely" means; an incident revealed a release was broken for hours before anyone noticed |
 
+## document-verification
+
+| Page | Load when |
+|------|-----------|
+| [editing-a-gated-document](document-verification/editing-a-gated-document.md) | Editing or rewording a document that grep/regex gates or a lint config check; a gate fails on wording whose meaning did not change; describing what an upstream spec says without tripping a "do not redefine it" gate; a check matches the pattern your own document quotes; recording an audit verdict inside the document that was audited; deciding which checks to re-run after editing a gated document |
+
 ## environments
 
 | Page | Load when |

From 0f848d8bf2beff4af6400a8ae8cbc033acd46a31 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: Thu, 30 Jul 2026 17:47:44 +0900
Subject: [PATCH 2/2] knowledge: add Decision Log block to the flush report

Required by the RTB decision-log PR gate: intent, rejected alternatives, and
reviewer focus points (including the qa/index.md conflict with PR #10).
---
 .dev-loop/INGEST_REPORT.md | 42 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/.dev-loop/INGEST_REPORT.md b/.dev-loop/INGEST_REPORT.md
index fd8052a..051256b 100644
--- a/.dev-loop/INGEST_REPORT.md
+++ b/.dev-loop/INGEST_REPORT.md
@@ -161,3 +161,45 @@ Body lines 83 / 82 / 93 (limit 120) · all four required sections present on eac
 every `related:` id and inline `[page-id]` resolves (13 checks, 0 misses) · each new page
 listed in its domain index with a multi-use-case "load when" · every index relative link
 resolves · no banned vague qualifiers · `log.md` entry appended.
+
+## Decision Log
+
+**의도**
+- 큐에 쌓인 4건을 3페이지로 인그레스트한다. qa 2건(자기참조 grep · 어휘 게이트/앵커)은
+  "게이트가 걸린 문서를 편집하는 쪽"이라는 **한 케이스**라서 1페이지로 합쳤다
+  (`AGENTS.md` "one case per page" + 지시문이 서로를 필요로 함).
+- 후보의 *why*를 그대로 받지 않고 실측으로 교정했다: `--body-file` 차단 원인은
+  "따옴표 때문"이 아니라 **따옴표=추출 실패 / 미확장 변수=경로 부재**의 두 모드이고
+  따옴표 친 리터럴 경로도 실패한다(본문 표에 5변형 실측).
+- 열린 PR(#6~#10)까지 중복 검사 대상에 포함했다. `main`만 보면 커버리지를 과소평가해
+  #10과 같은 페이지를 또 만들 위험이 있었다.
+- `INDEX.md`는 **의도적으로 건드리지 않았다** — #10이 이미 qa 라우트 라인을 갱신하므로
+  충돌면을 `wiki/qa/index.md` 한 훅으로 줄였다.
+
+**배제한 대안**
+- *insight 1을 `background-services`에 병합* → 기각. 그 페이지의 케이스는 *지속성*이고,
+  블로킹 stdin을 넣으면 "load when"이 흐려진다(양방향 링크로 대체).
+- *insight 2를 `portable-shell-scripts`에 병합* → 기각. 그 페이지 규칙은 "모든 확장을
+  따옴표로 감싸라"인데 이 건은 한 인자를 리터럴로 쓰라는 것 — 트리거도 다르고 규칙이
+  충돌해 보인다. 새 페이지에 "이 페이지는 인자 하나만 좁히며 다른 곳의 무따옴표 확장을
+  허용하지 않는다"를 명시.
+- *insight 2용 새 카테고리(`policy-gates`)* → 기각. 두 번째 멤버가 안 보이는 1페이지
+  카테고리. 메커니즘이 확장 타이밍이라 `shells`가 맞다.
+- *insight 1을 `debugging`으로* → 기각. 바꾸는 산출물이 호출 명령이므로 소유 도메인 우선
+  규칙에 따라 platforms(진단 절반은 `debugging-methodology-reproduce-first`로 링크).
+- *qa 2건을 #10 머지까지 보류* → 기각. 범위 축소는 소유자 판단이므로, 대신 충돌 위치와
+  해소법(헤딩 1개 + 행 2개)을 위에 명시했다.
+- *#10 브랜치를 base로 스택 PR* → 기각. 스킬이 `--base main`을 요구하고, 리뷰 diff에
+  #10 변경이 섞인다.
+
+**리뷰어가 볼 곳**
+1. `wiki/qa/index.md` — #10과 충돌 예정 지점. **헤딩 1개 유지 + 행 2개**로 해소.
+2. `wiki/qa/document-verification/editing-a-gated-document.md` — #10의 페이지와
+   중복이 아닌지(저자 측 vs 게이트 저자 측 분담)가 이 PR의 핵심 판단.
+3. `command-text-inspected-before-execution.md`의 Field context 5변형 표 —
+   `hooks/pre-flush-pr-gate.sh:56` 패턴으로 재현 가능.
+4. confidence 등급: 1·2는 `verified`(공식 문서+재현), 3+4는 `field-tested`
+   (위험은 문서화되어 있으나 지시문은 현장 유래). 1번 페이지의 클라이언트/서버 분리
+   단계는 문서 근거가 아니라 현장 유래라서 Field context로 분리해 표기했다.
+5. **[추정]** `platforms/shells`에 2페이지(이식성 + 확장 타이밍)가 앞으로도 응집적으로
+   남을 것이라는 판단은 추정이다 — 유사 사례가 더 쌓이면 별도 카테고리가 나을 수 있다.