diff --git a/.dev-loop/INGEST_REPORT.md b/.dev-loop/INGEST_REPORT.md
index 2ff3871..83fcbb5 100644
--- a/.dev-loop/INGEST_REPORT.md
+++ b/.dev-loop/INGEST_REPORT.md
@@ -1,35 +1,173 @@
-# 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`."
+Queue drained: 4 pending candidates from 2 sessions (`19484655…`, `dd51281c…`,
+harvested 2026-07-30). They land as **2 pages** — three candidates describe one
+situation (gates on a spec document) and merge into a single page; the fourth is a
+distinct mechanism (Unicode text matching) and gets its own.
+
+| # | Candidate (trigger) | Target |
+|---|---------------------|--------|
+| 1 | Fixing a grep gate for a document that does not exist yet | `qa/document-verification/spec-document-gates.md` (§Do this 1) |
+| 2 | Verifying an RFC/API-spec's requirements with grep gates | same page (§Do this 3, four axes) |
+| 4 | Prose-document checklist that keyword checks pass vacuously | same page (§Do this 3–4, modality/polarity + fail-closed) |
+| 3 | grep/regex over Korean text, stem-prefix pattern | `platforms/environment/unicode-text-matching.md` (new page) |
 
 ## 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**
+### Candidates 1, 2, 4 — gates on a specification document → `confidence: field-tested`
+
+**Claim.** A token-existence grep is not sufficient evidence that a document meets a
+requirement; gates need a positive control (a conforming sibling the pattern must
+match), a negative control (a mutated copy it must reject), four check axes
+(structure / modality+polarity / closed-set completeness / cross-reference), and
+fail-closed behaviour when the anchor is missing.
+
+**Sources checked.** Each mechanism is grounded in a tool or spec that implements it:
+
+| Directive | Source | What it says (checked, not assumed) |
+|-----------|--------|--------------------------------------|
+| Positive **and** negative control per gate | [eslint.org custom-rule-tutorial](https://eslint.org/docs/latest/extend/custom-rule-tutorial) (fetched) | Verbatim: "RuleTester requires that at least one valid and one invalid test scenario be present." |
+| Structure of a checker's test cases | [eslint.org nodejs-api](https://eslint.org/docs/latest/integrate/nodejs-api) (fetched) | `RuleTester#run()` takes `valid` + `invalid` arrays; invalid cases assert the expected error count |
+| Mutants prove detection, coverage does not | [testing.googleblog.com mutation testing](https://testing.googleblog.com/2021/04/mutation-testing.html) | Inserting faults and requiring failure measures whether checks detect bugs (already cited by `testing-quality-tests-that-cannot-fail`) |
+| MUST→SHOULD demotion changes the requirement while keeping every keyword | [RFC 2119](https://www.rfc-editor.org/rfc/rfc2119) (fetched) | MUST = "an absolute requirement of the specification"; SHOULD = there "may exist valid reasons … to ignore a particular item" |
+| Cross-reference axis is a real check class | [docs.vale.sh/checks/conditional](https://docs.vale.sh/checks/conditional) (fetched) | Verbatim: "Ensures that the existence of 'first' implies the existence of 'second'." |
+| Count + scope, not bare existence | [docs.vale.sh/checks/occurrence](https://docs.vale.sh/checks/occurrence) (fetched) | Enforces "the maximum or minimum number of times a particular token can appear in a given scope" (`scope: sentence`) |
+| Table structure gets parsed, not grepped | [markdownlint MD056](https://github.com/DavidAnson/markdownlint/blob/main/doc/md056.md) | Flags tables whose rows disagree with the header column count |
+| A delimiter count is not a parse | [markdownlint issue #1206](https://github.com/DavidAnson/markdownlint/issues/1206) | MD056 counts pipes inside backticks as separators and misreports — the failure mode of approximating a parse |
+
+**How verified.** Every ESLint/Vale/RFC quote above was fetched from the primary
+source in this session, not recalled. The *composite* methodology (the four-axis
+framing and its ordering) comes from the two harvested sessions, where an
+independent auditor built tampered copies that a 16-gate existence checklist
+accepted (5 of 8 passed all 16, including one with a row deleted from a 5-row
+`type` enum) and, after the axes were added, 32/32 mutants failed their designated
+check while the intact document passed 62/62.
+
+**Resulting confidence: `field-tested`, deliberately not `verified`.** Each
+individual directive is doc-backed, but the composite rule's evidence is in-house
+session experience a reader cannot re-run, so the page carries a `Field context`
+section stating that provenance instead of claiming external verification.
+
+### Candidate 3 — Unicode text matching → `confidence: verified`
+
+**Claim.** A stem-prefix pattern silently returns 0 hits on precomposed text
+(`아니` is not inside `아닌`), and `grep` applies no canonical equivalence, so an
+NFC pattern misses NFD data entirely.
+
+**How verified — reproduced locally** (macOS 15 / APFS, `grep` 2.6.0-FreeBSD, Python 3.13, 2026-07-30):
+
+```
+NFC 아닌 = U+C544 U+B2CC     NFC 아니 = U+C544 U+B2C8     substring? False
+NFD 아닌 = U+110B U+1161 U+1102 U+1175 U+11AB              substring? True   ← form-dependent
+grep -c '아니' <NFC file> → 0      grep -c '아닌' <NFC file> → 1   (same line)
+grep -c -f <NFC pattern> <NFD file> → 0      -f <NFD pattern> → 1
+```
+
+The NFC/NFD asymmetry was **not** in the harvested candidate; the reproduction
+surfaced it, and the page states it because it decides whether the naive pattern
+fails at all. A second reproduction (creating an NFC-named file, listing it, and
+opening it by both forms) confirmed APFS returns the name as written and resolves
+either form — matching Apple's documented behaviour rather than the folklore that
+macOS always decomposes.
 
-**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**
+**Sources checked.**
+- [UAX #15](https://unicode.org/reports/tr15/) (fetched) — Hangul syllables have special full-decomposition rules and are maintained under all normalization forms; canonically equivalent strings have different binary representations unless normalized.
+- [Unicode 16.0 core spec, chapter 3](https://www.unicode.org/versions/Unicode16.0.0/core-spec/chapter-3/) — §3.12 Conjoining Jamo Behavior: 11,172 precomposed Hangul syllables from `SBase = U+AC00`, algorithmic L/V/T decomposition.
+- [Apple File System Guide FAQ](https://developer.apple.com/library/archive/documentation/FileManagement/Conceptual/APFS_Guide/FAQ/FAQ.html) — APFS preserves the file name's normalization and is normalization-insensitive via hashes of the normalized form; HFS+ stores the normalized form.
+- [POSIX grep](https://pubs.opengroup.org/onlinepubs/9699919799/utilities/grep.html) — matching is specified by regular-expression rules over input lines; no canonical-equivalence folding.
 
-**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**
+The candidate's own note that its first diagnosis ("BSD grep ERE multibyte bug")
+was wrong is preserved as behaviour, not blame: the page's edge-case row says to
+dump code points before concluding the text is absent.
 
-**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).
+**Nothing left `unverified`.** No candidate was dropped.
 
 ## 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.
+**Pages read in full:** `INDEX.md`, `AGENTS.md`, `templates/page.md`,
+`skills/wiki-ingest/SKILL.md`, `wiki/qa/index.md`, `wiki/testing/index.md`,
+`wiki/platforms/index.md`, `wiki/debugging/index.md`,
+`wiki/testing/quality/tests-that-cannot-fail.md`, `wiki/qa/process/release-gates.md`,
+`wiki/qa/process/acceptance-criteria.md` (frontmatter + structure),
+`wiki/platforms/environment/timezone-and-locale.md`,
+`wiki/platforms/filesystems/paths-case-and-line-endings.md`.
+**Repo-wide greps** for `grep`, `unicode|normali[sz]|codepoint|multibyte|NFC|NFD`,
+and `specification|RFC|design doc|acceptance criteri` to find any page already
+holding this material.
+
+| Overlap candidate | Verdict |
+|---|---|
+| `testing-quality-tests-that-cannot-fail` | **Closest neighbour, not a duplicate.** It owns "prove a *test* can fail by mutating the code"; the new page owns "prove a *document gate* can pass and can fail". Same principle, different artifact and different failure table. Linked **both ways**, and the new page's edge-case table routes code deliverables back to it. |
+| `qa-process-release-gates` | Human release checklist (suite green / smoke / regression / rollback). No overlap with automated checks on a document's text. Left untouched. |
+| `qa-process-acceptance-criteria` | Adjacent: it defines testable criteria for a ticket; the new page verifies a document against them. Back-link added; the new page's "document deliberately relaxes a requirement" row points at it. |
+| `platforms-environment-timezone-and-locale` | Covers locale-dependent **casing and collation** (Turkish `İ`/`ı`, `LC_COLLATE`) — the environment as hidden input. Normalization form is a property of the **data**, not the environment, and the page is already 81 lines with a different "load when". Kept separate, back-linked. |
+| `platforms-filesystems-paths-case-and-line-endings` | Covers case-collisions and CRLF; mentions path normalization only as "use the platform's path API". Does not cover Unicode form. Back-link added. |
+| `platforms-tools-bsd-vs-gnu-cli` | Owns userland flag differences. The new page cites it inline for the multi-byte character-class caveat rather than restating it. |
+
+**Conflicts flagged: none.** No existing directive contradicts either new page.
+
+**Reciprocal `related:` links added** to `testing-quality-tests-that-cannot-fail`,
+`qa-process-acceptance-criteria`, `platforms-environment-timezone-and-locale`, and
+`platforms-filesystems-paths-case-and-line-endings`.
 
 ## 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`.
+### 1. `qa` / **new category `document-verification`** / `spec-document-gates.md`
+
+- **Domain — `qa` over `testing`.** `INDEX.md` scopes `testing` to *automated test
+  code for software behaviour*; the artifact here is a written deliverable and the
+  question is "does it meet its acceptance criteria, and is the gate that decides
+  that trustworthy" — which is `qa`'s remit (acceptance criteria, gates). Two of the
+  four candidates self-declared `qa`; the third declared `testing` and is routed
+  with them because it is a directive about the same gates.
+- **New category justified.** Existing `qa` categories are `process` (human release
+  workflow), `environments` (staging parity), `bug-reports`, `exploratory` — none
+  covers automated verification of an artifact. Folding it into `process` would put
+  a script-authoring page among human-workflow pages and blur that index's routing
+  lines. `document-verification` is where doc linting and link checking will also land.
+- **Merged 3 candidates into 1 page** per "one case per page": all three answer the
+  same situation (authoring/reviewing gates over a spec document) — candidate 1 is
+  the *control* step, candidate 2 the *what to check* step, candidate 4 the
+  *polarity + fail-closed* step. Splitting them would produce three pages sharing one
+  "load when" line, which is what the merge-before-create rule exists to prevent.
+
+### 2. `platforms` / existing category `environment` / `unicode-text-matching.md`
+
+- **Domain — `platforms` over `qa`** (the candidate's own hint): the mechanism is
+  text representation and its OS/producer-dependent form, not release quality. The
+  `platforms` index already owns "hidden environment inputs" and cross-OS filesystem
+  behaviour, and `environment/` is the sibling of `timezone-and-locale`, the existing
+  page for text operations that behave differently than assumed.
+- **No new category** — `environment` fits, so none was created.
+- Cross-linked to the `qa` page so a doc-gate author writing a Korean pattern is
+  routed here, and back so a 0-hit debugger reaches the gate-design rules.
+
+### Plumbing updated
+
+`wiki/qa/index.md` (new category block + route-line scope), `wiki/platforms/index.md`
+(page row + intro scope), `INDEX.md` (both domain route lines), `log.md` (2 ingest
+entries). Checked before commit: both pages under the 120-line body limit (77 / 54),
+every `related:` and inline `[page-id]` resolves, both pages listed in their domain
+index, no banned vague qualifiers, no prohibition without a paired replacement.
+
+## Decision Log
+
+- **의도** — 큐 4건을 드롭 없이 위키에 넣되 중복 페이지를 만들지 않도록 3건을 한
+  페이지로 병합하고, 각 지시문을 1차 출처로 검증했다. 하베스트된 주장을 그대로 옮기지
+  않고 ESLint·Vale·RFC 2119·markdownlint 원문을 실제로 열어 인용했다.
+- **배제** — ⓐ 4건을 4페이지로 쪼개는 안(“one case per page” 위반, load-when 라인이
+  겹침) ⓑ Unicode 건을 `timezone-and-locale`에 병합하는 안(그 페이지는 *환경*이 입력인
+  경우를 다루고 정규화형은 *데이터*의 속성 — 트리거가 다름) ⓒ 문서 게이트 페이지를
+  `qa/process`에 넣는 안(사람 프로세스 페이지 사이에 스크립트 작성 지침이 섞임)
+  ⓓ 복합 방법론에 `verified`를 부여하는 안 — 개별 지시문은 출처가 있으나 4축 조합의
+  근거는 사내 세션이라 `field-tested`로 낮추고 그 사실을 페이지에 명시.
+- **리뷰포인트** — ① 새 카테고리 `qa/document-verification` 신설이 과한지(대안: `process`
+  편입) ② 후보 3건 병합 경계가 맞는지 ③ `field-tested` 등급 판단 ④ Unicode 페이지의
+  NFC/NFD 비대칭은 원 후보에 없던 내용으로 이번 재현으로 추가된 것 — 원 후보 취지를
+  넘어서는지.
+
+Cross-Check: 각 지시문을 1차 출처 원문 fetch로 대조(ESLint RuleTester 문장·RFC 2119
+MUST/SHOULD 정의·Vale conditional/occurrence 설명·UAX #15·APFS FAQ), Unicode 주장은
+로컬 재현 2종(코드포인트 덤프 + grep 4회, APFS 파일명 생성/조회)으로 실측. 이번 세션은
+서브에이전트 사용이 배제되어 독립 에이전트 적대검증은 돌리지 않았음 — 검증은 출처
+대조와 재현에 근거한다.
diff --git a/INDEX.md b/INDEX.md
index 05789bd..aac33dc 100644
--- a/INDEX.md
+++ b/INDEX.md
@@ -14,10 +14,10 @@ follow the cross-pointers in their index or take the next matching seeded domain
 | [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) |
-| [qa](wiki/qa/index.md) | **seeded** | Release-quality process: release gates, regression scoping, bug reports, severity/priority triage, exploratory testing (writing automated test code → testing) |
+| [qa](wiki/qa/index.md) | **seeded** | Release-quality process: release gates, regression scoping, bug reports, severity/priority triage, exploratory testing, automated verification of document deliverables (spec/RFC gates) (writing automated test code → testing) |
 | [debugging](wiki/debugging/index.md) | **seeded** | Diagnosing a failure — finding what is wrong and why: reproducing, bisection, hypothesis testing, traces/logs, intermittent failures (fixing the diagnosed fault → its owning domain) |
 | [security](wiki/security/index.md) | **seeded** | Trust-boundary decisions: input validation, session-vs-token auth choice, per-resource authorization (IDOR), secrets hygiene, dependency trust, PII handling (XSS rendering → frontend; CI secrets → infrastructure; JWT implementation → backend/frontend auth) |
-| [platforms](wiki/platforms/index.md) | **seeded** | OS-level differences breaking code across macOS/Linux/Windows: shell portability, BSD-vs-GNU CLI, filesystem case/line endings, background services/cron, toolchain version pinning |
+| [platforms](wiki/platforms/index.md) | **seeded** | OS-level differences breaking code across macOS/Linux/Windows: shell portability, BSD-vs-GNU CLI, filesystem case/line endings, Unicode normalization in text/file-name matching, background services/cron, toolchain version pinning |
 | [mobile](wiki/mobile/index.md) | **seeded** | App-side iOS/Android/cross-platform: process death/state survival, offline-first sync, mobile-network calls, store rollout/hotfix strategy, startup time |
 
 All ten domains are seeded. New categories grow via `skills/wiki-ingest/SKILL.md`.
diff --git a/log.md b/log.md
index 587a5dc..9702020 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).
+## [2026-07-30] ingest | qa +1 (new category document-verification): document-verification/spec-document-gates — automated gates that decide whether a spec/RFC meets its requirements. Positive control on a conforming sibling before adopting a pattern (an absent target fails every pattern, so Red proves nothing), negative control per gate kept as a permanent mutant self-test, four check axes (table structure / sentence-scoped modality+polarity / closed-set exact count / cross-reference recomputation), fail-closed on a missing anchor. Derived from RFC-authoring sessions where an independent auditor's 8 tampered copies passed a 16-gate token-existence checklist (5/8 fully); after the axes, 32/32 mutants failed and the intact doc passed 62/62. Mechanisms sourced to ESLint RuleTester (valid+invalid required), Google mutation testing, RFC 2119 (MUST vs SHOULD), Vale conditional/occurrence checks, markdownlint MD056 + issue #1206.
+## [2026-07-30] ingest | platforms +1 (environment): environment/unicode-text-matching — non-ASCII grep/regex patterns. Precomposed syllables are single code points so a stem prefix is not a substring of its inflected form (`아니` not in `아닌`), and grep applies no canonical equivalence so an NFC pattern scores 0 on NFD data; enumerate surface forms, normalize both sides on write, assert expected hit counts, normalize file-name lists before cross-OS diffs. Reproduced on macOS 15/APFS 2026-07-30 (code-point dump + 4 grep runs); sourced to UAX #15, Unicode core spec §3.12 (11,172 precomposed Hangul syllables from U+AC00), APFS FAQ (preserving + normalization-insensitive; HFS+ stores a normalized form), POSIX grep.
diff --git a/wiki/platforms/environment/timezone-and-locale.md b/wiki/platforms/environment/timezone-and-locale.md
index a4927ef..729de7c 100644
--- a/wiki/platforms/environment/timezone-and-locale.md
+++ b/wiki/platforms/environment/timezone-and-locale.md
@@ -14,7 +14,7 @@ sources:
   - https://unicode.org/faq/casemap_charprop.html
   - https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap07.html
 last_verified: 2026-07-10
-related: [databases-schema-design-column-data-types, platforms-processes-background-services]
+related: [databases-schema-design-column-data-types, platforms-processes-background-services, platforms-environment-unicode-text-matching]
 ---
 
 # Timezone and Locale as Hidden Inputs to Date and Text Code
diff --git a/wiki/platforms/environment/unicode-text-matching.md b/wiki/platforms/environment/unicode-text-matching.md
new file mode 100644
index 0000000..cccb83b
--- /dev/null
+++ b/wiki/platforms/environment/unicode-text-matching.md
@@ -0,0 +1,68 @@
+---
+id: platforms-environment-unicode-text-matching
+domain: platforms
+category: environment
+applies_to: [general]
+confidence: verified
+sources:
+  - https://unicode.org/reports/tr15/
+  - https://www.unicode.org/versions/Unicode16.0.0/core-spec/chapter-3/
+  - https://developer.apple.com/library/archive/documentation/FileManagement/Conceptual/APFS_Guide/FAQ/FAQ.html
+  - https://pubs.opengroup.org/onlinepubs/9699919799/utilities/grep.html
+last_verified: 2026-07-30
+related: [platforms-environment-timezone-and-locale, platforms-filesystems-paths-case-and-line-endings, qa-document-verification-spec-document-gates]
+---
+
+# Matching Non-ASCII Text with grep and Regex
+
+## When this applies
+
+You are writing a grep/regex pattern that must match non-ASCII text (Korean,
+Japanese, accented Latin, emoji) in files, log lines, or file names; a pattern that
+looks correct returns zero hits on text you can see on screen; a search works on one
+machine and misses on another after the file crossed an OS, archive, or editor boundary.
+
+## Do this
+
+| Case | Do |
+|------|----|
+| Writing a literal pattern | Copy the exact substring out of the file rather than typing a stem you expect to be a prefix. In a precomposed script each syllable is a single code point: `아니` (`U+C544 U+B2C8`) is not contained in `아닌` (`U+C544 U+B2CC`), so a stem-prefix pattern returns 0 hits on a line that plainly contains the word |
+| The word occurs in several inflected forms | Enumerate the forms as alternatives — `(아닌\|아니다\|아니라)` — one alternative per surface form present in the text |
+| Pattern and data can come from different producers (editor, export, archive, another OS) | Normalize both sides to the same form before comparing (NFC for stored text), because `grep` matches code-unit sequences and applies no canonical equivalence: an NFC pattern scores 0 against the same word stored as NFD |
+| The match result gates a build, release, or verification checklist | Assert an expected non-zero count rather than "no error", so a normalization or stem mistake surfaces as a failed count ([qa-document-verification-spec-document-gates]) |
+| Comparing file names collected on macOS and Linux | Normalize both name lists in your code before diffing: APFS preserves the normalization it was given and looks up either form, HFS+ stores its own normalized form, and Linux filesystems store the bytes given — so the same name reaches your comparison in different forms |
+
+Measured on macOS 15 / APFS, 2026-07-30 (`grep` 2.6.0-FreeBSD, Python 3.13):
+
+```
+grep -c '아니' <NFC file>            → 0     # same line contains 아닌
+grep -c '아닌' <NFC file>            → 1
+grep -c -f <NFC pattern> <NFD file>  → 0     # no normalization by grep
+grep -c -f <NFD pattern> <NFD file>  → 1
+len('아닌') NFC = 2 code points, NFD = 5     # jamo L+V+T decomposition
+```
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| The pattern must survive both normalization forms | Match on a substring that contains no combining sequence (an ASCII token, an id, a number), or normalize the input through a filter before grep |
+| A pattern with a character class or quantifier over non-ASCII text | Test the exact pattern against a known-matching line first: BSD and GNU regex engines differ in multi-byte class handling, so a class that works on one userland can misfire on the other ([platforms-tools-bsd-vs-gnu-cli]) |
+| Zero hits and the cause is unclear | Print the code points of both the pattern and the target line (`python3 -c "print([hex(ord(c)) for c in open(f).read()])"`) before concluding the text is missing — it separates "word absent" from "different code points" |
+| The text is user-supplied and used as a key or a dedup identifier | Normalize to NFC at the trust boundary on write, so later equality and search compare one form |
+| The search happens inside a database rather than a file | Normalization is applied by the writer, not the engine — same rule: normalize on write, search the stored form |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Write a stem prefix (`아니`) expecting it to match its inflections | Copy the literal form present in the text, or enumerate the alternatives | Precomposed syllables are distinct code points, so the stem is not a substring of the inflected word |
+| Read a 0-hit grep as "the requirement is absent from the document" | Compare the code points of the pattern and the line before acting | 0 hits also means "different normalization form" or "different syllable" — an absence conclusion from that is a false negative |
+| Compare two file-name lists byte-for-byte across machines | Normalize both lists to NFC in code, then diff | Producers store different forms of the same name; APFS lookup hides this locally but a byte diff does not |
+
+## Sources
+
+- https://unicode.org/reports/tr15/ — normalization forms; Hangul syllables have special full-decomposition rules; canonically equivalent strings have different binary representations unless normalized
+- https://www.unicode.org/versions/Unicode16.0.0/core-spec/chapter-3/ — §3.12 Conjoining Jamo Behavior: 11,172 precomposed Hangul syllables from `SBase = U+AC00` decompose algorithmically into L/V/T jamo
+- https://developer.apple.com/library/archive/documentation/FileManagement/Conceptual/APFS_Guide/FAQ/FAQ.html — APFS preserves the file name's normalization and is normalization-insensitive via hashes of the normalized form; HFS+ stores the normalized form
+- https://pubs.opengroup.org/onlinepubs/9699919799/utilities/grep.html — grep matches patterns against input lines by the specified regular-expression rules; no canonical-equivalence folding is specified
diff --git a/wiki/platforms/filesystems/paths-case-and-line-endings.md b/wiki/platforms/filesystems/paths-case-and-line-endings.md
index e9a209e..0cf95c1 100644
--- a/wiki/platforms/filesystems/paths-case-and-line-endings.md
+++ b/wiki/platforms/filesystems/paths-case-and-line-endings.md
@@ -9,7 +9,7 @@ sources:
   - https://learn.microsoft.com/en-us/windows/win32/fileio/naming-a-file
   - https://git-scm.com/docs/gitattributes
 last_verified: 2026-07-10
-related: [platforms-shells-portable-shell-scripts]
+related: [platforms-shells-portable-shell-scripts, platforms-environment-unicode-text-matching]
 ---
 
 # Files That Break When a Repo Moves Between macOS, Windows, and Linux
diff --git a/wiki/platforms/index.md b/wiki/platforms/index.md
index dccb400..35abf3e 100644
--- a/wiki/platforms/index.md
+++ b/wiki/platforms/index.md
@@ -3,8 +3,9 @@
 Route here for: OS-level differences that break code and scripts moving between
 macOS, Linux, and Windows — shell portability, BSD-vs-GNU CLI flags, filesystem
 case/line-ending/path behavior, file permissions and exec bits across
-git/archives/containers, hidden environment inputs (timezone/locale, per-context
-PATH resolution), keeping processes alive as services or scheduled jobs, and
+git/archives/containers, hidden environment inputs (timezone/locale, Unicode
+normalization form in text and file names, per-context PATH resolution), keeping
+processes alive as services or scheduled jobs, and
 pinning toolchain versions across machines. Application logic stays in backend;
 SQL stays in databases.
 
@@ -27,6 +28,7 @@ Match your situation to a "load when" line; load only matching pages.
 | Page | Load when |
 |------|-----------|
 | [timezone-and-locale](environment/timezone-and-locale.md) | Date/time or text-processing code behaves differently across machines (passes locally, fails in CI or vice versa); a cron/scheduled job fires at the wrong hour or double-fires/skips around DST; reviewing code that formats, parses, or compares dates or strings; writing tests that touch time; building case-insensitive keys, sorted output, or number parsing that must agree across machines |
+| [unicode-text-matching](environment/unicode-text-matching.md) | A grep/regex pattern over non-ASCII text (Korean/Japanese/accented Latin/emoji) returns zero hits on text you can see; writing a pattern that must match an inflected or precomposed word; a search or name comparison works on one machine and misses after the file crossed an OS/archive/editor boundary; deciding where to normalize (NFC/NFD) user-supplied text used as a key |
 | [path-resolution](environment/path-resolution.md) | "command not found" though the tool is installed; a different version runs than the one installed; sudo/CI/cron/GUI apps/ssh can't find a command the interactive shell finds; two installations of the same tool conflict; deciding how a script should locate its correctness-critical tools |
 
 ## filesystems
diff --git a/wiki/qa/document-verification/spec-document-gates.md b/wiki/qa/document-verification/spec-document-gates.md
new file mode 100644
index 0000000..c2e3959
--- /dev/null
+++ b/wiki/qa/document-verification/spec-document-gates.md
@@ -0,0 +1,95 @@
+---
+id: qa-document-verification-spec-document-gates
+domain: qa
+category: document-verification
+applies_to: [general]
+confidence: field-tested
+sources:
+  - https://eslint.org/docs/latest/extend/custom-rule-tutorial
+  - https://eslint.org/docs/latest/integrate/nodejs-api
+  - https://testing.googleblog.com/2021/04/mutation-testing.html
+  - https://www.rfc-editor.org/rfc/rfc2119
+  - https://docs.vale.sh/checks/conditional
+  - https://docs.vale.sh/checks/occurrence
+  - https://github.com/DavidAnson/markdownlint/blob/main/doc/md056.md
+  - https://github.com/DavidAnson/markdownlint/issues/1206
+last_verified: 2026-07-30
+related: [qa-process-acceptance-criteria, testing-quality-tests-that-cannot-fail, platforms-environment-unicode-text-matching]
+---
+
+# Automated Gates on a Specification Document
+
+## When this applies
+
+You are writing or reviewing automated checks (grep/script) that decide whether a
+written deliverable — RFC, API spec, schema doc, design doc — satisfies its stated
+requirements; a document passed its checklist but a reviewer still found the
+requirement unmet; you are fixing gate patterns for a document that is not written yet.
+
+## Do this
+
+1. **Prove the gate on a conforming sample before adopting it.** Run each pattern
+   against a sibling document that already satisfies the same spec and require the
+   exact expected count (`7` sections, `2` signature mentions). A run against the
+   not-yet-written target only shows the file is absent — a typo'd or mis-anchored
+   pattern produces the identical result. ESLint's `RuleTester` encodes this pairing:
+   it "requires that at least one valid and one invalid test scenario be present."
+2. **Prove the gate on a mutated copy.** Plant the exact defect the gate claims to
+   catch, require FAIL, and keep the mutant as a permanent self-test — this is
+   mutation testing applied to the checker instead of the code.
+3. **Check on four axes.** Token existence alone passes documents that violate the spec:
+
+| Axis | What the check does | Defect that a token-existence check misses |
+|------|---------------------|--------------------------------------------|
+| Structure | Parse the table and assert rows, columns, and non-empty cells | The whole table is deleted while the token survives in a nearby paragraph |
+| Modality and polarity | Within one sentence scope, assert the requirement is neither negated nor demoted (MUST→SHOULD, 필수→권장/원칙적으로) | "X is not required" and "X is recommended" both contain every keyword |
+| Set completeness | Assert the exact member count of a closed set (`enum has exactly 5 rows`) | One enum row is deleted; the token count is still ≥ 1 |
+| Cross-reference | Assert that a statement in one section implies its counterpart elsewhere, and recompute a derived value from its inputs | Two sections disagree, or an `Examples` block silently stands in for the deleted normative rule |
+
+4. **Fail closed when the anchor is missing.** When the section heading, table, or
+   derivation input a check needs cannot be located, report FAIL. A check that
+   reports "nothing to check" disappears the moment someone rewords the sentence it
+   keyed on.
+5. **Write patterns against the literal text in the document**, not a remembered
+   stem — for non-ASCII text apply [platforms-environment-unicode-text-matching].
+
+## Edge cases
+
+| Case | Then |
+|------|------|
+| No conforming sibling exists (first document of its kind) | Author a minimal conforming fixture, run the gate against it, require PASS, and keep the fixture next to the gate |
+| The gate must exist before the document (plan-first workflow) | Take the positive control from the sibling or fixture; treat the target's failing run as evidence of absence only |
+| The `Examples` section satisfies the check while the normative section does not | Scope the check to the normative section (heading range), so examples cannot stand in for the rule |
+| The check counts delimiters (`\|`) as a stand-in for parsing | Use a Markdown parser — markdownlint's own MD056 misreports when a pipe appears inside backticks (issue #1206) |
+| The document deliberately relaxes a requirement | Change the gate and the acceptance criteria in the same commit, and record the relaxation in the PR ([qa-process-acceptance-criteria]) |
+| The deliverable is code, not a document | Apply [testing-quality-tests-that-cannot-fail] — same red-run proof, expressed as tests |
+
+## Instead of
+
+| If you are about to | Do this instead | Why |
+|---------------------|-----------------|-----|
+| Adopt a pattern because it exits non-zero against the unwritten target | Run it against a conforming sibling and require the expected count | An absent file fails every pattern; the red run cannot distinguish a correct pattern from a mistyped one |
+| Treat "the keyword is present" as the requirement being met | Add the sentence-scoped polarity and modality check | RFC 2119 makes MUST and SHOULD different requirement levels, so a demotion keeps every keyword while dropping the obligation |
+| Assert a token appears at least once for a closed enum | Parse the table and assert the exact row count | Deleting one member leaves the count ≥ 1, so the gate stays green on an incomplete set |
+| Skip a check whose anchor sentence was not found | Report FAIL and name the missing anchor | A skipped check is indistinguishable from a passed one in the summary line |
+| Verify a cross-section value by matching the number as written | Recompute it from its inputs and compare | Matching the written number passes when both sections were edited to the same wrong value |
+
+## Sources
+
+- https://eslint.org/docs/latest/extend/custom-rule-tutorial — "RuleTester requires that at least one valid and one invalid test scenario be present" — a checker is proved by code that must pass plus code that must fail
+- https://eslint.org/docs/latest/integrate/nodejs-api — `RuleTester#run()` takes `valid` and `invalid` case arrays; invalid cases assert the expected error count
+- https://testing.googleblog.com/2021/04/mutation-testing.html — inserting faults and requiring failure measures whether checks detect defects; coverage alone does not
+- https://www.rfc-editor.org/rfc/rfc2119 — MUST is "an absolute requirement"; SHOULD permits ignoring the item with justification — distinct requirement levels behind the same vocabulary
+- https://docs.vale.sh/checks/conditional — "Ensures that the existence of 'first' implies the existence of 'second'" — the cross-reference axis as a first-class check
+- https://docs.vale.sh/checks/occurrence — enforces the minimum/maximum number of times a token appears "in a given scope" — count and scope, not bare existence
+- https://github.com/DavidAnson/markdownlint/blob/main/doc/md056.md — MD056 flags tables whose rows disagree with the header's column count (structural table checking)
+- https://github.com/DavidAnson/markdownlint/issues/1206 — MD056 counts pipes inside backticks as separators: a delimiter count is not a parse
+
+## Field context
+
+Distilled from RFC-authoring sessions in this repo (2026-07): an independent
+auditor built 8 tampered copies of an RFC that a 16-gate token-existence
+checklist accepted — 5 passed all 16 gates, including a copy whose 5-row `type`
+enum had a row removed. After the four axes above plus fail-closed anchors were
+added, all 32 tampered copies failed their designated check while the intact
+document passed 62/62.
diff --git a/wiki/qa/index.md b/wiki/qa/index.md
index 92b7105..f1a1382 100644
--- a/wiki/qa/index.md
+++ b/wiki/qa/index.md
@@ -2,7 +2,8 @@
 
 Route here for: release-quality process — acceptance criteria, gates,
 regression scoping, test-environment parity, post-release verification, bug
-reports, manual/exploratory testing. Writing automated test code → wiki/testing/;
+reports, manual/exploratory testing, and automated verification of document
+deliverables (specs/RFCs). Writing automated test code → wiki/testing/;
 rollout/canary/rollback mechanics → wiki/infrastructure/.
 
 Match your situation to a "load when" line; load only matching pages.
@@ -17,6 +18,12 @@ 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 |
+|------|-----------|
+| [spec-document-gates](document-verification/spec-document-gates.md) | Writing or reviewing automated checks (grep/script) that decide whether a spec/RFC/schema document meets its requirements; a document passed its checklist but the requirement is still unmet; choosing what a doc gate must assert beyond keyword presence (table structure, MUST-vs-SHOULD demotion, closed-set completeness, cross-section consistency); validating a gate pattern for a document that does not exist yet |
+
 ## environments
 
 | Page | Load when |
diff --git a/wiki/qa/process/acceptance-criteria.md b/wiki/qa/process/acceptance-criteria.md
index c1704c5..d72562b 100644
--- a/wiki/qa/process/acceptance-criteria.md
+++ b/wiki/qa/process/acceptance-criteria.md
@@ -10,7 +10,7 @@ sources:
   - https://www.agilealliance.org/glossary/acceptance/
   - https://www.agilealliance.org/glossary/definition-of-ready/
 last_verified: 2026-07-10
-related: [qa-process-regression-scope, qa-process-release-gates]
+related: [qa-process-regression-scope, qa-process-release-gates, qa-document-verification-spec-document-gates]
 ---
 
 # Writing Acceptance Criteria That Settle "Done" Before Development
diff --git a/wiki/testing/quality/tests-that-cannot-fail.md b/wiki/testing/quality/tests-that-cannot-fail.md
index f55c155..dbe18ff 100644
--- a/wiki/testing/quality/tests-that-cannot-fail.md
+++ b/wiki/testing/quality/tests-that-cannot-fail.md
@@ -11,7 +11,7 @@ sources:
   - https://martinfowler.com/bliki/TestCoverage.html
   - https://testing.googleblog.com/2013/05/testing-on-toilet-dont-overuse-mocks.html
 last_verified: 2026-07-10
-related: [testing-quality-minimum-case-set, testing-quality-behavior-not-implementation, testing-mocking-what-to-mock, testing-async-async-testing]
+related: [testing-quality-minimum-case-set, testing-quality-behavior-not-implementation, testing-mocking-what-to-mock, testing-async-async-testing, qa-document-verification-spec-document-gates]
 ---
 
 # Proving a Test Can Fail