docs: add Team Scout's testing/tooling half of the gap backlog

docs/BACKEND_GAP_BACKLOG.md

@@ -1,8 +1,9 @@
-# Backend Context Gap Backlog (M4 — Team Falcon)
+# AI-Context Gap Backlog (M4 — Teams Falcon and Scout)
 
-Backend half of next quarter's AI-context gap backlog, drafted as part of the Backend AI-Context
-Rock (epic #37124, milestone M4 / #37129). To be merged with Team Scout's testing/tooling half
-into one document, reviewed with the team, and confirmed as next quarter's Rock.
+Next quarter's AI-context gap backlog, drafted as part of the Backend AI-Context Rock (epic
+#37124). Items 1–5 are Team Falcon's backend half (M4 / #37129); items 6–10 are Team Scout's
+testing and tooling half (M4 / #37580). To be reviewed with the team and confirmed as next
+quarter's Rock.
 
 Each entry below is a **gap to be filled next quarter** — not something fixed as part of this
 Rock. Sizes are rough (S/M/L) estimates of the documentation-writing effort, not the underlying
@@ -126,10 +127,138 @@ fall cleanly under either team's current split.
   M1–M3 audited Falcon's backend docs — it may have the same kinds of staleness/fabrication
   issues found elsewhere; (2) it still needs to be linked into `CLAUDE.md`'s navigation index.
 
+## 6. `.claude/skills/` — Unowned and Unaudited — Size: L
+
+No team owns `.claude/skills/`, and no milestone in this Rock read it — yet at **14,052 lines
+across 23 first-party skills** it is the single largest AI-context surface in the repository,
+larger than `docs/backend/`, `docs/testing/` and `docs/core/` combined.
+
+- The ownership split (#37124) assigns Team Scout `.claude/commands/` — **7 files, 629 lines**. It
+  does not mention `.claude/skills/` at all.
+- The commands are thin wrappers over the skills. `.claude/commands/create-issue.md` is **5 lines**
+  and says only "use the `dot-issue-manage` skill in CREATE mode"; the skill it points at,
+  `.claude/skills/dot-issue-manage/SKILL.md`, is **630 lines**. Auditing `.claude/commands/` in M1
+  audited the pointers, not the payload.
+- Governance exists and is real — `CATALOG.md` is generated by `.claude/tools/gen-skills-catalog.mjs`,
+  `skill-lint.mjs` runs as a CI gate (`.github/workflows/cicd_pr_skill-lint.yml`), and
+  `CONTRIBUTING.md` documents the process. But the lint validates **naming, frontmatter and catalog
+  freshness**. Nothing checks whether a skill's instructions match real source.
+- **A fabrication was found on the first spot-check.**
+  `.claude/skills/dot-cicd-diagnose/ISSUE_TEMPLATE.md:90` tells the reader to run
+  `./mvnw test -Dtest=ContentTypeAPIImplTest#testCreateContentType`. The class is real, but
+  `testCreateContentType` does not exist in it — and the class lives in `dotcms-integration`, so it
+  needs `verify ... -Dit.test=`, not `test -Dtest=`. That is the **same `-Dtest=` vs `-Dit.test=`
+  error** found in `INTEGRATION_TESTS.md` during M3 (#37128), alive in a surface no milestone
+  covered.
+- Skills already carry an `Owner` field in frontmatter (`@dotcms/scout`, `@dotcms/falcon`,
+  `@dotcms/platform`, `@dotcms/maintenance`, individuals). Per-skill ownership exists; what is
+  missing is ownership of *auditing them against source*, and any process that keeps them current.
+
+## 7. 86 Integration Tests That Never Run in CI — Size: M
+
+`scripts/validate-integration-test-registration.sh` (added in #37590) measures it: **632 classes
+registered** in a CI suite against **693 concrete `*Test.java`** in `dotcms-integration` — **86 that
+exist, compile, and are never executed by any CI job.**
+
+- This is standing backlog, not an active leak. The flow is clean: **0 of the 33** integration tests
+  added in the last 30 days went unregistered. The registration failure mode this Rock set out to
+  stop is not currently recurring.
+- They accumulated over more than two years — `WorkflowProcessorTest` dates to January 2024.
+- Nobody knows what they are. Each is one of three things: dead code that should be deleted,
+  redundant coverage already provided elsewhere, or **genuinely missing CI coverage on a real
+  subsystem**. The third case is a silent quality gap, because the class looks like a passing test
+  to anyone reading the source.
+- The gap is the triage, not the mechanism. Detection already exists and can run in CI with
+  `--strict`; what is missing is a pass over the 86 and a decision on each.
+
+## 8. No Content-Accuracy Check for Any AI-Context Surface — Size: M
+
+Every automated check that exists validates **structure**. Nothing validates whether a claim is
+true — which is the failure mode this entire Rock was created to address.
+
+- Surfaces found carrying fabricated content so far: `INTEGRATION_TESTS.md` and
+  `REST_API_PATTERNS.md` (#37128), `BACKEND_UNIT_TESTS.md` (#37666), `CLI_OVERVIEW.md` (#37664),
+  `CICD_PIPELINE.md` (#37711), `.claude/commands/gh-issue-troubleshoot.md` (#37610), five files
+  under `.cursor/rules/` (#37590 and #37629 merged; #37715 open), and
+  `.claude/skills/dot-cicd-diagnose` (item 6, still unfixed). **Every one found by reading. None by any tool.**
+- The shape is consistent: **real API names arranged around a fictional subject.** A reviewer
+  spot-checking `WorkflowAPI` gets a hit and moves on, while `WorkflowManager` in the same snippet
+  does not exist. `PushContext` is real but an interface, not the record the doc showed.
+  `HealthStateManager` is real but has `getLivenessHealth()`, not `getLivenessResponse()`.
+- Two variants are worth designing against, because a naive "does this symbol exist" check would
+  miss both. `CICD_PIPELINE.md`'s **prose was accurate** — phase model, zero-trust design, artifact
+  strategy all matched — and only the YAML blocks were invented, so a reader skims the text, trusts
+  it, and copies the part that is wrong. And `e2e-rules.mdc` (#37715) was not invented from nothing
+  but **duplicated from a real doc and then drifted**: it told readers to *always* use `data-testid`
+  and *never* CSS selectors, where its source of truth
+  (`core-web/apps/dotcms-ui-e2e/AGENTS.md`) puts `getByRole` first and permits CSS inside the Dojo
+  iframe, which is where `data-testid` does not exist. Every symbol in it resolved; the guidance was
+  still wrong, and the Cursor rule auto-loads while the doc has to be opened deliberately.
+- Existing gates do not catch any of this: `skill-lint` checks frontmatter, the reachability check
+  proposed in #37709 catches unreachable files, and CI compiles production code — but no doc
+  example is ever compiled or resolved.
+- **The failure reproduces even under active hunting.** The replacement example written for #37666,
+  in a PR whose entire purpose was removing fabricated content, stubbed
+  `healthStateManager.getLivenessResponse()` — a method that does not exist. Caught in review, not
+  by a tool.
+- Worth scoping, not assuming: compiling every fenced Java block is expensive and probably the
+  wrong first cut. Resolving each `ClassName`, `#method`, `-Dflag` and Maven profile named in an
+  AI-context file against the source tree, and failing on those with no referent, is cheap — and
+  would have caught seven of the nine above.
+
+## 9. No Upkeep Process for `.cursor/rules/` and `.claude/commands/` — Size: S
+
+Both directories drifted out of sync with `docs/` because nothing connects them. They are updated by
+hand, by whoever remembers.
+
+- `.cursor/rules/java-context.mdc` still said `Core: Java 11 syntax. CLI: Java 21 ok` months after
+  M1 (#37126) corrected the same claim in the docs — the rule sat outside Falcon's slice and was
+  left behind (fixed in #37590).
+- `.cursor/rules/README.md` indexed five of six rules, and documented one rule's glob as what it
+  *should* have been rather than what the frontmatter said (fixed in #37629).
+- `doc-updates.mdc` — the rule whose entire job is telling a reader where to update documentation —
+  had a glob of `**/*.mdc`, so it never loaded when editing a `.md` (fixed in #37629).
+- `e2e-rules.mdc` is 223 lines restating `core-web/apps/dotcms-ui-e2e/AGENTS.md` rather than
+  pointing at it — the only rule with zero `@docs` pointers — and has drifted into contradicting it
+  on locator strategy while describing a `src/config/` directory that does not exist. **#37715 is
+  open** and folds it to a ~50-line pointer; until that merges the contradiction is still what
+  Cursor loads on every spec edit.
+- The fixes are merged; the process gap is not. Nothing makes a change to `docs/testing/` prompt a
+  look at `test-context.mdc`, and the CI gate that exists for skills has no equivalent for rules or
+  commands. The reachability check in #37709 (approved, not yet merged) covers `docs/`, not these
+  two directories.
+
+## 10. The Rock's Own Ownership Split Omitted 10 Files — Size: S (decision, not a content fix)
+
+The split in #37124 covers `docs/backend/` (18 files), `docs/infrastructure/` (1), `docs/testing/` (5)
+and nine of the ten then in `docs/core/` — **43 of the 53 `.md` files under `docs/` at the time**.
+
+- Unassigned to either team: `docs/claude/` (4 files), `docs/cli/` (2), `docs/integration/` (1),
+  `docs/test-cases/` (1), `docs/README.md`, and one `docs/core/` file, since the split's own
+  arithmetic gives 4 + 5 = 9 against 10. (`docs/frontend/`, 10 files, is out of scope for both teams
+  by design.)
+- Confirmed with Jose (Sep 21): the split was built from the file list available when the epic was
+  written, and he has not touched any of the unassigned files.
+- `docs/integration/` was already caught this way during M0 and is item 5 above — the **second time
+  the same omission surfaced**, which makes it a process finding rather than a one-off.
+- Resolved since: `docs/cli/` taken by Scout, audited and linked (#37664, merged — it found the
+  Java language level stated as 21 against a pom pinned to 11), and `docs/claude/` and
+  `docs/test-cases/` likewise (#37721, merged). #37709 links every remaining orphan and rebuilds
+  `docs/README.md` as a complete index of all 54 files; it is approved but **not yet merged**.
+  Still needing an owner: the loose `docs/core/` file.
+- The durable fix is not a better manual split. It is the complete index plus an automated
+  reachability check that fails a PR leaving any doc unreachable — an index over every file makes an
+  omission visible, and the check stops one being added silently. Both are in #37709, approved and
+  awaiting merge; until it lands, nothing prevents the next omission.
+
 ## Next Steps
 
-1. Merge this with Team Scout's testing/tooling half into one backlog document.
-2. Review with the team; confirm scope and priority order for next quarter's Rock.
+1. ~~Merge this with Team Scout's testing/tooling half into one backlog document.~~ Done — items
+   6–10 added.
+2. Review with the team; confirm scope and priority order for next quarter's Rock. Two items are
+   candidates for promotion rather than plain backlog entries: **item 4** (upgrade tasks, the same
+   silent-registration failure as the MainSuite rule) and **item 8** (no content-accuracy check —
+   the gap that let every fabrication this Rock found reach `main`).
 3. `docs/integration/` (item 5): ownership model is now decided (joint, cross-team touch-base
    required) — link it into `CLAUDE.md` and audit its content against real source before or as
    part of whichever Rock picks it up.