docs: split CLAUDE.md's Cold Storage and Closed Backlog into non-imported files

.markdownlint-cli2.jsonc

@@ -10,20 +10,19 @@
 // MD029 (ordered-list-item-prefix) is the one exception, kept enabled here: it is the
 // exact rule that caught a real, if minor, semantic issue on PR #391 -- CLAUDE.md's
 // "Active Backlog" section uses numbers as STABLE IDENTIFIERS (cross-referenced from
-// Closed Backlog by number), not literal sequential list order, so a second numbered
-// item added there can silently collide with markdown's own ordered-list numbering
-// rules. See docs/agent-lessons-learned.md and AGENTS.md's "Recurring tooling" section
-// for the convention this backstops: new Active Backlog items should be bullets with
-// the number spelled out in prose, not literal "N." markers -- reserve literal ordered-
-// list syntax for the historically-grandfathered item 7.
+// docs/agent-closed-backlog.md by number), not literal sequential list order, so a
+// second numbered item added there can silently collide with markdown's own ordered-
+// list numbering rules. See docs/agent-lessons-learned.md and AGENTS.md's "Recurring
+// tooling" section for the convention this backstops: new Active Backlog items should
+// be bullets with the number spelled out in prose, not literal "N." markers.
 //
-// One PERMANENT, EXPECTED finding: item 7's own solo "7." entry. Every MD029 style
-// (tested: one, ordered, one_or_ordered, zero) requires a list -- even a one-item list --
-// to start at "1."/"0.", so there is no style option that accepts "7." as a valid lone
-// starting number. Rewriting item 7's own giant, heavily-cross-referenced entry to a
-// bullet (like item 15) was judged disproportionate risk for a lint cosmetic given its
-// size -- so this ONE finding on line ~473 is a known, accepted baseline, not a bug.
-// Treat any OTHER finding (a 2nd hit) as the thing actually worth investigating.
+// Formerly one PERMANENT, EXPECTED finding here: item 7's own solo literal "7." entry,
+// grandfathered as a known baseline because rewriting its giant, heavily-cross-
+// referenced text to a bullet was judged disproportionate for a lint cosmetic. Item 7
+// was closed and moved out of CLAUDE.md entirely (to docs/agent-closed-backlog.md) on
+// 2026-07-31, taking its literal "7." marker with it -- this check should now come back
+// fully clean with ZERO findings. Any finding here going forward is a real regression
+// worth investigating, not an expected baseline.
 //
 // Run via `tools/run_sanity_sweep.sh` (advisory only, never fails the sweep) or directly:
 //   markdownlint-cli2 CLAUDE.md

AGENTS.md

@@ -240,7 +240,7 @@ THEN stop and open/append a PR. One loop = one change set.
   - Preferred actionlint install: `curl -sSLO https://github.com/rhysd/actionlint/releases/latest/download/actionlint_linux_amd64.tar.gz && tar -xzf actionlint_linux_amd64.tar.gz actionlint && ./actionlint -oneline .`
   - If the release tarball resolves to "Not Found" due to proxy filtering, install with Go instead: `go install github.com/rhysd/actionlint/cmd/actionlint@v1.7.1` and add `/root/.local/share/mise/installs/go/1.24.3/bin` to `PATH` before running `actionlint`. On this runner `go env GOPATH` resolves to `/root/go`, so the compiled binary also lives under `/root/go/bin` -- add that directory to `PATH` if the mise shim is absent.
   - JSON: `jq -e .` over `*.json`.
-- Markdown (`CLAUDE.md` specifically): `npm install -g markdownlint-cli2`, then `markdownlint-cli2 CLAUDE.md` (or via `tools/run_sanity_sweep.sh`, which runs this automatically -- see "Recurring tooling" below). Deliberately narrow: `.markdownlint-cli2.jsonc` at the repo root disables every rule except `MD029` (ordered-list-item-prefix). This repo's docs are a long-form historical record, not GFM-polished prose -- a full default pass produces ~2700 hits on `CLAUDE.md` alone, almost all line-length (`MD013`), a rule this repo has no interest in enforcing. `MD029` is kept because it backstops a real convention: `CLAUDE.md`'s "Active Backlog" section uses item numbers as **stable cross-referenced identifiers** (e.g. "(Active Backlog item 14)"), not literal sequential list order, so a new numbered item can silently collide with markdown's own ordered-list renumbering rules if written as a literal `N.` marker. **New Active Backlog items should be bullets (`- **...`) with the number spelled out in prose, not literal ordered-list markers** -- reserve literal `N.` syntax for the historically-grandfathered item 7, which the lint config already treats as a known, permanent, accepted baseline finding (see that file's own comments for why). This check is advisory only (never fails the sweep) and scoped to `CLAUDE.md` only -- `README.md`/`docs/*.md` use their own numbering schemes (e.g. "Finding N") that were not reviewed against this convention.
+- Markdown (`CLAUDE.md` specifically): `npm install -g markdownlint-cli2`, then `markdownlint-cli2 CLAUDE.md` (or via `tools/run_sanity_sweep.sh`, which runs this automatically -- see "Recurring tooling" below). Deliberately narrow: `.markdownlint-cli2.jsonc` at the repo root disables every rule except `MD029` (ordered-list-item-prefix). This repo's docs are a long-form historical record, not GFM-polished prose -- a full default pass produces ~2700 hits on `CLAUDE.md` alone, almost all line-length (`MD013`), a rule this repo has no interest in enforcing. `MD029` is kept because it backstops a real convention: `CLAUDE.md`'s "Active Backlog" section uses item numbers as **stable cross-referenced identifiers** (e.g. "(Active Backlog item 14)"), not literal sequential list order, so a new numbered item can silently collide with markdown's own ordered-list renumbering rules if written as a literal `N.` marker. **New Active Backlog items should be bullets (`- **...`) with the number spelled out in prose, not literal ordered-list markers.** (Item 7 -- the one entry that historically used a literal `7.` marker instead, grandfathered as a known MD029 exception -- was closed and moved to `docs/agent-closed-backlog.md` on 2026-07-31, so this check should now come back fully clean with zero findings; a finding here again is a real regression, not an expected baseline.) This check is advisory only (never fails the sweep) and scoped to `CLAUDE.md` only -- `README.md`/`docs/*.md` use their own numbering schemes (e.g. "Finding N") that were not reviewed against this convention.
   - Generic paired-delimiter scan for `.bat`, `.cmd`, `.ps1`, `.py`, `.yml`, `.yaml`, `.json`:
     - `tools/check_delimiters.py` validates (), {}, [], and quotes " ' (handle escapes and ignore comments when practical), plus a growing set of targeted `.bat`/`.cmd`/`.ps1` heuristics for regressions that have actually bitten this repo -- each one has a `# derived requirement:` comment in the source explaining the exact incident it guards against. Current heuristics: `-like` PowerShell wildcard patterns embedded in batch strings with an unmatched `[`; `set VAR="value"` instead of `set "VAR=value"`; unquoted `%VAR%` in a file-system command (`del`/`if exist`/`mkdir`/`copy`/`move`/`pushd`); unescaped `|` inside a `for /f` backtick subshell; PowerShell `-or`/`-and` outside a conditional; and (added 2026-07-25, see docs/agent-lessons-learned.md's "`rem` needs a space after it" entry) a `rem` comment line whose very next character is not whitespace, which cmd.exe parses as an attempt to run a literal command rather than a comment.
     - For `.bat/.cmd`, treat `^` as escape and `REM`/`::` as comment starts; avoid over-parsing redirection symbols.