fix: local corpus paths were unreadable on Windows (backslashes eaten)

gsub() in regex mode treats backslashes in the REPLACEMENT string as
escape sequences and silently drops them. A Windows corpus path is full of
them, so C:\Users\RUNNER\AppData\... was substituted into the view SQL
as C:UsersRUNNERAppData... and every DuckDB read failed with 'No files
found that match the pattern'.

Effect: uscogdata could not read a LOCAL corpus on Windows at all -- the
bundled fixture included, so the entire test suite failed there, and any
cog_mirror() copy was unusable. Remote https URLs were unaffected, having
no backslashes, which is part of why it stayed hidden.

The bug predates the {long_files} token; it lived in the original {url}
substitution since that code was written. Nothing ever ran on Windows
until the GitHub mirror's check matrix existed, which found it on its
first run: 14 of 15 Windows failures were this, the 15th a downstream
consequence of view registration failing.

fixed = TRUE treats pattern and replacement as literal text. The
regression test reproduces on any platform -- it is string handling, not
filesystem behaviour, so it needs no Windows runner.

.github/workflows/R-CMD-check.yaml

@@ -0,0 +1,61 @@
+# Multi-platform R CMD check, running on the GitHub mirror.
+#
+# This exists because the canonical Gitea runner is Linux-only, and this
+# package hard-depends on duckdb and httr2 -- both compiled, both with real
+# platform variance -- while having never been checked on Windows or macOS.
+# A large share of the audience is on Windows.
+#
+# Gitea reads .gitea/workflows and GitHub reads .github/workflows, so this
+# file is inert on the canonical repo and coexists with the Gitea CI that
+# remains authoritative for deploys.
+#
+# The suite needs NO credentials: tests/testthat/setup.R points USCOGDATA_URL
+# at the bundled fixture corpus. That is exactly why inst/extdata/fixture_corpus
+# must never be added to .Rbuildignore.
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+name: R-CMD-check
+
+permissions: read-all
+
+jobs:
+  R-CMD-check:
+    runs-on: ${{ matrix.config.os }}
+    name: ${{ matrix.config.os }} (${{ matrix.config.r }})
+
+    strategy:
+      fail-fast: false
+      matrix:
+        config:
+          - {os: macos-latest,   r: 'release'}
+          - {os: windows-latest, r: 'release'}
+          - {os: ubuntu-latest,  r: 'devel', http-user-agent: 'release'}
+          - {os: ubuntu-latest,  r: 'release'}
+
+    env:
+      GITHUB_PAT: ${{ secrets.GITHUB_TOKEN }}
+      R_KEEP_PKG_SOURCE: yes
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: r-lib/actions/setup-pandoc@v2
+
+      - uses: r-lib/actions/setup-r@v2
+        with:
+          r-version: ${{ matrix.config.r }}
+          http-user-agent: ${{ matrix.config.http-user-agent }}
+          use-public-rspm: true
+
+      - uses: r-lib/actions/setup-r-dependencies@v2
+        with:
+          extra-packages: any::rcmdcheck
+          needs: check
+
+      - uses: r-lib/actions/check-r-package@v2
+        with:
+          upload-snapshots: true
+          build_args: 'c("--no-manual")'

R/views.R

@@ -122,9 +122,25 @@
 #' fallback -- correct for the local temp corpora the direct-execution tests
 #' build.
 #' @noRd
+#' `fixed = TRUE` is load-bearing, not a style choice.
+#'
+#' In regex mode, `gsub()` interprets backslashes in the REPLACEMENT string as
+#' escape sequences and silently drops them. A Windows corpus path is full of
+#' them, so `C:\Users\RUNNER\AppData\...` was substituted in as
+#' `C:UsersRUNNERAppData...` and every DuckDB read failed with "No files found
+#' that match the pattern". `fixed = TRUE` treats pattern and replacement as
+#' literal text, which is what a filesystem path needs.
+#'
+#' This is why the package could not read a LOCAL corpus on Windows at all --
+#' including the test fixture, hence the entire suite, and any `cog_mirror()`
+#' copy. Remote https URLs were unaffected, having no backslashes, which is
+#' part of why it stayed hidden: the bug predates the `{long_files}` token and
+#' lived in the original `{url}` substitution, unnoticed because nothing ever
+#' ran on Windows until the mirror's check matrix existed.
+#' @noRd
 .render_view_sql <- function(sql, url, manifest = list()) {
-  sql <- gsub("\\{long_files\\}", .long_files_sql(url, manifest), sql, fixed = FALSE)
-  gsub("\\{url\\}", url, sql, fixed = FALSE)
+  sql <- gsub("{long_files}", .long_files_sql(url, manifest), sql, fixed = TRUE)
+  gsub("{url}", url, sql, fixed = TRUE)
 }
 
 #' Register DuckDB views from inst/sql/ SQL files

inst/extdata/fixture_corpus/manifest.json

@@ -1,7 +1,7 @@
 {
   "schema_version": 6,
-  "built_at": "2026-07-31T00:47:27Z",
-  "pipeline_commit": "aadb46b",
+  "built_at": "2026-08-03T16:51:32Z",
+  "pipeline_commit": "e7394a4",
   "fixture_note": "Four-year (2011, 2012, 2019, 2020) fixture for uscogdata tests. Full corpus available via USCOGDATA_URL. Regenerated from the sparsified schema-v6 corpus: the wide era (<= FY2011) no longer stores explicit zeros, so FY2011 absence means Census published $0 while FY2012+ absence means not reported. representation.parquet and code_set.parquet carry that rule and ship in full, as do every other metadata table in the publish tree. 2011/2012 straddle both the wide-aggregate -> modern-leaf format boundary (exercised by basis=\"harmonized\" and recipe= queries) and the dense -> sparse representation boundary (SB194); 2019/2020 retain the prior per-capita/CPI regression anchors. Regenerated via data-raw/regenerate_fixture_corpus.R.",
   "data_vintage": {
     "source_vintages": {
@@ -105,7 +105,7 @@
       },
       {
         "path": "data/series_breaks.parquet",
-        "sha256": "06dcc995ff533e57cc65fa25086cc9bf83ba592c58bf7cc99269dc2576f69944",
+        "sha256": "731998516cd802f63fcf7fb66053c7a62b7be955ab0794cad4a4979cb7628b87",
         "description": "series_breaks.parquet"
       },
       {

tests/testthat/test-long-files.R

@@ -62,3 +62,36 @@ test_that("registered `long` view reads through the enumerated list", {
     expect_true(all(c(2011, 2012, 2019, 2020) %in% yrs))
   })
 })
+
+test_that("a Windows-style corpus path survives token substitution", {
+  # gsub() in regex mode treats backslashes in the REPLACEMENT as escape
+  # sequences and silently drops them, so a Windows path went in as
+  # C:\Users\RUNNER\... and came out as C:UsersRUNNER..., after which every
+  # DuckDB read failed with "No files found that match the pattern".
+  #
+  # That made a LOCAL corpus unreadable on Windows -- the bundled fixture
+  # included, so the whole suite failed there -- while remote https URLs
+  # worked fine, having no backslashes. It went unnoticed for the life of the
+  # package because nothing ever ran on Windows.
+  #
+  # Reproducible on any platform: this is string handling, not a filesystem
+  # behaviour, so it does not need a Windows runner to catch.
+  win <- "C:\\Users\\RUNNER~1\\AppData\\Local\\Temp\\Rtmp123/"
+
+  out <- uscogdata:::.render_view_sql(
+    "FROM read_parquet('{url}data/summary_categories.parquet')", win
+  )
+  expect_true(grepl("C:\\Users\\RUNNER~1\\AppData", out, fixed = TRUE))
+  expect_false(grepl("C:Users", out, fixed = TRUE))
+
+  # The same must hold through the {long_files} path, which embeds the url
+  # once per enumerated partition.
+  manifest <- list(files = list(long_partitions = list(
+    list(year = 2011L, path = "data/long/year=2011/part-0.parquet")
+  )))
+  out2 <- uscogdata:::.render_view_sql(
+    "FROM read_parquet({long_files}, hive_partitioning = true)", win, manifest
+  )
+  expect_true(grepl("C:\\Users\\RUNNER~1\\AppData", out2, fixed = TRUE))
+  expect_false(grepl("C:Users", out2, fixed = TRUE))
+})