Package test projects for tbRegExp and tbFileSysTools, callout labels, and both packages moved to their newest versions

.github/actions/run-gates/action.yml

@@ -135,8 +135,9 @@ runs:
       run: node --test test/example-batches.test.mjs
     # Unit tests for scripts/lib/tb-ports.mjs, which claims the DevTools ports
     # the harness's IDEs use: two claims at once are disjoint, a port in use
-    # and a live lock are passed over, a stale lock is taken over. Starts no
-    # IDE. No browser, no built tree, about 2 s.
+    # and a live lock are passed over, a stale lock is taken over; and for what
+    # ending an IDE removes of its temp folder and profile. Starts no IDE. No
+    # browser, no built tree, about 2 s.
     - name: Unit-test the harness's port claims (test/ports.test.mjs)
       shell: bash
       run: node --test test/ports.test.mjs
@@ -215,6 +216,14 @@ runs:
     - name: Verify the book-coverage warnings (check_book_coverage.mjs)
       shell: bash
       run: node scripts/check_book_coverage.mjs
+    # Every NOTE and WARNING of a contributed package's section names the test
+    # cases behind it, and each package test project under test/package-tests/
+    # is in order: its cases named, well formed, in expect.json for the
+    # documented version and in its generated RunAll module. Sources only: no
+    # browser, no built tree, no install.
+    - name: Verify the package test markers and projects (check_package_tests.mjs)
+      shell: bash
+      run: node scripts/check_package_tests.mjs
     # The symbol index (tB/symbols.json) is read by the IDE help add-in. A
     # build that indexes the reference cleanly says nothing about the rules
     # that did not fire on it, so these probes assert each against the case

WIP.A11y.md

@@ -98,7 +98,7 @@ It has one blind spot: it compares a candidate against a baseline produced by th
 
 It also compares *which* findings axe produces, never their shape. The `no-html` scheme passes the gate on every audit and would still crash the reporter, because `noHtml: true` makes `node.html` null and `check_a11y.mjs` calls `.slice()` on it. Treat the gate as necessary, not sufficient.
 
-Cost attribution for the scan lives in [perf/ab-axe.mjs](perf/ab-axe.mjs), with [perf/probe-axe-dom.mjs](perf/probe-axe-dom.mjs) for DOM operation counts and [perf/probe-axe-scaling.mjs](perf/probe-axe-scaling.mjs) for the size curve. The investigation, [builder/PLAN-axe-perf.md](builder/PLAN-axe-perf.md), is complete. Its conclusion: one vendored source patch, `SOURCE_PATCHES['plain-color-fields']` in `axe-scan.mjs` (see above), is worth **26 %** across a realistic page set and **30 %** on large pages. It is gated audit-by-audit by [check_a11y_fingerprint.mjs](scripts/check_a11y_fingerprint.mjs) and verified value-by-value by [check_axe_patch_equiv.mjs](scripts/check_axe_patch_equiv.mjs) -- both are needed, because the fingerprint gate compares `incomplete` as a rule-id set and would not notice a colour error that shifted ratios without flipping a pass/fail. The patch asserts an exact occurrence count at each substitution point, so an axe-core bump fails loudly rather than silently reverting. Before touching the scan:
+Cost attribution for the scan lives in [perf/ab-axe.mjs](perf/ab-axe.mjs), with [perf/probe-axe-dom.mjs](perf/probe-axe-dom.mjs) for DOM operation counts and [perf/probe-axe-scaling.mjs](perf/probe-axe-scaling.mjs) for the size curve. The investigation, [builder/PLAN-axe-perf.md](builder/PLAN-axe-perf.md), concludes that one vendored source patch, `SOURCE_PATCHES['plain-color-fields']` in `axe-scan.mjs` (see above), is worth **26 %** across a realistic page set and **30 %** on large pages. It is gated audit-by-audit by [check_a11y_fingerprint.mjs](scripts/check_a11y_fingerprint.mjs) and verified value-by-value by [check_axe_patch_equiv.mjs](scripts/check_axe_patch_equiv.mjs) -- both are needed, because the fingerprint gate compares `incomplete` as a rule-id set and would not notice a colour error that shifted ratios without flipping a pass/fail. The patch asserts an exact occurrence count at each substitution point, so an axe-core bump fails loudly rather than silently reverting. Before touching the scan:
 
 - **Audit cost is super-linear in page size** -- `k = 2.73` on real pages, almost entirely inside `color-contrast`. The four largest pages in the site cost ~5.9x an average sample page each.
 - **Every axe-core upgrade needs the fingerprint gate run across it** -- axe ships new and revised WCAG rules between minors, and the source patches are pinned to the bundle's current text.

WIP.Build.md

@@ -112,9 +112,9 @@ Three `.py` files stay, and none is an oversight:
   account of the blocked JavaScript port.
 - **`scripts/subset_font.py`** cuts each font an SVG screenshot embeds down to the
   characters it draws (`shoot_docs --svg`, [WIP.Screenshots.md](WIP.Screenshots.md)). It is
-  fontTools, the subsetter `build_fonts.py` already relies on. harfbuzzjs was tried first and
-  removed (owner: "don't mess with harfbuzz"); it is the library whose CFF2 defect keeps
-  `build_fonts.py` in Python. It loads a font without recalculating glyph bounds, because the
+  fontTools, the subsetter `build_fonts.py` already relies on. It does not use harfbuzzjs
+  (owner: "don't mess with harfbuzz"), the library whose CFF2 defect keeps `build_fonts.py` in
+  Python. It loads a font without recalculating glyph bounds, because the
   IDE's icon font stores loose glyph boxes and recalculating them moved icons a pixel, and it
   keeps the hinting (owner's decision) and the font's timestamp, so a picture's SVG is the same
   bytes from one run to the next.
@@ -141,7 +141,7 @@ a change to `src/` is packed again by hand. See [Tools and
 Scripts](docs/Documentation/Tools-IDE.md#imagestrip).
 
 **Its own unit tests are twinBASIC `[TestFixture]` modules in `src/Sources/Tests/`**, with
-the images they build in `TestImages`, run by `tbrun --tests` (owner, 2026-10-10), which reads
+the images they build in `TestImages`, run by `tbrun --tests` (owner), which reads
 every `.twin` under `Sources/`. `TestImages.Cat` copies each `ParamArray` element through a
 `Variant` before assigning it to an array, around a twinBASIC defect queued in
 BUGS-TO-REPORT.md.
@@ -153,7 +153,7 @@ build without one rule) fails it. The two are kept in step by hand: a change to
 strips is a change to both. `compactImage` in the same module is what the SVG screenshots
 (`scripts/svgshot/snapshot-svg.mjs`, [Pictures as SVG](#pictures-as-svg)) embed their images
 through, so a change to what it strips reaches their SVGs on the next `shoot_docs --svg` run.
-What both strip was decided by the owner, 2026-10-10:
+What both strip was decided by the owner:
 
 - **An embedded image keeps its pixels and any notice of whose it is, and nothing else.**
   The IDE's icons carry GIMP's EXIF, IPTC and XMP profiles (12 KB around a 28-pixel icon),
@@ -413,6 +413,15 @@ admonition's lines by their `>` markers and strips them, and a masked fence
 inside an admonition takes its markers with it into the stash. It asks
 `blockRegions`, with the same parser, which lines are code instead.
 
+It also reads the label, the rest of the `[!TYPE]` line (`lib/callout-labels.mjs`),
+and throws on one it cannot parse: it used to drop that text silently, which is
+how a typo in a label would vanish. A package version in a label is checked against
+`page.packageVersion`, the section's `indexed_from`, which the main thread stamps
+on every page before the render chunks go to the workers, since a chunk need not
+hold the section's landing page. The code-region gate calls the chain with no
+`packageVersion` at all, which skips that check rather than failing every
+labelled page: "" means a page in no section.
+
 ### Whitespace inside inline code is content
 
 A documented value can be a padded string: [Partition](docs/Reference/Default/VBA/Interaction/Partition.md)
@@ -723,8 +732,8 @@ build.
 CI workflows: over a hundred probes on fixtures, no tree, no install. The scanner's probes are
 the traps the packages' `.twin` sources contain, the community packages' included, and what
 a declaration's record says (signature, attributes with their arguments, description, return
-type, accessor, the project name). A block of them holds the forms an audit of `parseTwin`
-found it losing without a word (lists after `Inherits` and `Implements`, a qualified CoClass
+type, accessor, the project name). A block of them holds the forms `parseTwin` could lose
+without a word (lists after `Inherits` and `Implements`, a qualified CoClass
 interface, `[Hidden(False)]`, modifier-named and initialised fields, brackets in a Type, a
 result type before `Overrides`, an unmodified `Declare`, and the `#If` branches as
 alternatives), each with the compiler's answer in its comment, checked with `tbrun --tests`
@@ -1148,7 +1157,7 @@ SVGs: online and book do, offline (and so the help archive) does not. Why:
   and the index audit agree without knowing about pictures. That adds a few hundred literal
   patterns, so `offlineExcluded` (and its copy in `check-tree.mjs`) compares a pattern with no
   wildcard as a string instead of compiling it. The archive is a stored zip, 37 MB without
-  the SVGs, and the 310 of them (2026-10-10) would add 18 MB.
+  the SVGs, and the 310 of them would add 18 MB.
 - **A picture's SVG is never inlined.** `figurePlugin` inlines a diagram's SVG in its figure; a
   picture's (`isPictureSvg`: an `.svg` with its `.png` beside it) stays an `<img>` in the
   figure's container, and `tbdocs.mjs` does not put it in the `svgContentsMap` the workers

WIP.ExamplesBuild.md

@@ -96,7 +96,7 @@ const lang = tok.info ? tok.info.trim().split(/\s+/)[0] : "";
 ```
 
 The first whitespace-separated token is the language and the rest is discarded, so anything
-after `tb` is invisible. This was verified against the real pipeline (`createMarkdownIt`
+after `tb` is invisible. This holds in the real pipeline (`createMarkdownIt`
 plus `initHighlighter`, not a bare markdown-it; see WIP.md's [Source
 dashes](WIP.md#source-dashes)). Four properties are probes that ride along on every run:
 
@@ -158,7 +158,7 @@ judgement rather than with a silence. An unknown reason is refused the way a bad
 and `inert` beside `check_build` is refused as a contradiction.
 
 The census prints three numbers: marked, inert by reason, and **undecided**. Only the last
-is a backlog, and it is the one to drive to zero. The inert count should sit still.
+is a backlog, and it is the one to keep at zero. The inert count should sit still.
 
 **`blocked` is the one reason that is not about the fence**, and **nothing carries it
 today**. It marks correct code that a *product* defect stops compiling. It exists so that
@@ -330,9 +330,8 @@ the repository's to commit: four of the nine ship no licence file, and WinReg em
   one project, VBA_JSON's `Dictionary` class hid the `Scripting.Dictionary` that WinReg's
   samples name.
 - **A project does not see the names a package's own references declare.** WinReg embeds
-  WinDevLib, and its samples' `REGTYPES`, `BOOL` and `RegOpenKeyEx` (and, before 1.0.1.1,
-  `HKEY_CURRENT_USER`, which WinReg's own `RegKeyspace` now supplies) were unresolved
-  until `contributed-WinReg` referenced WinDevLib (at the version WinReg embeds, 8.8.0.507)
+  WinDevLib, and its samples' `REGTYPES`, `BOOL` and `RegOpenKeyEx` are unresolved unless
+  `contributed-WinReg` references WinDevLib (at the version WinReg embeds, 8.8.0.507)
   and the Scripting Runtime itself. The WinReg landing page tells the reader the same.
 - **An error inside an embedded package is the package's, and fails nothing**
   (`splitPackageFaults`). WinReg 1.0.1.1 and ArrayListLib 1.6.151.1 do not compile on BETA
@@ -412,37 +411,34 @@ form with a `TextBox` on it, and impossible to compile alone, because the design
 than the code declares `Text1`. Declaring those instances lets the compiler check what the
 sample asserts: that the member exists, that it takes those arguments, that the types line up.
 
-Against a bare template 99% of the 263 `Reference/Built-In` fences fail; with every
-package referenced 78%; with the stage set 67%.
-
 **The list is written out rather than inferred, and that is the point.** A rule over "an
 identifier ending in a digit" would also declare `Var1`, `Arg1`, `Line2`, `SQLITE3`, `VBA7`,
 `MySub1` and `IShellView2`, none of which is a control. The stage set is the tool's one
 deliberate fiction, so it stays explicit and greppable.
 
 **Stage entries are the biggest lever.** A flat tail of undeclared *names* is not a flat
-tail of *work*: 45 undeclared VB control instances (`lblName`, `mnuFileSaveAs`, `optPlain`,
-`dlgOptions`, `fraLeft`), each on one or two pages, cost one file edit and fixed 22 samples
-in `Default/VB`.
+tail of *work*: a VB control instance (`lblName`, `mnuFileSaveAs`, `optPlain`,
+`dlgOptions`, `fraLeft`) used on one or two pages costs one file edit and fixes every
+sample that names it.
 
 ## Batching, which is the whole cost question
 
 One project per fence is unaffordable. What makes it tractable is that **IDE cost is flat in
 project size**: what is paid for is startup, not compilation.
 
 **Fill the lanes, not the batches.** Filling each batch to `--batch` before opening another
-put 120, 55, 4 and 3 samples on four lanes, and a run takes as long as its biggest batch. The
-batcher sizes to `ceil(total / jobs)` instead.
+would put 120, 55, 4 and 3 samples on four lanes, and a run takes as long as its biggest
+batch. The batcher sizes to `ceil(total / jobs)` instead.
 
 ### A sample can pass on its neighbour's declarations
 
 Several samples share a generated project, and a `module`-slot sample's declarations are
 **visible to every other sample in it**. So a sample can compile because an unrelated one
 defined what it referenced. Which samples share a project depends on the selection, so the
 same page can pass one run and fail the next. `Tutorials/Testing-with-Assert.md` defines
-`PadLeft` in one fence and tests it in three others: a `--propose` survey compiled all 30 of
-its fences together and passed, while the gate (marked fences only) put the definition in a
-different project and the tests failed on `Unrecognized symbol 'PadLeft'`.
+`PadLeft` in one fence and tests it in three others: a `--propose` survey compiles all 30 of
+its fences together and passes, while the gate (marked fences only) can put the definition in a
+different project, where the tests fail on `Unrecognized symbol 'PadLeft'`.
 
 `projname=` is the answer. It is markup rather than a heuristic because the grouping is a
 fact about the page that no scan recovers:
@@ -619,8 +615,8 @@ the Form page's `True`).
   both. No template changes `project.warnings` today (the defaults leave every warning on),
   so the attribute is insurance. An error canary puts an error in every batch, so `tbbuild`
   never exits 0. An undefined-symbol canary is only the warning TB0002 under **Option
-  Explicit off**, where the name is declared implicitly, and stopped a full run on
-  `Core/Deftype.md:56`, a sample that compiles, in the `[implicit]` template. A batch that
+  Explicit off**, where the name is declared implicitly (a sample that compiles, such as
+  `Core/Deftype.md:56` in the `[implicit]` template, would fail it). A batch that
   crashes the compiler returns no rows at all, canary included; `runBatch` hands a crash to
   `isolateCrash` before it reads the canary, so that case never reaches it.
 
@@ -824,8 +820,7 @@ that fails only the second is one LLVM cannot generate code for.
 **Count diagnostics, then try the fix.** A guess from counting diagnostics is not a result.
 "`[Me]` wants a fourth slot" (27 samples, TB5025): a bare `Class`
 wrapper fixes **5**, because `Me` becoming legal does not make `Me.Caption` resolve: the
-wrapper has no `Caption` (`TB5027 [tbx_…] does not contain 'Caption'`). The largest win was
-the stage set.
+wrapper has no `Caption` (`TB5027 [tbx_…] does not contain 'Caption'`).
 
 ### The Class row, and why it needs a base
 
@@ -933,7 +928,7 @@ Language rules:
 
 ### What cannot be marked
 
-The 52 `inert` fences are not programs: syntax skeletons with
+The `inert` fences are not programs: syntax skeletons with
 `<placeholder>` names (`Interfaces-CoClasses.md` writes `Inherits base_interface`),
 property-assignment lines shown outside their `With` (`WinServicesLib/ServiceManager.md`
 documents each property as `.Name = "..."`), continuation fences that deliberately reuse the
@@ -958,6 +953,6 @@ a reader writes.
   one line each, because its purpose is "the control instances the samples assume". The
   invented *classes* go in `hidden` fences, because a class is not a control instance and a
   page is where it belongs. The line between the two is a judgement, not a rule.
-- Which pages should carry `check_run` (33 samples do)? A sample states its output in its own
+- Which pages should carry `check_run` (88 samples do)? A sample states its output in its own
   text, as a trailing comment on a `Debug.Print` line or as the comment lines under
   `' Output:`; its comparison is only as useful as the pages that state a printed value.

WIP.Fonts.md

@@ -131,12 +131,12 @@ There is a secondary design point for upstream HarfBuzz --- a subsetter that can
 CFF2 instance metrics should refuse rather than emit a font with mismatched outlines and an
 invalid `head` --- but the actionable report is harfbuzzjs's.
 
-**Not yet reported upstream (last checked 2026-09-21).** A full sweep of both trackers
+**Not yet reported upstream.** A full sweep of both trackers
 (every open issue and PR, closed-issue term searches, Discussions, and the `subset-font` /
-`subfont` / `cn-font-split` wrappers) found nothing. Two look-alikes are different bugs: #6068
+`subfont` / `cn-font-split` wrappers) finds nothing. Two look-alikes are different bugs: #6068
 (HVAR advances frozen on partial instancing; fixed) and the open #4029 (glyf `hhea` bounds
 not recomputed when the metrics map is empty --- the same shape, a different table). Both
-repos' `main` still carried the defect when checked.
+repos' `main` carried the defect when last checked.
 
 **Status.** Decided: wait for the harfbuzzjs fix rather than take the Pyodide route or
 leave the script Python for good. harfbuzzjs 1.6.2 is a pure HarfBuzz 14.5.0 bump
@@ -170,7 +170,7 @@ built by fontTools and are correct. When the port happens:
   The only thing that meets all six is fontTools itself under Pyodide (`@web-alchemy/
   fonttools`, ~15 MB, pinned to a stale 4.56.0) --- CPython in WASM, which satisfies "no
   Python" as a technicality and nothing more.
-- One workaround was proposed and **declined**: stop pinning `opsz` on Source Serif, keep
+- One workaround is **declined**: stop pinning `opsz` on Source Serif, keep
   the axis, and have `print.css` ask for `font-variation-settings: "opsz" 11`. It would
   sidestep the bug and cost ~70 KB in a PDF-only face, but it changes what the generator
   produces to suit a tool defect. Do not re-propose it; the pin stays. Switching Source