The new empty-grid spec built a grid from `data: []` alone, and Handsontable
reports zero columns when the data is empty and no `columns` setting defines
the schema - so buildGridPdfInput() correctly returned null and the spec failed
in CI while passing against the stub harness. The real grids declare their
columns, so the fixture does now too.
Two findings from the second review.
A row taller than the printable area was drawn straight through the footer and
off the page, losing the tail of the cell unannounced - MAX_COL_CHARS bounds a
column, not a value, so one long text field wraps to more lines than a page
holds. Rows are now clamped to the tallest row that fits the page, with ' ...'
marking the cut, and the height budget accounts for both margins so an
exactly-fitting row stays on the page it started on instead of leaving the
first page with headers only.
exportGridPdf() also returned silently for a grid with no rows, although the
writer renders an empty table as a valid header-only page. The input is now
built by buildGridPdfInput(), which returns null only when there are no
columns, so a rowless grid downloads a page saying "0 rows x N columns".
Tests: the tall-row spec asserts the marker, the single page and that no text
start falls outside the printable band; the empty-grid spec asserts the input
is built rather than skipped.
Two findings from review.
First, a regression in the CSV/Excel path: getExportRange() defaults `left`
to skipLeadingCols, so comparing it against skipLeadingCols made every
no-selection export look like a whole-grid export and the plugin put the
leading housekeeping column back in. The whole-grid test now compares the left
edge against visual column 0, so a skipped column stays out.
Second, the PDF row cap was applied after collectGridTable() had already read
every cell, so a very large grid still froze before the cap took effect.
collectGridTable() now takes a maxRows limit that clamps the row loop before
any cell is read, and exportGridPdf() still reports the true total in the
subtitle.
Both are covered: a spec asserts the CSV export receives range
[0, skipLeadingCols, lastRow, lastCol], and one asserts the cap stops cell
reads rather than collecting and slicing.
CI caught this on the real grid specs: getPlugin('hiddenColumns') returns the
plugin instance even when it is switched off, and isHidden() then dereferences
a map that enablePlugin() never created, so any grid without a hiddenColumns
setting threw instead of exporting.
Ask the plugin's enabled flag first and treat a failure as "visible" - an
export is better off with a column the user could not see than with an error
dialog. The spec adds a regression case for exactly the CI failure, and its
afterEach no longer destroys an already-destroyed instance when a test creates
no grid.
Adds a PDF entry to the right-click Export submenu, next to CSV and Excel,
so the grid as displayed can leave the browser as a printable document.
Implemented directly against PDF 1.4 rather than adding a PDF library: the
output is a text table, which needs no embedded fonts, no images and no
compression. The base-14 Courier font is used because its glyphs are exactly
600/1000 em wide, so column widths are arithmetic on character counts instead
of a font-metrics table - a 1200 x 12 grid is laid out in ~60 ms.
Layout: A4 landscape, bold header band repeated on every page with "Page n of
m" footers, numeric columns right-aligned, values word-wrapped inside their
column, and long tables paginated. Wide tables are fitted by shrinking the
body font first (7pt down to 5pt, so a 20-column table keeps full headers at
6pt) rather than squeezing columns; only at 5pt are columns scaled, and they
stop at 4 characters - past that trailing columns are dropped and the first
page says how many were kept, so nothing is omitted silently.
Encoding is WinAnsi, the widest set the base-14 fonts cover without embedding;
characters outside cp1252 print as '?'. A typeset, branded report is a
different feature and belongs on the server, where SAS formats and DC metadata
are available.
The export keeps the existing "what you see" contract: current selection (when
it spans more than one cell), otherwise the whole grid, hidden columns
omitted, leading housekeeping columns skipped. The range calculation that CSV
and Excel already used is now a shared getExportRange(), so the three formats
cannot drift apart. Rows are capped at 5000 and the subtitle says so.
Tests: pdf-export.util.spec.ts covers document structure (xref offsets, page
count, footer numbering), string escaping, WinAnsi mapping and fallback,
wrapping, alignment, multi-line values, pagination, font fitting and the
page-width invariant; hot-export.util.spec.ts covers grid extraction against a
real Handsontable instance.
A MIN_TOTAL_MB typo reached the integer test under set -e and aborted with "integer expression expected"; it now prints usage and exits 2. A kernel without MemAvailable no longer reports 0 MB available in the summary or the warning, so the oversubscription warning stays meaningful.
release.yaml (all three jobs), sites.yaml and release-assets.yaml run the same cold production build that the review jobs do, so they get the same pre-install check.
The header now says what the check rules out (an undersized box) and what it does not (two jobs sharing one box), so the MemAvailable warning is not read as a gate.
An undersized runner does not fail honestly: the kernel OOM killer takes
the build, so the job reports exitcode 137 at ng build, or Karma
disconnects mid-suite after twenty minutes. Both read as a flaky test
rather than as a box that cannot hold the job, and both cost a whole CI
slot.
The check runs first in the three client jobs (Build-and-ng-test,
Build-and-test-development, lighthouse), prints total and available RAM,
vCPU count and any cgroup limit, and exits 1 when the effective size is
under 8 GB - the supported minimum. A cold production build measures
3.8 GB peak on its own, so 8 GB leaves roughly 2x headroom.
An uploaded file's preview is validated in place, so the cells that fail
a rule carry Handsontable's invalid cell meta. Discarding the file
restores the loaded data, but cell meta is keyed by row/column and
survives that restore - without a re-validation the grid keeps painting
cells that are no longer invalid.
cancelEdit re-runs the validation (hot.validateCells()) after restoring
the data source, which is what clears them; this test drives the flow and
pins it: import a file that fails validation, discard it, then assert the
row count is back, no .htInvalid cell remains and the page is read-only
again. The manual-edit path (test 25) keeps the parsed values by design.
Refs #122
The multi-load page rendered the loading stage only inside a branch guarded by
`parsedDatasets.length`, and that collection is empty for the whole read and
sheet-scan phase - so two of the four stage labels ("Reading the file",
"Searching for the data in the file") were assigned but never reached the DOM,
and a large file sat on screen with no feedback at all.
The stage is now shown next to the file name and size from the moment a file is
attached, and again in the content area while the sheets are being parsed. The
labels cover every phase the user waits through: reading the file, searching
for the data in the file, file ready to load, fetching the dataset details (one
getdata call per dataset - the longest wait), extracting the data, and
extracting the data finished. A spinner accompanies the stages that have work
running; the idle stages and the finished state show the label alone.
Two corrections fall out of that: the stage only moves to `extracted` once
every sheet has been parsed rather than when the first one finishes, and the
read's continuation only advances the stage if the read is still the phase in
progress (the sheet scan runs in parallel and can finish first).
excel-multi-load.cy.ts gains test 4, which asserts the file size, the ready
stage, the fetching stage (the getdata calls are held open so it is on screen
long enough to assert) and the handover to the "select a dataset" hint.
Closes#121
Enables noMacroIfInOpenCode and promotes it to "error". Open-code %if is a SAS
9.4M5 feature, so the 19 sites it found in nine files were hiding a version
requirement the codebase does not document (the requirements page says 9.4M3).
Each is rewritten so the construct no longer appears in open code:
- Six blocks become a small macro called unconditionally, which is where the
conditional then lives: drop_cas_tables (demodata), add_derived_cols
(getdata), build_sasdata4 (getdata test 4), drop_derived (stagedata),
build_staged (getrawdata), and one per option in the 20261003 migration.
- Four blocks in the two makedata STPs get the same treatment: check_outfmt,
check_keepme, emit_response and self_destruct. Their `call symputx` writes
are now explicit about the global symbol table, because a %let or a plain
symput inside a wrapper would be local to it.
- Two guards in stagedata and getrawdata are removed rather than wrapped: the
SQL only selects into a list that is already %let blank, and with no derived
variables it returns no rows, so the guard was doing nothing. The trailing
drop keeps a conditional, now in a macro and tested with %trim, since a
zero-row select can leave a single blank in the variable.
- validatefilter's two %let branches become one data step with call symputx,
which writes to the global table the later %mp_filterstore reads.
Verified: `npm run lint:sas` clean with the rule as an error; `sasjs compile`
succeeds; on sas.4gl.io the full suite passes 64/64, including the derived
variable tests in getdata test 4, stagedata tests 3 and 4 and validatefilter,
which are the paths these wrappers touch.
The tree is clean under all of them, so the remaining warn-level rules can be
promoted without a single code change: hasDoxygenHeader, hasMacroNameInMend,
hasMacroParentheses, noNestedMacros, strictMacroDefinition, noSpacesInFileNames,
lowerCaseFileNames, noTabIndentation, noTrailingSpaces, noGremlins and
noNestedCommentsInHeader.
With that, every rule @sasjs/lint 4.9.0 can raise is an error - only
noEncodedPasswords was an error by default - so any new violation fails
`npm run lint:sas` and the CI step that runs it, instead of printing a warning
that leaves the exit code at 0.
Checked: a trailing space appended to a file reports
"Error Line contains trailing spaces" and a non-zero exit, where it used to be
a Warning.
sas/.sasjslint takes maxLineLength from 180 to 80 - the linter's own default -
and promotes it to "error".
53 of the 258 .sas files already hold a line longer than 80. Rewrapping them
would be a large, reviewable-only-in-theory diff, so each keeps its current
longest line as an exact ceiling via a per-file override, e.g.
@sasjslint {"maxLineLength": 95} in services/validations/tables_all.sas. A
file with no override is held to 80, and an overridden file cannot grow past
the length it has today: the ceiling is the file's longest line, not a round
number.
Both directions were checked: a 101 character line added to an
un-overridden file fails with "Line exceeds maximum length by 21 characters",
and a 114 character line added to macros/mpe_execute.sas (ceiling 101) fails
by 13. `npm run lint:sas` is clean with the setting in force.
Rung 6 of the #376 ratchet: `"lineEndings": "lf"` is enabled and promoted to
"error". Its 37 findings are cleared and nothing but line endings changed.
- 34 files ended without a newline; each now ends with exactly one LF.
- 3 files were stored with CRLF (services/public/getgroups.sas and the two
sas9 lineage services, getmetacols and getmetatables), against the repo's
own `* text=auto eol=lf` in .gitattributes. They are converted to LF.
Checked mechanically that this is whitespace only: for every changed file,
HEAD's content and the new content are identical once CRLF is folded to LF
and the trailing newline is stripped.
VERIFICATION
- `npm run lint:sas`: 0 warnings, 0 errors, exit 0.
- `sasjs compile`: succeeds.
- On sas.4gl.io, against a scratch appLoc seeded with makedata:
"Tests that pass: 64/64 (100%)".
The rule flags the %if blocks service bodies run in open code, which need
SAS 9.4M5. The documented minimum is 9.4M3, so enabling it would mean
raising the floor or wrapping every service body in a macro - neither is
a lint change. Written down so the next agent does not enable it by
mistake.
Rung 5 of the #376 ratchet. Both rules are off by default; enabling them
surfaced 10 findings, all now cleared, and both are promoted to "error".
- Nine `* text;` comment statements become block comments: three disabled
debug lines (a putlog in bitemporal_dataloader, an alternative type
assignment in mpe_loader, a put in demodata) and six explanatory
comments in the two sas9 usernav services.
- loadfile.sas assigned `libname approve "&dir"` and never used the
libref; stagedata.sas and restore.sas assign the same libref because
they write approve.jsdset, loadfile writes through a fileref instead.
The statement is removed.
VERIFICATION
- `npm run lint:sas`: 0 warnings, 0 errors, exit 0.
- `sasjs compile`: succeeds.
- On sas.4gl.io, against a scratch appLoc seeded with makedata:
"Tests that pass: 64/64 (100%)".
Rung 4 of the #376 ratchet. It clears the last 5 findings and promotes the
three rules that produced them to "error": minMacroNameLength,
lowerCaseMacroNames and indentationMultiple. `sasjs lint` now reports no
warnings and no errors across the project.
- viewdata.sas wraps its service body in a helper macro named `x`; it
becomes viewdata_body. The wrapper is what keeps the %global and the
%return in one macro scope.
- usergroupsbymember.sas (sas9) builds work.iwant inside a helper macro
named `x`; it becomes build_iwant.
- refreshtablelineage.sas (sas9) defines `getTables` with uppercase; it
becomes `gettables`, at the definition, the %mend and its seven calls.
- getstagetable.test.sas: one comment line carried three leading spaces
where the file uses two.
VERIFICATION
- `npm run lint:sas`: 0 warnings, 0 errors, exit 0, "All matched files
use @sasjs/lint code style!".
- `sasjs compile`: succeeds.
- Deployed to sas.4gl.io and ran the SAS suite there against a scratch
appLoc (/Public/app/dc-lint-ratchet) seeded with makedata: "Tests that
pass: 64/64 (100%)". The viewdata and usergroupsbymember suites are in
that run, so the two renames are covered by passing tests.
Rung 3 of the #376 ratchet: `sas/.sasjslint` promotes noBannedStrings to
"error". The rule flags the literal words "error " / "warning " anywhere in
a file, comments included, because they turn up in a grep of the SAS log
and bury the real messages. Its 13 findings are cleared:
- Nine are prose and are reworded, so the sentence both reads normally and
stops matching: a truncation that aborts a build, a syntax fault, an
abort condition, "only records the failure in work.mp_abort_errds" (four
copies of the mp_abort-inside-%include comment), a test description, a
hook comment.
- Two are the text the user sees. loadfile.sas reports
"(mf_abort.error file found)" - the marker file carries the word in its
name - and buildtermsas9.sas reports "Some final error in &_program";
both are reworded to say the same thing without the literal, since the
text sits in a macro argument where the %str(WARN)ING split does not
survive the quoting.
- Three keep the word because the word is the content: the doxygen
@warning tags in macros/bitemporal_dataloader.sas and the two SAS 9
admin STPs (makedata, makelib). Each carries
@sasjslint {"noBannedStrings": false}, as
targets/server/services_server/admin/makedata.sas already did.
VERIFICATION
- `npm run lint:sas`: 5 warnings, 0 errors, exit 0.
- `sasjs compile`: succeeds.
- Deployed to sas.4gl.io and ran the SAS suite there:
`sasjs test` reports "Tests that pass: 64/64 (100%)". The run used a
scratch appLoc (/Public/app/dc-lint-ratchet) seeded with makedata, so
the shared /Public/app/dc deployment was left alone; the scratch target
was removed from sasjsconfig.json before this commit.
5 warnings remain: macro name length (2) and case (2),
indentationMultiple (1) - the last rung.
Rung 2 of the #376 ratchet: `sas/.sasjslint` promotes noUndeclaredMacros to
"error". Its 25 findings on release/7.22 are cleared, and two of them were
real faults rather than missing prose.
WHAT WAS BROKEN
Comparing the compiled program against the macros it calls (compile, then
list every project macro each generated service calls but does not define)
found three services calling macros that were absent from their own
compiled program:
- services/validations/tables_all.sas calls %mf_getuniquename() on its
third line and never declared it.
- services/validations/mpe_tables.dsn.sas calls %mf_getuniquename() and
%dc_assignlib() and declared neither.
- services/hooks/mpe_tables_postedit.sas calls %dc_assignlib(), which is
supplied by the service that includes the hook (see below).
The two validation services are configured as rule_value entries in
MPE_TABLES (mpe_makedata.sas and the v3.12/v3.13 migrations), so the calls
run when a user reviews or filters a library - each would have raised
"Apparent invocation of macro ... not resolved". Both now declare what
they call, and the definitions are present in the compiled program.
WHAT CHANGED
- 14 files declare the macros they call but had been relying on another
declaration to bring in: mf_getquotedstr, mf_nobs, mf_getuser (twice),
mf_getapploc, mf_getattrn, mp_abort (six services), mf_getuniquename
(twice), dc_assignlib, bitemporal_closeouts. These were defined in the
compiled program already; the declarations make the dependency visible
and the two validation services correct.
- 6 hook scripts get a file-local @sasjslint {"ignoredMacros": [...]}
(mp_abort, plus dc_assignlib in mpe_tables_postedit) instead of a macro
list. Hooks are included at run time by the service that owns them, so
the macros are already in that session, and these files are copied and
edited by end users - they should not carry a declaration section to
maintain. Recorded in AGENTS.md.
- bitemporal_dataloader.test.2.sas loses its test 30 block: it is gated
with `and 0` ("skip until have a db") and its only setup line calls
%assign_direct_lib, a macro that exists nowhere in this project or in
@sasjs/core. Enabling it would fail on the first line.
VERIFICATION
- `npm run lint:sas`: 18 warnings, 0 errors, exit 0.
- `sasjs compile`: succeeds; the called-but-undefined list drops from
three services to the one hook exception; the only definition-count
changes are additions.
- 18 warnings remain for the next rungs: noBannedStrings (13), macro name
length (2) and case (2), indentationMultiple (1).
sas/.sasjslint now promotes noUnusedMacros to "error", so the rule fails
the build rather than printing a warning. The 54 findings it reported are
gone:
- 46 header entries nothing calls are dropped. Each removal was checked
against the compiled artefacts rather than the linter alone: the
project was compiled before and after, and the two sasjsbuild trees
compared. 37 macro definitions disappear from the generated programs -
every one of them either a duplicate of a definition the file still
reaches through another declared macro, or a macro no generated program
calls. No macro that a generated program calls lost its definition.
- 8 entries are kept, in five files that carry
@sasjslint {"noUnusedMacros": false} with the reason written above it:
the three serviceInit programs and serviceterm.sas declare mpeinit and
mpeterm without calling them, because that declaration is how a service
body reaches those macros at all; loadfile.sas declares mp_binarycopy
for the fallback that is commented out on SAS 9.4 and later.
- Four headers are left with no macros to declare, so their empty
<h4> SAS Macros </h4> section is removed rather than kept as a stub.
@sasjs/cli is bumped 6.2.0 -> 6.5.0 in sas/package.json. The lint engine
ships inside the CLI and is not a direct dependency, so the pinned CLI is
what fixes the rule set: 6.2.0 resolves @sasjs/lint 4.3.0, 6.5.0 pins
4.9.0 exactly. Both versions were compiled against this tree and the
sasjsbuild output is byte-identical, so the bump changes the lint engine
and nothing else.
`sasjs lint` is also now reachable as `npm run lint:sas`, and a CI step in
build.yaml runs it. Until now nothing ran SAS lint automatically, and a
bare `sasjs` resolves whatever CLI is installed globally - a different
lint version reports a different rule set, which is why the counts in
issue #376 (taken with 6.2.5) do not match the pinned 6.2.0.
43 warnings remain under the rules not yet promoted: noUndeclaredMacros
(25), noBannedStrings (13), macro name length and case (4),
indentationMultiple (1). They are the next rungs of the ratchet.
The pre-commit Prettier gate could not run: it needs client/node_modules,
which this worktree does not install, and the fix it attempted touched
three client files that this change does not modify (reverted). No client
file is in this commit.
`cy.get` inside `.within(<tree node>)` only searches that node, and the
metadata accordion renders in the content area - so `.clr-accordion-header-button`
was never found however long it retried. The spec queries the accordion at
document level, selects the first metadata type by its class rather than by
node index, and clicks the header button directly (Clarity puts
`clr-accordion-header-button` on the button itself).
The assertions are on the app's own markers: the single open panel
(`clr-accordion-panel-open`, `aria-expanded`), the association tree inside the
panel, and the attribute grid that `metanav/metadetails` fills.
metanav moves from EXCLUDED to SPECS in `.gitea/scripts/cypress-specs.sh`.
Closes#395.
The case took the id from the stage route and looked for a download button
carrying it. For an upload the route does not hold a table id at all - it holds
a load reference, the CI failure naming DC20261031-201657.470_601010_4449 - so
the button it looked for could never exist, however long it waited or however
many times the screen was reloaded. That is why the row "never appeared" while
the approval was in fact there.
The row is now found the way hook-programs finds its own: by the table it names.
The id the download is named after is read from the button, which is the only
place that knows it.
Verified against the mock estate locally, with the licence the spec now plants:
7 passing, exit 0.
The estate ships no licence key, so it runs on the free tier, and the free tier
disables file upload - which is exactly what the audit case needs. Whether it
could run depended on another spec having planted a key earlier in the run: the
key persists in the estate config, so the case passed or failed on spec order.
loginAndUpdateValidKey(true) plants one here. csv-limited, which asserts the
free-tier modal, runs first in cypress-specs.sh and is unaffected; download runs
last, so nothing downstream inherits the key.
Found by running the spec against the mock estate locally: the upload input
never rendered, the modal read "Locked Feature (File Upload) ... site id (123)",
and the run went 7/7 green once the key was planted.
mpe_abortmessage existed to assemble the message mpe_execute aborts with - the
program, the state and the child message - inside a data step, so that a percent
sign, an ampersand or a quote in the child message was not macro resolved on the
way into a symbol. mpe_execute already runs a data step over the status row for
state and errmsg, so the text is built there now and the helper, plus the %let
fallback that duplicated it, is gone.
What the helper did is kept: the values come from the status row through
symget/symputx, never through %let, and a message already at the 32767 limit is
cut to the room the prefix leaves.
The comma and percent cases its tests covered now live in core's mx_execute
test, where the message is captured from a real child; what stays here is the
assembly DC owns - the prefix, the state and the child message from the status
row.
The audit-file case uploads a fixture and then looks for the approval it should
create. An upload identical to what the table already holds produces no
changes, and a changeset with no changes leaves nothing waiting for approval -
so the case finds no row. A retry lands in exactly that state: the first
attempt's rows are already in the table, so the second upload is a no-op and the
failure looks like a missing approval rather than a duplicate upload.
The fixture is now copied per attempt with its primary keys renumbered from a
random base, so the upload is always an insert. The copy is removed in
afterEach - deleteFolder cannot be used there because the source fixture lives
in the same directory.
Core 7.0.3 fills outstatus.errmsg on both platforms - the SASJS branch since
7.0.2, the Viya branch from the job result as of 7.0.3 - so a canceled call now
always carries the reason it was canceled.
That removes the last platform difference from the assertion: while Viya wrote a
blank errmsg, an expected message could only be checked on one platform and the
state alone had to stand in on the other. A comparison against a message that
was never recorded now fails instead of passing silently, which is the point of
asking for it.
The macro decided an aborted call by scanning the child's response text for a
sasjsAbort object and then recovering the message from that text - anchors for
the separator, gating on whether the payload was escaped, cuts for the log
extract, and a capture width that had to be widened more than once. Every one
of those steps existed to undo the escaping in the response.
Core now publishes the abort reason in outstatus.errmsg for both the SASJS and
Viya branches, with the whole message and no truncation, so the macro reads the
row instead: STATE says whether the child was canceled, ERRMSG is the reason.
The cut at the appended log extract is the only thing left, because that text is
part of the message the child raised; it is stripped as characters, as an
escaped pair or as a doubled pair, whichever the producer wrote.
The macro therefore takes outstatus= rather than outref=, and the call sites in
the service tests now pass the status dataset they already asked %mpe_execute
for. stagedata.test.4's two aborting calls had no status row at all and now
have one.
A platform that reports the state but no message - SAS Viya until core fills
errmsg from the job result - is judged on the state alone, so the same tests run
on both platforms.
@mpe_assert_abort.test covers the twelve cases directly against status rows
instead of building response files, including the doubled separator the estate
actually writes, a message past the old 256 character width, quotes and
backslashes.
Requires @sasjs/core 7.0.2, which carries the SASJS half of this.
The upload and submit now work; the failure moved to the approve grid, which is
loaded when the screen opens and does not yet contain the row just submitted.
The screen is reloaded up to five times, two seconds apart, until the row with
this test's id is present.
The mock getddl aborts when ds is missing, as production does (getddl.sas:43-45),
instead of quietly returning a library-level DDL - a client that dropped ds would
otherwise stay green against the mocks.
The comment above the case no longer claims a leftover row cannot matter: the
grid paginates, so a row pushed past the first page would not be found.
The upload does parse now - CI showed the info modal appearing - but the helper
treated it as an abort because the info modal and the abort modal are the same
component (.clr-abort-modal). The title is what separates them: a good upload
raises "Table Found", so that is asserted and anything else fails with its own
text instead of being clicked through.
After Submit, the stage route is already current, so waiting on the URL proved
nothing; the editor leaving preview mode is the signal that the submit was
accepted. Also trims the comment that narrated the old route.
The first case failed in CI with "Expected to find element:
input[type=file]#file-upload", because the editor renders that input inside its
upload modal - the Upload button in the button bar opens it. Clicking the button
by its label rather than by position, as excel.cy.ts does.
The viewer download cases (2-7) passed, so the rest of the spec is sound.
Follow-ups from the review of #414.
The credential guard now checks each staged .npmrc rather than only the client
one: the root and sas/ files are tracked policy as well, so a token pasted into
either was equally one `git add -A` away. Verified by staging a token in
sas/.npmrc - the hook exits 1.
$RUNNER_TEMP/npmrc is removed at the end of each job that writes it, with
`if: always()` so a failed job cleans up too.
The root `install` script is gone: the root .npmrc sets ignore-scripts, so npm
never ran it. CONTRIBUTING already spells out the per-project commands, and is
now the single source.
The release workflows build the published client from the same tree, so they run
`check:archiver` like build.yaml does.
dc
allan