v0.8.4

• A failing report piped into another program arrives whole. The CLI
  ended a run with process.exit( ), and a write to a pipe is asynchronous
  on POSIX: whatever the pipe had not taken yet was dropped with the
  process. abap2ui5lint src --json | jq on a failing corpus handed jq
  exactly 65,536 bytes of a 175 KB document, and in a workflow the GitHub
  annotations, printed last, were the first thing lost - on exactly the runs
  whose report matters. A run now sets process.exitCode and ends when its
  output is out; the early exits (--explain, --help, --version,
  --init) wait for stdout and stderr to drain first. Exit codes unchanged.

• abapGit's metadata XML is no longer read as a view. A file named on
  the command line is collected when it starts with <, and every bulk way
  of naming files names abapGit's sidecars too - abap2ui5lint src/*, a
  pre-commit hook handing over git diff --cached --name-only. Each
  zcl_app.clas.xml / package.devc.xml was judged as a view whose root
  control is abapGit: an aggregation-in-aggregation error and exit 1 for
  every class a commit touched. A document whose root element is abapGit
  is not collected now, and checkXmlSource( ) (and so --stdin) returns
  it with nothing to check; isAbapGitXml( ) is exported from ..

• a( v = … n = … ) reconstructs like a( n = … v = … ). ABAP passes
  named parameters in any order, but the reconstructor read the value of
  v, b and t as everything after its = up to the closing paren - so
  with the name written second the value came out as
  client->_bind_edit( mv_text ) n = \value``, was dropped as unresolvable
  with a note, and the attribute (a binding, here) was missing from every
  gate's view of the class. The arguments of an attribute call are read by
  name now, with the paren- and literal-aware splitter the rules use. The
  reconstructed documents of samples-controls, abap2UI5 and the fixtures (856
  classes) are unchanged - none of them writes the other order.

• --fix no longer promises the next run work it will not find. An
  edit inside text another fix of the same pass deleted was counted as
  "deferred to the next run (overlapping)": a CRLF class with five dead
  view_model_update( ) lines reported 5 deferred - the \r of each
  deleted line - and the next --fix had nothing to do. Such an edit is moot
  and no longer counted; one overlapping REPLACED text still is.

• The --fix summary counts problems, not edits. "fixed N problem(s)"
  added up fix SPANS, and one crlf-line-ending finding carries a span per
  line: a CRLF class with nine fixable findings said "would fix 45
  problem(s)" under a dry-run list of nine. Both counts are findings now;
  applyFixes( ) returns the deferred ones as deferredFindings
  (additive).

• --watch no longer loses a save made during its first run. The
  watchers were set up after the first run, which with the render gate is
  seconds of browser launch and UI5 boot: a file saved in that window, after
  the run had collected its files, was never seen, and the loop kept
  reporting the old state until the next save. The watchers now exist
  before the first run, and a change during it is one more run after it.
  The watch/render test in test/review/watch.mjs raced the same window
  and failed on a loaded machine.

• An output file that cannot be written is exit 2, and the report still
  arrives whole. --sarif-out / --json-out into a path that cannot be
  created threw out of the run - a stack trace and exit 1, the findings
  code, on a clean run - and so did --update-baseline; a badge that could
  not be written exited through process.exit(2) after the report had been
  printed, which cut a piped report off at the pipe buffer again. Each is
  one line on stderr and exit 2 now, and the files written beside the report
  set the exit code instead of exiting.

• The render gate works under pnpm, and no longer hangs when it cannot.
  The resource server served "the folder @openui5/sap.ui.core is in", which
  in a flat npm tree holds every library and under pnpm holds sap.ui.core
  alone - so the bootstrap waited for sap.m forever and the run never
  ended. Every @openui5 package of the runtime is now resolved on its own,
  and both waits are bounded (openRenderer({ bootTimeout, renderTimeout }),
  90 s and 60 s): a UI5 that does not boot is ERR_RENDER_RUNTIME_BROKEN
  naming what the page logged, a document that does not finish rendering is
  its own render error and its page is reloaded.

• The render runtime is also found in the project the run starts in.
  npx --yes @abap2ui5/linter src in a repository with
  @abap2ui5/linter-render as a devDependency said "12 of 12 packages
  missing", and a global linter with a local runtime exited 2 under
  --render: the runtime was looked for next to the linter only. It is now
  looked for there first and in process.cwd( ) second.

• A Chromium that will not start is a missing runtime, not a crash. With
  @abap2ui5/linter-render installed and npx playwright install chromium
  skipped, the run ended in Playwright's browserType.launch: Executable doesn't exist banner, a stack trace and exit 1 - the code for findings.
  Now it is one sentence (ERR_RENDER_BROWSER_MISSING, the fix named) and
  exit 2 where the gate was asked for (--render, "render": true), and the
  property gate alone with a notice where the gate was only left on - what a
  missing runtime package already got. openRenderer also closes the HTTP
  server it had started when the launch fails; it used to keep the process
  alive.

• A render runtime of another UI5 release than the snapshot is named.
  The property gate judges against the snapshot (1.152), the render gate
  against the runtime; @abap2ui5/render-runtime 0.1-0.6 serve 1.151, and a
  1.152 member then failed view creation as "unknown setting" - a render
  error that read like a broken view. The CLI now says so on stderr before
  the run (runtimeSnapshotMismatch( ) in ./render).

• A baseline only speaks for the files a run linted. Every entry no
  finding matched was STALE, whether or not its file was part of the run:
  after --update-baseline over src (57 entries), linting one file with
  --baseline reported "56 STALE entries" and exited 1 - also under
  --advisory, whose help promises exit 0, and under --stdin. Now only an
  entry of a file the run looked at can be stale - one it linted, or one
  under a path it walked that it no longer collects (a class that stopped
  building a view cannot have a finding, so its entries are stale like a
  deleted file's) - or of a file that is gone from disk (nothing will ever
  match it again). --update-baseline on one file replaced the whole
  baseline with that file's entries; it now replaces the entries of the
  files the run looked at and keeps the others, dropping those of deleted
  files. --advisory / --fail-on never report stale entries and exit 0.
  --json carries an additive baseline block (file, suppressed,
  byRule, stale: [{ key, count }]), so a document saying failing: 0
  next to exit 1 names the reason. updateBaseline( ) and keyFile( ) are
  new in ./baseline; applyBaseline( ) and updateBaseline( ) take
  { scope }, the paths the run walked.

• --fix no longer deletes a namespace a helper still uses.
  unused-namespace-declaration judged the reconstructed view as if it were
  the whole view. abap2UI5's own z2ui5_cl_ui5_app_start declares
  xmlns:form for the SimpleForm its create_layout_form( ) helper adds - a
  RETURNING helper whose result is assigned, which the reconstructor does not
  follow - and without its waiver comments the finding came with a deleting
  fix that broke the view. The rule now stands down for a class whose
  reconstruction is known incomplete (a builder call it could not place,
  unplacedTokens on prepareAbap( ) - helperTokens without a mere second
  stringify( ) of a finished view), and for a prefix the class writes in
  more builder literals (ns = \form`, `form:X`, n = `core:require`) than its documents carry. A prefix written nowhere is still reported. A waiver of the rule where it stood down is unjudged, not unused-directive (applyDirectives( … { stoodDown })`): the rule did not judge that class,
  and app_start's own waiver would otherwise have been reported as "remove
  the directive" on every run over abap2UI5.

• New rule malformed-xml (error). A raw *.view.xml /
  *.fragment.xml with a duplicate attribute, a </contnt> for <content>
  and no closing root tag passed --no-render with "Success!" and exit 0:
  the XML reader behind the property gate is lenient on purpose, and nothing
  asked whether the document was well-formed - which the browser's parser
  decides before UI5 sees a control, refusing the whole view. parseXml( )
  now records mismatched, unclosed and stray closing tags and duplicate
  attributes as root.malformed (tags inside CDATA and comments are text),
  and checkNodes( ) reports each once, so the VS Code extension's and
  mcp-server's self-assembled pipelines get it too. The tree the other rules
  judge is built exactly as before.

• A CDATA section in a raw view is character data, to its ]]>. The XML
  reader skipped comments but read a CDATA section as markup: a <Txt> in
  one became an element (and an unknown-control), and a <!-- in one -
  script text testing for "<!--" - opened a comment that swallowed the
  real tags behind the section up to the next -->, so a view xmllint
  accepts was three malformed-xml errors and the elements after the
  section were never judged. Unchanged on the 1,390 *.view.xml /
  *.fragment.xml files of the OpenUI5 samples and the corpora.

• unused-directive no longer names a waiver for a rule that did not
  run. Under --no-properties, or with the rule switched off or excluded
  in the rules block, a directive for it suppressed nothing because nothing
  was judged - and the hint told the author to delete the waiver the full
  run needs. Such an id is now unjudged rather than unused. applyDirectives( … { ran }) takes the caller's answer (the entry points pass "not a
  property-walk rule" when the property gate is off - WALK_ONLY_RULES in
  ./properties, gated against the emit sites), and ruleRuns( ) in
  ./findings is the rules block's.

• A bare directive is not called unused by a run without the property
  gate either. Under --no-properties a directive with no id (so: every
  rule) over a line whose only finding is a property-walk one still read
  "suppressed nothing … remove the directive". It is unjudged whenever the
  run left a gate out (parseDirectives( … { gatesRan })); a rule the
  rules block switched off does not make it so, since the config speaks
  for every run of the repository.

• ignore and the render gate's exclude mean the same thing however
  the run is started. #35 made a rules[id].exclude pattern
  match every spelling of a path - the absolute one a config's paths
  produce and the relative one abap2ui5lint src does. ignore and
  rules['render-error'].exclude kept testing the path only as it was
  reached, so abap2UI5's own "ignore": ["/src/99/"] dropped the frozen
  package from abap2ui5lint and not from abap2ui5lint src, where its 17
  classes came back with 86 findings and exit 1, and a render-error waiver
  written the same way waived nothing. All three go through one matcher now
  (pathMatches( ) in ./findings).

• --cache no longer replays a render verdict of another runtime. The
  cache key held the linter version, the snapshot and the settings, but not
  the UI5 release @abap2ui5/linter-render serves - which moves on its own,
  and is exactly what the runtime-mismatch notice above tells a reader to
  upgrade. The run after npm i -D @abap2ui5/linter-render@latest replayed
  the old runtime's render errors. A rendering run's cache is now keyed by
  that release too; a property-only run's is not, and survives the upgrade.

• --format markdown keeps the tags its messages quote. A table cell
  passed a message through as raw HTML, and GitHub's sanitizer drops an
  element it does not know: in a PR comment or a job summary
  <Page> carries the attribute title twice read " carries the attribute
  title twice", close <content> with </content> read "close with ", and
  <mvc:View> became a link to mvc:View. &, < and > are escaped as
  entities now.

• --format sarif writes each location as a URI. abapGit names a
  namespaced object's file with # for its namespace slashes
  (#abc#cl_app.clas.abap), and SARIF's artifactLocation.uri is a URI
  reference, where # starts the fragment: the location named src/, and
  code scanning had no file to put the alert on. A blank in a path was no
  URI character either. Every segment is percent-encoded now.

• A --no-render run over a corpus is about a quarter faster. Every
  reader of DATA/TYPES/CONSTANTS declarations split the whole class into
  statements again - three parseData( ) and four staticAttributes( )
  splits per class, over the same scrubbed source - which was 121 MB of
  splitting for the 12 MB of samples-controls' 642 classes and a quarter of
  the run. splitStatements( ) is memoized like scrub( ) (its shared
  result frozen), and the run went from ~13 s to ~10 s with byte-identical
  findings over samples-controls and abap2UI5.

• A byte-order mark no longer breaks the config or the baseline. A file
  saved by Notepad or PowerShell's Out-File started with U+FEFF and failed
  as "Unexpected token"; it is stripped before parsing - also from the
  preview data --screenshot-model names and from the <class>.mock.json
  the screenshot reads by convention.

• --init writes a $schema the new file can reach. It was always
  ./node_modules/@abap2ui5/linter/…, wrong in a monorepo package and
  pointing at nothing under npx or a global install. It is now the relative
  path to the nearest node_modules/@abap2ui5/linter that is the running
  linter, and otherwise the published schema of exactly this version
  (https://unpkg.com/@abap2ui5/linter@<version>/…) - also where npx's own
  copy sits below the new file (--init in the home directory, or with
  npm's cache inside the project): a path into a cache is none to commit.

• --init no longer shadows an abap2ui5lint.json. Discovery reads
  abap2ui5lint.jsonc before abap2ui5lint.json, and --init only checked
  for the first: beside an existing .json it wrote a fresh .jsonc, which
  from then on silently won (a repository with "failOn": "never" went from
  exit 0 to exit 1). It refuses beside either spelling now.

• A missing extends target says so. It was reported as "no such file -
  check the --config path", also when no --config was given; the message
  now names the extends value and where it was looked for, and a directory
  is named as one.

• scripts/generate-metadata.mjs finds the sources where they are
  installed. The shipped generator looked in
  <linter>/node_modules/@openui5 only, which a consumer's tree never has (npm
  hoists above it, pnpm keeps them beside @abap2ui5/linter-render). It now
  also resolves each package through @abap2ui5/linter-render, through its
  own package and through the current directory.

• The @abap2ui5/linter-render peer range starts at its first
  published line. >=0.7.0 <0.9.0 named a 0.7.0 the registry never had -
  the name starts at 0.8.0; 0.7.0 is the last @abap2ui5/render-runtime.
  FLOOR in scripts/peer-range.mjs is 0.8.0 and the range
  >=0.8.0 <0.9.0. The legacy @abap2ui5/render-runtime range stays
  >=0.1.0 <0.8.0 on purpose: its 0.1-0.6 lines serve UI5 1.151 against the
  1.152 snapshot, but narrowing an optional peer is an ERESOLVE for every
  project still on them, so the version gap is named on stderr instead (see
  above).

• The docs no longer send npx to an unregistered name. The issue
  templates asked for npx abap2ui5lint --version; abap2ui5lint is the
  command the package installs, not a package on npm, so outside a project
  that installed it npx would download whatever someone publishes under that
  name. Where an install is implied (the README's "stay green" block, the
  release smoke test, the rules page, the templates) it is now
  npx --no-install abap2ui5lint, and where none is, the scoped
  npx @abap2ui5/linter. AGENTS.md now records samples-controls' npm
  dependency as it is (^0.8.3, no longer a ^0.5.1 git dependency).