CI limits concurrency to two
workers. A browser failure blocks the shared validation and publishing jobs.
Failed cases retain a Playwright trace and screenshot in the existing CI report
artifacts. The input-bounds, native-italic and native-composition suites also attach
native-selection-events.json: the last 256 focus, keyboard, input, composition
and selection events, including bounded selection text, endpoints and timing.
Endpoints include rendered field offsets so identical per-glyph text nodes can
be distinguished. Native composing flags and the pixel-rendering fallback mode
are recorded alongside them.
These observers do not alter input or selection. Open a retained trace with
npx playwright show-trace path/to/trace.zip; inspect its JSON attachment to
compare native event ordering with the failed assertion. Passing cases discard
their traces and do not attach the event history. Tests are not automatically
retried; intermittent failures still block validation.
The Windows GitHub Actions job additionally uploads windows-resource-diagnostics
for seven days, including passing runs. Its numeric JSONL snapshots contain TCP
state counts across the runner and for fixture port 4173, aggregate process
working/private bytes and handles for the test runner and fixture connection
owners, and available/committed/commit-limit/pool memory counters. The initial
sample also records configured IPv4/IPv6 dynamic TCP port ranges. It records no
addresses, individual connections, process names/arguments, test titles or
environment. Missing measurements are null with fixed error categories.
The reporter samples at suite start/end, after unexpected failures and at test
starts at most once per 15 seconds. Queries never overlap; each has an eight
second timeout and a 16 KiB output limit. The artifact is capped at 120 samples
and 128 KiB, with a final summary. Collection is read-only, enabled only on
Windows GitHub Actions, and does not change retries, assertions or test status.
Compare observations with the retained failing trace; a high TimeWait count
alone does not establish port exhaustion or explain ERR_NO_BUFFER_SPACE (#89).
Windows test Chromium temporarily disables TcpPortRandomizationWin, following
Playwright’s upstream fix
for the Windows loopback socket bug.
Playwright 1.64.1 is still the stable version. Its complete disabled-feature list
is preserved in one replacement argument; simply appending another
--disable-features argument would discard its defaults. A real CDP command-line
check compares against an unmodified launch and fails if defaults drift or any
other launch switches change. Remove the helper when a stable Playwright update
includes the upstream fix. This affects test browsers only, changes no OS setting,
and adds no assertion retries, skips or weaker image comparisons. The mitigation
does not certify every previous resource failure’s exact cause.
Cases that require an exact source caret establish it explicitly before native
typing or selection. A programmatic focus followed immediately by Control+End
can race ProseMirror’s delayed focus correction before the native selection
notification arrives. The affected overflow/placeholder/italic fixtures use
the existing rendered-range helpers; native shortcut, arrow, deletion, clipboard
and undo assertions remain unchanged. This avoids relying on an unverified
starting caret rather than retrying a failed edit or changing production input.
Composition completion also verifies that deferred document/access recovery
restores supported pixel-glyph painting. A back-to-back lifecycle case keeps the
newer preedit on live text until its own end. Native Chromium CDP cases run on
Linux and Windows; they remain automated lifecycle checks, not physical IME or
screen-reader certification.
Linux CI uses Xvfb with software OpenGL so Firefox can create a real WebGL
context. Without an X11 display its headless build correctly uses the static
fallback, which does not exercise the animated panorama renderer.
Linux GitHub Actions also retains graphics-startup-diagnostics for seven days.
This single startup snapshot records memory, the test process’s file descriptor
count and resource limits, installed Mesa/Xvfb versions and whether the expected
display/software-rendering environment is present. It does not create a GL
context or retain display addresses, process arguments or environment values.
After the native API observer tests, the same artifact also retains separate
GLX worker observations for each engine, with timestamps and unique filenames.
Writes publish complete JSON files atomically; startup removes only prior
generated GLX records. These are best effort: utility installation,
query timeout/missing-command/output-limit/format errors and file-write failures
cannot fail browser rendering checks. Only fixed capabilities or a fixed reason
are retained, never raw command output. Native WebGL and panorama assertions
remain required and unchanged; no new skips or retries are introduced.
The panorama suites observe native WebGL calls and creation/loss events. On an
unexpected failure, graphics-diagnostics.json joins the retained test trace:
at most 32 events, attempt/outcome counts, numeric capabilities, classified
renderer and fixed known failure codes/blocked-path flags. Arbitrary driver
messages are excluded. Linux CI adds a fresh host snapshot and a read-only
glxinfo -B probe with a two-second/16 KiB limit; only booleans, a renderer
category and numeric GL version are retained. Each JSON artifact is capped at
16 KiB; missing queries use null values and fixed categories. Diagnostic errors
never replace test status. There are no new retries, preferences or weakened
ready-state assertions. These observations support #90’s investigation; they
do not establish its cause. A separate failed GLX probe and a Firefox policy
refusal are distinct observations even when they occur on the same runner.
A focused Chromium job on Windows also checks regular/bold baselines, mixed
and fully bold two-line MOTDs, descenders, and the React formatting toggle.
Supported per-glyph bundled text uses locally generated pixel masks to avoid
OS antialiasing differences. The Windows job checks complete native reference
foreground/shadow masks at GUI scale 2 as well as editing behavior. Custom fonts
and text outside that renderer retain browser rasterization. Passing the Ubuntu
matrix alone does not establish correct Windows text geometry or certify an
actual Windows Java client; that capture remains outstanding in #70.
PLAYWRIGHT_CHROMIUM_EXECUTABLE overrides only the Chromium executable. It does
not change the Firefox or WebKit executable. A local Vite server is reused;
CI starts an isolated one.
Covered behavior
Every engine exercises DOM editing and React controlled echoes, focus and unmount/remount, selection formatting and undo, source validation, properties, icon upload/crop/corrupt input, alpha preservation, asset providers, local-only editing, synthetic composition and lifecycle/late-import boundaries. Menu geometry and the six existing Java text-reference comparisons run in each engine too. Panorama coverage includes native 26.3 projection probes, moving rows, pause and speed changes, CORS authorization, GPU loss and static fallbacks. Twenty rows verify that independent previews share one document’s GPU context. The Windows Chromium job also runs these panorama and example cases. The paste boundary test supplies an explicit clipboard-data property on its synthetic event. Firefox discardsClipboardEvent constructor-supplied data;
this harness difference is separate from handling a real user paste. These
checks do not exercise OS clipboard permission prompts or a real IME.