Add tests for directory-watcher's default node_modules/dotfile ignoring
and AbortSignal-based scan interruption, plugin-progress's render
mutex/throttling/foreign-write isolation, and plugin-testops's global
attachment/error upload deduplication. Shorten the multi-line comments
added earlier in the session down to single lines per the repo's
comment style.

yoctocolors wraps the bolded count in the middle of the "Found N new test
results" message, splitting it with escape codes. That's invisible locally
(no TTY, no colors), but CI runners enable color output, so the plain
stringContaining assertion never matched there.

isIgnoredByDefault was splitting on node:path's sep, which is a backslash
on Windows. chokidar normalizes paths to forward slashes regardless of
platform, so the split never found a segment boundary there and
node_modules/dot-directories went unfiltered. Split on either slash instead.

…watch

Classify TestOps upload failures into transient (5xx/429/timeout),
recoverable ("launch is closed"), or terminal (auth/validation/payload/
conflict) kinds, and retry only the retryable ones with exponential
backoff. Wired into all three upload call sites (test results, global
attachments, global errors) — previously a failed request just logged
and gave up with no retry at all.

Add a leaky-bucket pacer with independent per-window budgets for
requests, files, and bytes sent to TestOps, enabled by default with
sane defaults (20 req/s, 1000 files/s, 1 GiB/s) since the server-side
limits it approximates are real regardless of whether anyone configures
for them. Disable via a new uploadRateLimit: false plugin option, or
tune it via an explicit uploadRateLimit config.

Replace the ad-hoc SIGINT handler in run --watch, and add SIGTERM
handling (which it didn't have), with a small shared signals helper
(notifySignals/waitForAbort/boundedTerminationSignal) — conventional
exit codes (130/143), a graceful-shutdown deadline, and the same
kill-tree-with-escalation behavior as before. The `open` command's
temp-report cleanup now handles SIGTERM too.

allure run now watches source files and reruns just the changed spec by
default when it's not running in CI (checked via the same CI-provider
detection plugin-testops already uses, plus a CI=true fallback) — running
it locally feels like starting a watch session, matching how the command
is actually used day to day. In CI it still runs once and exits with a
pass/fail code, unchanged. --watch / --no-watch still override the
default explicitly either way.

Clearer than relying on clipanion's implicit --no-watch negation.
--watch and --once together is a usage error.

…only dropping newly discovered directories

terminationOf only listened for 'exit', so a failed spawn (e.g. bad command)
emitted an unhandled 'error' event and crashed the whole watch session
instead of being reported as a failed run.

allure watch --new-only was passing ignoreInitial to every dynamically
discovered allure-results directory, including ones that appear after watch
has already started — dropping their backlog even though it's new content,
not a pre-existing one to skip.

Defaulting to watch mode outside CI would silently change behavior for
every existing "allure run" workflow that isn't run in CI, since the
command is what actually executes the test suite. Back to plain --watch,
no CI auto-detection, no --once.

…ed directory's .gitignore

Blanket-ignoring anything starting with "." could silently drop test data
users keep in a dot-prefixed directory. Now only node_modules/dist/build/
out/coverage/.git are ignored unconditionally, plus whatever plain names
the watched directory's own .gitignore lists.

Also flips allure watch's --new-only to on by default (pass --no-new-only
to ingest the existing backlog instead), since starting a watch session and
getting flooded with every already-existing result is rarely what's wanted.

remove unnecessary logs from testops plugin (change log level) in realtime
simplify testops uploader logic
fix some progress plugin issues

TestError has no stable id, but the store only ever appends to
allGlobalErrors() and never replaces existing entries, so the same
logical error keeps the same object reference across calls. A
Set<TestError> keyed by that reference dedupes the same way global
attachments already do (Set<string> by attachment.id), without
depending on array position staying meaningful.

Covers the exact scenario the at(-1)-based bug hit: several new errors
landing between updates, another update with nothing new, then one
more. Verified this fails against the old at(-1) logic (drops one of
two errors added in the same cycle) and passes with the current
Set-based dedup.