Changed the logic to add attachment to the zip in onEnd rather than
onTestEnd because attachment files can be deleted if e.g. preserveOutput
option is specified. Instead we add files once all workers have been
shut down. On a simple run with 1000 tests each adding 1Mb attachment
the overall time difference is 49s (streaming attachments) v 1m9s
(attachments added in the end).
Also:
- remove `blob-report` directory at the start;
- markdown's `report.md` next to package.json;
- use default location in playwright's workflows.
References #24451.
We store shard number in the report metadata event and then sort shard
files by shard number. This guarantees that within each project sharded
events will always go in stable order.
For linux tests without tracing blob-report-1.zip takes 19M, while
unpacked size is 228 MB. That size is counted for GitHub artifact
billing:
<img width="434" alt="image"
src="https://github.com/microsoft/playwright/assets/9798949/5bc32511-6686-4581-a348-acb6a54cd99b">
We zip individual .jsonl reports so that they still have unique names
and can be easily uploaded into the same artifacts directory without
name collisions.
Attachments with `path` are stored as `resources/sha1.ext` zip entries
and extracted under `report-dir/tmp` when merging. This way normal
fs.readFile keeps working as before even thought the file path is
different. The clients should rely on `attachment.name` instead of
`attachment.path` when deriving user visible titles in the UI. If this
turns out not to be the case we can reconsider later.
#10437
