Postgres' WAL is analogous to SQL Server's transaction log, which is also truncated only on a checkpoint. Both PG's WAL and SQL Server's transaction log are only truncated when all those pages are written to the data files and nothing else depends upon those log files. For more info see the docs.
This allows PostgreSQL to fully restore the page even if the data file contains a torn or partially written page caused by a power loss or crash.
Only if the WAL hasn't been checkpointed yet. It's not a backup which can restore arbitrary pages, it's just a way to harden the transaction to disk quickly without writing changes to all different parts of the data files. That happens long before the page is actually written to the file, which only happens during the checkpointing process.
How does SQL Server handle torn or partially written pages?
Is torn page detection or recovery handled differently in SQL Server compared to PostgreSQL?
Similarly to Postgres, it only truncates once the full checkpoint has happened. If it crashed then it will just try to checkpoint again on restart.
Does SQL Server store full page images in the transaction log?
Not always, for small modifications it only stored the changed rows. You can see more in this article.
I know SQL Server has page checksums and torn-page detection, but I want to understand whether the storage engine logs enough information (like PostgreSQL does with full-page writes) to fully reconstruct a damaged page during crash recovery.
Checksums only tell you whether a page isn't written correctly, and only when it's read. All that will happen is an error is thrown. DBCC CHECKDB (which also does more thorough checks) can recover data from the transaction log if it's there.
Torn page detection (which covers the case of a chcekpoint going bad) will detect and redo any partial page update. So if only part of a page got updated then the checkpoint redo will only write over the rest of the page. So Postgres' full_page_writes logic isn't necessary.
