[build] Address aspect-build/rules_ts#664, drop Windows workaround

[fhanau]

Finally figured out how to solve this. We need to ensure that rules_ts enables the composite setting and avoid writing to the same path twice since Windows does not have a proper Bazel sandbox.
Also update to the latest Bazel release and enable --rewind_lost_inputs, both of these should help with the Bazel remote cache issues we've seen.

=============

This issue has hounded workerd CI and made the Windows build slower than it needs to be for more than a year, glad that we can get rid of the workaround. I did a bunch of testing on #1528 (my perpetual build systems draft PR) to ensure that this being solved is not just a mirage based on getting the right outputs from the remote cache, but also happens when caching is disabled.
As a fun side effect, we also get to re-enable a test that was disabled due to the same root cause.

[fhanau]

I'm dropping this comment on purpose: isolated_typecheck would allow us to get rid of a type check action in cases where it is not needed which is why I was interested in it, but as the comment states this would require quite a bit of work and the gain is relatively small. This comment is an AI-generated replacement for a comment I originally wrote. There is no need to describe it in this much detail and for our purposes not having isolated_typecheck is not a big loss.

[ask-bonk]

Both callers that used wd_ts_project already explicitly set composite = True, so the default change is safe. The wd_ts_test calls from types/BUILD.bazel also explicitly set composite = True.

Now let me look at the src/wpt/BUILD.bazel -- it uses wd_ts_project but does NOT pass composite. It was previously defaulting to False, and now will default to True. This is a behavioral change. Let me verify if this is safe:

The wd_ts_project at src/wpt/BUILD.bazel:15 does not pass composite, so it gets the default. Before this PR, default was False; after this PR, default is True. This might affect the WPT build. However, the PR is specifically enabling composite everywhere to fix the rules_ts issue, so this is intentional.

Let me now do a final review of all the changes as a whole. The changes look correct and well-motivated. Let me verify the types/tsconfig.json diff is purely a cleanup (composite is inherited from base):

The types/tsconfig.json removes:

1. "composite": true -- now inherited from base.tsconfig.json (no behavioral change)
2. An empty line -- cosmetic cleanup

Both are correct.

This PR looks clean. The key changes are:

1. Bazel 9.1.0 -> 9.1.1
2. --rewind_lost_inputs for CI cache resilience
3. composite = true enabled globally for TypeScript (fixing rules_ts#664)
4. Per-test ts_project in wd_test.bzl replaced with a single consolidated ts_project in BUILD.bazel
5. Windows workaround build step removed from CI
6. form-data-test-ts re-enabled

All changes are build-system-only, no runtime behavior changes, no backward compatibility concerns.

LGTM

github run

[npaun]

The WPT test config is also TS. Was the trick already used there?

[fhanau]

The WPT test config is also TS. Was the trick already used there?

Likely more as a lucky accident, but yes – it has a single ts_project, so there's no concern about files being generated twice. We still had //src/wpt:wpt-all@tsproject in the set of targets where the prior workaround was being applied – I think it might not have been needed for all of them, or it might have been needed for importing node correctly since composite needs to be set.

[npaun]

@fhanau

Putting it all into one project was intentional, but I didn't know I needed composite. Thanks for getting this sorted out.

[petebacondarwin]

Approving on behalf of wrangler for the "typings" files.