perf: make updater faster by streaming instead of reading into memory

[ottenhoff]

Fix system crash (OOM) caused by updater spending time in folders like docs/

• Skip hidden and vendor directories (node_modules, vendor, docs, etc.) to avoid OOM
• free ASTs immediately
• ts-morph dependency bumped to ^27.0.2

Closes #3483

[marvinhagemeister]

Let's store the found paths, so that we don't need to walk the file system again a second time.

[ottenhoff]

Agreed. I just pushed a new commit.

[marvinhagemeister]

What's the difference between HIDE_FILES and SKIP_DIRS? Aren't both used for the same thing? Let's unify that.

[ottenhoff]

yup, done.

[bartlomieju]

Good fix for a real OOM problem. The core idea — stream files one-by-one via walk + forget()/removeSourceFile() instead of loading everything via addSourceFilesAtPaths — is solid. A few things to address:

1. userFileCount vs totalFiles are always equal (logic bug)

Since walk is called with skip: [SKIP_DIRS], no file from a skipped directory will ever appear in filesToProcess. That means the if (!SKIP_DIRS.test(entry.path)) userFileCount++ check will always be true for every entry, so userFileCount === totalFiles always. The distinction between the two is dead code after the refactor.

Same applies to the modifiedFilesToShow filter — it's filtering against paths that can never match since walk already excluded them.

You can simplify by dropping userFileCount entirely and just using filesToProcess.length.

2. docs in SKIP_DIRS is too aggressive

```
node_modules|vendor|docs|.git|.next|.turbo|_fresh|dist|build|target|.cache
```

A user's Fresh 1.x project could easily have a docs/ directory containing actual .tsx routes or components that need migration. Unlike node_modules/vendor/dist/build which contain generated or third-party code, docs is a common name for user-authored content. I'd remove docs from this list.

3. The old code comment said "sequential" but was actually parallel

The original code had:

// process files sequentially to show proper progress
await Promise.all(sfs.map(async (sourceFile) => { ... }));

That was parallel via Promise.all, not sequential. The new for loop actually is sequential — good fix, but worth noting this was also a latent bug (progress bar was racing).

4. Minor: build and target in skip list

build is a pretty common directory name that could appear in user projects (e.g., routes/build/index.tsx). Consider whether matching it only at the project root level would be safer, or just remove it. target (Rust convention) is unlikely to contain TS/TSX files so that's fine.

5. Test change looks correct

The updated expectations for node_modules/foo/bar.ts and vendor/foo/bar.ts to retain old imports (not be transformed) is the right behavior — these files should now be skipped entirely.

Overall this is a good improvement. The main things to fix are #1 (dead code from the refactor) and #2 (docs being too aggressive).