Harden Metro bundle retry against file-watcher races (#56530)

[rubennorte]

Summary:

Each Fantom test writes a unique \-<test>.js entrypoint into
.out/js-builds/ and then asks Metro to bundle it. Metro's file watcher
(metro-file-map's FallbackWatcher on Linux, debounced 100 ms) does not
always observe the new entrypoint by the time the HTTP request arrives,
especially when multiple workers are writing entrypoints concurrently.
The previous retry logic was too narrow:

• Only HTTP 404 was treated as retryable. Metro returns 404 only when the
  entry file path itself can't be resolved; an unresolved transitive
  dep (e.g. setUpDefaultReactNativeEnvironment) returns HTTP 500 with
  {type: 'UnableToResolveError'} — we'd throw immediately on that.
• Only 3 attempts with a flat 500 ms wait (~1 s total), which is not
  enough on a busy host with 8 workers writing entrypoints at once.

This results in ~30 spurious "Failed to request bundle from Metro: Unable
to resolve module ..." failures per run.

Refactor createBundle into a focused fetchBundleWithRetry helper that:

• Parses Metro's JSON error envelope ({type, message, ...} from
  formatBundlingError) once per response and uses type to decide
  whether to retry. Retries on HTTP 404, on HTTP 500 with
  UnableToResolveError or ResourceNotFoundError, and on transient
  fetch network errors. All other failures (transform errors, syntax
  errors, real config issues) throw immediately so we don't waste seconds
  on them.
• Uses exponential backoff (100 ms → 200 → 400 → 800 → 1.6 s, capped at
  2 s) with up to 10 attempts (~11 s total worst case).
• Surfaces a clean error message (parsed from the JSON envelope) when
  retries are exhausted.

Changelog: [Internal]

Reviewed By: andrewdacenko

Differential Revision: D101791796

[meta-codesync]

@rubennorte has exported this pull request. If you are a Meta employee, you can view the originating Diff in D101791796.

[meta-codesync]

This pull request has been merged in c1ffe9f.