fix(utils): strip timestamp query when followed by a hash

[dfedoryshchev]

removeTimestampQuery() stops stripping the ?t=<13-digit> HMR query when the URL also has a hash fragment, so the timestamp leaks into the module graph.

utils.ts:341's terminator (?:&|$) accepts & or end-of-string but never #. This is a regression from #23364: the previous /\bt=\d{13}&?\b/ matched before a # because of the word boundary.

It is the inverse of injectQuery(), which has a dedicated hash-only branch and emits exactly file?t=<ms>#frag, so the round trip no longer round-trips.

Against a dev server with import iconUrl from './icon.svg#shape', the module-graph URL after an asset update is /icon.svg?t=1787868494725#shape instead of /icon.svg#shape. A new module-graph node is created on every HMR update of that asset, and transformRequest.ts:502's importedMod.url !== moduleUrl check never matches. Fragment URLs are a normal shape here - SVG sprites and #iefix fonts - and the playgrounds use them.

The fix only changes the terminator; \d{13} preceded by ? or & still defines a timestamp. A differential over old and new across query shapes differs on exactly one class, a 13-digit t= followed by a hash: current-t=, my-t=, 12- and 14-digit values, #t=, duplicate timestamps and ?raw&import are all untouched.

Worth noting #22343 was closed as expected behaviour for widening \d{13} to \w+; this does not widen what counts as a timestamp.

Test added to the existing removeTimestampQuery block. The utils suite goes 124 to 125 passing.

[santusht06]

Tested this locally across all query and hash permutations (?t=...#hash, ?import&t=...#hash, ?t=...&import#hash, mid-query timestamps, etc.).

The lookahead (?=#|$) correctly catches timestamps preceding hash fragments without leaking into the module graph or affecting non-13-digit parameters. All unit tests pass cleanly. LGTM!

[bluwy]

If this is causing a bug in your app, please open an issue explaining it. What is the usecase for import iconUrl from './icon.svg#shape' etc. Otherwise this seems more of a theoretical fix that we might not hit in practice.