fix(#939): derive the service worker's cache prefix from one authored string

[TortoiseWolfe]

Closes #939.

The ticket's premise had already changed, so I checked before porting

#922 (PR #946) already fixed the live bug. Verified on a throwaway fixture carrying both halves of the coupling, rebranded with the fixed script:

public/sw.js                  const CACHE_VERSION = 'geolarp-v1.0.0';
                              cacheName.startsWith('geolarp-')
scripts/stamp-sw-version.mjs  const REQUIRED_PREFIX = 'geolarp-';

.mjs is inside the sweep now, so both halves rebrand together and agree. The failure #939 described cannot reproduce on a fork rebranded with the fixed script. I have corrected #939's body and revised its severity rather than leave it claiming a live leak.

Why this is still worth landing

The two strings now agree because both happen to be rewritten in the same sweep. That is incidental, not structural. One rebrand:keep, one hand edit, or one move into a file the sweep does not reach, and they part again — silently, because the damage is only visible in a browser's storage.

Deriving one from the other removes the second copy instead of re-synchronising it. CACHE_VERSION is now the only place the brand appears; CACHE_PREFIX derives from it, and both purge sites and IMAGE_CACHE derive from that.

Two things wrong independently of #922

The guard could not fail.

const cacheVersion = `${REQUIRED_PREFIX}v${version}-…`;
if (!cacheVersion.startsWith(REQUIRED_PREFIX)) {   // tautology

cacheVersion was built from REQUIRED_PREFIX. That is why the divergence shipped with a check sitting right beside it. It now refuses to stamp when it cannot parse a prefix out of the worker — a condition that can actually occur.

Nothing related what the build stamps to what the worker purges. So nothing could have caught it.

The new test never names a brand

That is deliberate: sw-cache-prefix-agrees.test.js asserts the worker derives rather than hard-codes, the stamp script reads rather than declares, and — behaviourally — that stamping a worker carrying an arbitrary brand yields a version with that same brand. The third would have caught the original defect precisely because it does not care what the brand is.

tests/unit/sw-images.test.ts required IMAGE_CACHE to be a literal starting with the brand. That check was the coupling — two hand-synchronised strings. It now asserts IMAGE_CACHE is built from CACHE_PREFIX, and that CACHE_PREFIX strips the -v tail so #438 cannot return through a per-build image cache name.

Mutation-tested on this tree

mutation	result
worker hard-codes its purge prefix	1 fail
stamp script re-declares a prefix (the shipped bug)	2 fails
IMAGE_CACHE from the per-build version (#438)	2 fails

All return green on restore.

Behaviour upstream is unchanged, which is the point

A real production build still stamps scripthammer-v0.0.1-<sha>, and all three cache names remain purge-eligible. This is hardening, not a behaviour change.

568/568  pnpm test:scripts
  7/7    tests/unit/sw-images.test.ts
 exit 0  type-check, production build
  clean  gitleaks

Ported from TortoiseWolfe/geolarp#35, where it was built and proven first.