fix(magecommand): static verify — classify a bundle re-split as expected

[pingiun]

A pipeline turned on FAST_STATIC_DEPLOY_VERIFY and it failed on bundles. The compile is not wrong; the verifier was.

Not a bug — measured, not asserted

Magento keeps one .min-sibling cache for a whole setup:static-content:deploy process, so in a combined multi-locale invocation the first locale poisons the plain name for every later one — its bundles drop files a solo deploy of that same locale keeps. magecommand scopes that cache per (area, locale), deliberately: it makes the rayon fan-out deterministic and equals N separate deploys. The pipeline invokes both sides combined, so the two disagree.

On the store that failed:

	reference	ours
bundled modules	1576	1579
modules missing from ours	—	0
shared modules with differing source	—	0

A strict superset. The three extras (csso.js, jquery.ui.touch-punch.js, slick.js) each have a deployed .min.js sibling — the only reason the reference could have dropped them. A module left out of the bundle is fetched from that sibling; both trees work, ours makes fewer requests.

And the oracle is the order-dependent side: deploying a locale alone gives a different result than deploying it second. Chasing that would trade away the determinism the deploy module is built around.

So the verifier learns the signature

A package's bundle files move to a new resplit bucket — clean, and unaffected by --strict, since the modules and their sources are identical and there is no stricter reading under which they are a defect — when, across the package's whole bundle set:

• every module the reference bundles is in ours with byte-identical source, and
• every module we bundle that the reference does not has a deployed .min sibling.

Only the bundle files move; anything else differing in the package keeps failing on its own merits. (I had this too strict at first — requiring every difference in the package to be a bundle — which meant unrelated CSS differences suppressed the classification entirely.)

Negative controls

Three, because a rule that explains everything is worthless:

• a module the reference bundles and we lack — a real defect;
• an extra module with no .min sibling — nothing could have dropped it;
• a shared module whose source differs — the split is irrelevant, the content is wrong.

None may be explained away.

Effect

On the affected store the combined-invocation verdict now equals the per-locale one exactly: 7 changed / 0 extra, from 12 changed + 1 extra.

What remains there is unrelated and untouched: three CSS files per locale (the url() fallback relocation and cssmin serialization gaps) plus the sri-hashes.json they feed. That gate cannot pass on that shop until the CSS work lands, so running it report-only until then is worth considering.

🤖 Generated with Claude Code