[6.x] Ensure set preview images are updated when assets are renamed/replaced/deleted

[duncanmcclean]

This pull request ensures that Set Preview Images on Bard & Replicator fields are updated when assets are renamed/replaced/deleted, like we do with content.

I've also added Blueprint::all() which returns blueprints from all collections, taxonomies, etc.

Closes #12680

[jasonvarga]

Unless I've done something silly, this doesn't seem to work at all. I can't get it to rename anything.

Plus, I think it might not work on a second level if you've defined statamic.assets.set_preview_images.folder as the path won't get saved in the blueprint - just the filename. We might need to change that.

[jasonvarga]

This review was generated via a Claude-powered code review.

Solid, well-tested PR overall. The asset-listener flow is straightforward and the test coverage is comprehensive for the happy paths. A few real bugs and notes below.

---

Warning — Container is not filtered for blueprint/fieldset updates

src/Listeners/UpdateAssetReferences.php:117-135 calls the updater without filterByContainer($container), and src/Assets/AssetReferenceUpdater.php:76-134 (updateBlueprintFields) never checks the asset's container against Sets::previewImageConfig()['container']. The folder string is the only thing being matched.

Consequence: if statamic.assets.set_preview_images.container is previews but an asset is renamed in some other container (say documents) where the path happens to start with the same folder/, the blueprint's set.image (which logically points at the previews container) will be rewritten — corrupting the reference.

Suggested fix in updateBlueprintFields:

if (
    ! ($config = Sets::previewImageConfig())
    || $this->container !== $config['container']
    || ! Str::startsWith($this->originalValue, $config['folder'].'/')
) {
    return;
}

…and have the listener call ->filterByContainer($container) before ->updateReferences(...) on the Blueprint::all() / Fieldset::all() loops, mirroring the entry loop above.

A regression test where two containers share the same folder/filename would catch this.

---

Warning — Blueprint::all() misses several real namespaces

src/Fields/BlueprintRepository.php:27-37 hardcodes 'navigation', 'assets', 'globals', 'forms' plus collections/taxonomies/additional namespaces, but does not include:

• user and user_group (root-level, used at src/Auth/UserRepository.php:66 and src/Auth/UserGroupRepository.php:19)
• anything stored in the root via Blueprint::in('')

User blueprints can contain bard/replicator fields, so a preview image inside a user blueprint won't be updated. Worth either including the user/user_group blueprints explicitly, or iterating every file under directory() rather than enumerating namespaces.

The hardcoded list is also brittle for the future — any new core blueprint namespace will silently be missed by every reference updater.

---

Note — setParent(null) in makeBlueprintFromFile is unrelated and undocumented

src/Fields/BlueprintRepository.php:350 adds ->setParent(null) to every blueprint coming out of makeBlueprintFromFile. There's no test exercising it and the PR description doesn't mention it. It only runs when BLINK_FROM_FILE misses (so it doesn't clobber a parent set on a later find against the same path), but it will clobber any parent set during construction by something else that holds the same instance.

If this was needed because Blueprint::all() iterations were producing parent-leaked instances, that motivation should be either in a code comment or split into its own commit so the reasoning is recoverable.

---

Note — Performance: full blueprint/fieldset enumeration on every asset save

UpdateAssetReferences::replaceReferences now walks Blueprint::all() and Fieldset::all() on every save/replace/delete, regardless of whether statamic.assets.set_preview_images is configured. The early-return inside updateBlueprintFields saves the inner work, but instantiating every blueprint/fieldset still pays disk I/O.

A cheap up-front check at the top of replaceReferences would skip both loops entirely when the feature is off:

if (Sets::previewImageConfig()) {
    Blueprint::all()->each(...);
    Fieldset::all()->each(...);
}

(Combined with the container fix above, this can also short-circuit on container mismatch.)

---

Note — findFieldsInBlueprintContents matches anywhere type is a string

src/Data/DataReferenceUpdater.php:102-116 walks all nested arrays and triggers on any 'type' => 'bard' / 'type' => 'replicator' key/value pair. It doesn't constrain to "this is actually a fieldtype declaration." In practice, blueprint contents are unlikely to have these strings in validate/if/etc., and the downstream if (! isset($fieldContents['sets'])) skip makes false matches harmless. But it's worth a comment noting the heuristic — or restricting to keys living under field to make the intent obvious.

---

Note — Old set schema (no group wrapper) is silently skipped

updateBlueprintFields assumes the modern grouped structure (sets.<group>.sets.<setHandle>). A blueprint authored with the older flat schema (sets.<setHandle> directly, no group) hits the ! isset($setGroup['sets']) branch and returns untouched, even if it has an image. If 6.x still loads old-format blueprints from disk this is a gap; if not, ignore.

---

Tests look good

The cases in tests/Listeners/UpdateAssetReferencesTest.php cover rename, move-out-of-folder, delete, no-config, and the wrong-folder negative case. Two cases worth adding given the findings above:

1. Two containers: configure set_preview_images.container = 'A', place a same-named asset in container B, rename it, assert no blueprint changes. (Would catch the missing-container-filter bug.)
2. User blueprint with a bard set image. (Would catch the Blueprint::all() namespace gap.)

---

Misc

• src/Fields/FieldsetRepository.php:176, 187 — refreshing the $fieldsets cache on save/delete is a nice correctness fix; useful independently of this PR.
• Blueprint::all() is a generally useful new public API; might want a short PHPDoc explaining the namespace ordering since callers (and the it_gets_all_blueprints test) rely on it.