FEATURE: Synchronization

[JamesAlias]

Overview

Rule-based, automatic reconciliation of translations across workspaces and/or dimensions, built on the stale-translation projection that came with Retranslation. Where the inspector's Retranslate button is a per-node, on-demand action, a synchronization rule keeps a target dimension continuously in step with its source: content published into live/en is translated into de-review/de, reviewed there, and published from there.

Configuration is based on #54 by @mficzel:

Sitegeist:
  LostInTranslation:
    nodeTranslation:
      synchronization:
        - sourceWorkspaceName: live      # a publish into this workspace triggers the rule
          sourceDimension: en            # must equal the target dimension preset's referenceLanguage
          targetWorkspaceName: de-review # translated variants are written here
          targetDimension: de
          scope: Document                # Document (documents + content) | Content (content only)
          mode: auto                     # auto (translate on publish) | ask (defer to a manual "sync now")

Rules are opt-in — the default is [], and with no rules configured nothing in this PR runs.

Under the Hood

The design rests on one core decision:

The target dimension is a projection of the source dimension. The source language is the single source
of truth for properties, structure and subtree tags. Target-side changes are overwritten when a
synchronization runs — that is intended, and nothing tries to preserve them.

A rule has only two knobs:

• scope governs what synchronization may create, and nothing else. Content only fills in content below Documents that already exist in the target (adopting a Document into a language stays a deliberate manual act); Document also creates the missing Document variants.
• mode governs when the translation runs, never what synchronization does. For the same source-side change both modes reach the same target state. Everything cheap runs inline on every publish under either mode — subtree tags, deletions, the cross-workspace rebase. Only translating, which costs a DeepL call per node, is deferred by ask.

Two consequences worth reviewing explicitly:

• Deletion is tagging. Neos 9 soft-removes by applying the removed subtree tag, so deletions ride the tag mirror — and restoring from the trash bin mirrors back, which the old hard-removal mirror could never do. Any SubtreeTag is mirrored, not a fixed list; sync owns the target dimension's whole tag space.
• Tags sync as state, properties sync as deltas. Tags are diffed and target-side changes reverted; properties are stale-driven, so a target-side edit survives until the source changes that property. Both obey the axiom — they differ in how aggressively they reassert it, because reasserting a property costs a DeepL call and reasserting a tag costs nothing.

Entry points

Driver	Trigger	Coverage
SynchronizationCommandHook	a publish into a rule's source workspace	the publish delta only — incremental, no tree walk
WorkspaceSynchronizer	"Sync now" / "Sync all" in the backend module, post-publish prompt in the Neos UI, lostintranslation:synchronize	every stale record in the target dimension
FullWorkspaceSynchronizer	lostintranslation:synchronize --full	the whole source subtree, stale or not

• Backend module gains a Synchronization overview: every rule with its pending count, or why it cannot
  run at all (target workspace missing — rules never create it — or target not based on the source workspace,
  which a cross-workspace rule requires). A blocked rule reports no pending count, because nothing would act
  on it. A fresh install without ./flow cr:setup gets a guidance banner instead of an error.
• Neos UI raises a post-publish prompt for mode: ask rules ("translations are out of date — synchronize
  now?"). auto rules need none.
• CLI: lostintranslation:synchronize (--full, --dry-run, --skip-existing),
  lostintranslation:retranslate-node, lostintranslation:reconcile (prunes orphaned stale rows left by
  descendants of deleted aggregates).

Notes

• uriPathSegment slug coercion moved out of two hardcoded if ($name === 'uriPathSegment') sites into a
  per-property TranslatedPropertyPostProcessorInterface (options.translationPostProcessor, read parallel to
  options.automaticTranslation). UriPathSegmentPostProcessor is registered on
  Neos.Neos:Document.uriPathSegment but inactive unless the integrator also enables automaticTranslation
  there.
• A missing language dimension no longer bricks the Content Repository when the package is disabled. Both
  command hook factories return a no-op DisabledCommandHook when the feature is off; they still throw when it
  is on, because a missing dimension under an enabled package is a real misconfiguration.
• AI attribution/authorization state no longer leaks past a throwing cascade onto unrelated commands
  (SynchronizationCommandHook matches on set membership, and clears the tracker at the start of every publish).
• --dry-run is honest now. It previews variants, property updates, removals and tag changes without
  dispatching anything, without rebasing the target, and — the part that was actually broken — without paying
  the full DeepL bill to produce a report of zeros.
• Docs: Documentation/RetranslationAndSynchronization.md (811 lines) covers the projection, both features,
  all workflows, cross-workspace mechanics, superseded decisions and known gaps. README.md gains an
  integrate-into-an-existing-project section and per-command reference. TESTING.md documents the harness.
• Test harness & CI: a self-contained Docker Behat setup (make e2e, ~9s for the full suite) plus a Behat
  job in the GitHub workflow across PHP 8.2–8.4 × Neos 9.0–9.1; the build now also runs on 3.* branches.