fix github actions build

[Tsyklop]

This branch overhauls how framework lessons persist and propagate learner changes across stages, replacing the legacy diff-based storage with a git-like content-addressable history (snapshots / commits / refs), and fixes the resulting navigation, course-update, and submission flows. It also gets CI and the :intellij-plugin:test suite back to green.

Why
The legacy diff-based framework storage lost or mis-propagated learner code across stages (ALT-10961, ALT-10998, ALT-10961 follow-ups). A self-contained snapshot history makes navigation/update deterministic and restart-safe, and fixes the data-loss bugs the failing tests captured.

What changed

1. Framework lesson storage — git-like architecture (ALT-10961)
   FrameworkLessonManagerImpl reworked to store full per-stage snapshots as content-addressable blobs with commits and stage_ / step_ refs and a HEAD (see framework/impl/FRAMEWORK_STORAGE.md). Full contents (not diffs) are stored so state survives IDE restart / offline.
   FrameworkStorage / LegacyFrameworkStorage / framework/storage/UserChanges updated; legacy records are migrated on the fly.
   Navigation propagation reworked: snapshot-based target state, ancestor checks to decide when a merge is needed, and a Keep / Replace conflict path (PropagationConflictDialog) with auto-Keep/auto-Replace heuristics based on whether each side has learner changes.
2. Course update preserves learner changes
   FrameworkTaskUpdateInfo, FrameworkLessonHistory, UpdateUtils: when a course update adds/removes/changes a framework task file, the learner's local edits are preserved instead of being overwritten; author content is kept in the task model for later revert.
3. Supporting changes
   SolutionLoaderBase, TaskNavigationAction, handlers/handlersUtils, HyperskillUtils, HyperskillOpenInIdeRequestHandler, submissions/utils, ext/TaskExt/StudyItemExt, VirtualFileExt adjusted for the new storage model (original test/template file caching, ref handling, rename/move).

[meanmail]

Review of the framework-lesson storage rework.

The change looks well thought out and is generously commented, and the RemoveTaskFile.apply fix (state -= path instead of state[path] = text) is a real, correct bug fix. A few cross-cutting points in addition to the inline comments:

• Mixed line endings. FrameworkLessonManagerImpl.kt now mixes CRLF and LF (base had ~1689 CRLF lines, this branch ~1479), so many -/+ pairs in the diff are byte-for-byte identical and only differ by EOL. This roughly doubles the diff and hides the real changes. Suggest adding a .gitattributes (*.kt text eol=lf) and normalizing EOLs in a separate, content-free commit.
• LOG.warn for debug output. Several LOG.warn("saveExternalChanges: ... keys=...") / ChangeFile.apply: ... messages are debug-level detail logged at WARN. This will be noisy in production logs; please downgrade to debug/trace.

[meanmail]

Re-review — approved ✅

Thanks for the thorough follow-up. I re-checked the changes against the new revision (49d0371) rather than just the commit messages, and the previous round of feedback is handled well.

Addressed (verified in code):

• Removed the test-only currentTask.record fork in FrameworkLessonManagerImpl; the update test now seeds the legacy record itself.
• Dropped the EDU-XXXX placeholder wrappers — YamlFormatSynchronizer.saveItem is called directly.
• HyperskillOpenInIdeRequestHandler now injects the stages executor/loader, so the CompletableFuture / timeout / cancellation path runs under tests (HyperskillLoadStagesWithTimeoutTest covers Ok / timeout / failure / ProcessCanceledException).
• LegacyFrameworkStorage: java.io.ByteArrayInputStream imported; the deliberate version fall-through documented in both readers.
• Debug-level LOG.warn traces downgraded to LOG.debug; remaining WARN logs are genuine failures/caught exceptions.
• .kt line endings normalized to LF in a dedicated commit — the diff is reviewable now.
• pluginVerification.failureLevel: accepted the rationale (keep the marketplace-validated build green).

The RemoveTaskFile.apply fix (state -= path) remains the core correctness win.

Non-blocking, left open for your call:

• PropagationConflictDialog.kt:198 still has an isUnitTestMode branch — same injectable-seam idea as the Hyperskill handler would let tests exercise the real result path. A test-quality nit, not a correctness issue.
• RemoteEduTaskYamlMixin.checkProfile switched ALWAYS → NON_EMPTY: just confirm no backend/round-trip consumer requires check_profile to always be present.

Neither is a blocker, so approving. Could you resolve the addressed threads on your side? I don't have permission to resolve them in this repo.