[Yariminal]

Scans can suggest risky changes when image similarity is mistaken for a disposable copy, a filename is borrowed from unrelated content, or a linked resource is moved on its own. Review and execution also need to survive interrupted analysis and keep concurrent CLI/server writes from racing approval. This change makes proposals evidence-aware and session-scoped, requires explicit review for non-exact duplicate candidates, previews the ordered file changes before commit, protects known resource bundles, and journals selected operations for undo. Server-owned run plans persist progress and support explicit resume or retry.

The review API and web UI now show source/keeper evidence, analysis limits, protection reasons and projected final paths. Stored MD5 grouping is identified separately from unavailable SHA-256 evidence; execution rechecks live SHA-256 before exact-copy trash. The naming and organizer gates preserve project identifiers and numbered sequences and suppress unsafe grouping suggestions. CLI and web writers share a cross-process operation lock.

Validation through 2026-09-28: root independently ran 263 tests with zero skips; JavaScript syntax, Git whitespace and ddh --help checks also passed. The suite includes exact-copy trash/undo, legacy database migration and two-session CLI scope. Two local web servers rejected review/commit while a separate CLI writer held the lease; after it exited, commit and CLI undo restored the synthetic asset. A two-interpreter six-file restart retained two completed analyses and processed only the four remaining files after explicit resume; the AI provider responses in that proof were stubbed. Across six fresh disposable corpora, 1,337 files (4,431,413,374 extracted bytes) yielded 1,321 indexed AI attempts: 780 analyses, 541 explicit skips, zero provider failures or unprocessed indexed files. Small, Photo and Links selected 13 total actions; all passed dry-run, commit and undo. Medium, Hebrew and Large had no safe file operation and passed zero-operation checks. Root's final audit matched all 1,337 original file hashes/paths, 1,321 indexed database paths/names, six source ZIP hashes, directories and foreign keys. Medium's 503 numbered frames kept their folder and names. See the validation report for per-corpus evidence and limits.

Cross-platform CI passed all nine Windows, Ubuntu and macOS jobs on Python 3.10, 3.11 and 3.12 for commit ae65f17. CI run. A separate clean Windows installation using only .[dev,web] also passed 263 tests and the CLI entry point.

Known limits: Medium still yields 476 non-exact duplicate candidates requiring individual review, so candidate precision is not established. Browser visual and accessibility checks were unavailable after the earlier local-browser permission denial. The Medium run was not crash-tested against live Ollama; the separate restart proof used stubbed model responses. Conservative resource protection does not prove every external application reference resolves.