feat(files): begin Filesystem Collections ingestion

[devnull03]

Overview

This draft begins the Filesystem Collections release: qrate will build archival components from files and folders, accept dropped material in open projects, and expose nested descriptions as collapsible rows.

The first implementation slice adds a shared filesystem inventory used by project creation and open-project file resolution. Recursive discovery now preserves relative paths, so files such as photographs/001.jpg and documents/001.jpg remain distinct instead of collapsing to the same basename. The photo index understands those relative paths while retaining basename matching for existing projects.

The design plan in docs/dev/filesystem-collections-plan.md covers the remaining schema v4 migration, standards-aware field roles and labels, import preview, drag-and-drop surfaces, hierarchy editing, interchange, and release sequencing. Parent descriptions are normal archival components and will participate in search, diagnostics, galleries, exports, plugins, and Google Sheets.

Implemented in this draft

• Add a UI-independent file-ingest crate with deterministic recursive discovery.
• Preserve directories, parent relationships, and relative paths in the inventory.
• Skip filesystem cruft and directory symlinks, with bounded warnings for partial scans.
• Use the shared scanner in project-wizard folder discovery.
• Use the shared scanner in the table photo index and resolve duplicate basenames by relative path.
• Add focused coverage for nesting, recursion controls, ordering, ignored files, and duplicate filenames.

Verification

• cargo fmt --all --check — passed
• cargo test -p file-ingest -p project-wizard -p table --target-dir target/rust-analyzer — passed (101 tests)
• cargo clippy -p file-ingest -p project-wizard -p table --all-targets --target-dir target/rust-analyzer -- -D warnings -A dead_code — passed
• git diff --check — passed

The PR remains a draft while the higher-level schema and UI slices are implemented.

---

Import duplicates (ASNT-77)

Closes #87 (Notion ID: 3b821d32-b13b-818d-a1b6-c542b16fe2d7)

Re-importing a batch folder no longer copies what the project already holds. file_ingest::duplicates matches an import against existing components by stored source path, then by a single filename-key match; anything several rows could claim stays ambiguous and is never merged automatically.

• Skip (default), Update, or Add as new decide only the matches. Update re-links the row and leaves its metadata and its place in the arrangement untouched — qrate links files in place, so nothing is overwritten.
• A skipped folder still adopts new files below it, so a second batch lands inside the series the first one made.
• The wizard asks only when a folder holds a file that several rows name; the open table asks only when a drop repeats material. The answer is saved per project.
• The wizard and the table now match files with the same key rule, so a multi-part item counted as matched in the Files step is also linked at creation.

Plan and the two deliberate deviations: docs/dev/import-duplicates-plan.md.

Audit fixes in this branch

• Hierarchy is written in the same transaction as its rows. It was saved ahead of them, so structure naming an unsaved row failed its foreign key and the next save restored the stale copy — imported hierarchy was lost on reopen.
• Inserted and duplicated rows appear beside their neighbour instead of at the end of the table.
• Dropped roots queue after existing siblings rather than landing among them.
• Deleting a parent promotes its children into its place, not to the top level.
• Export fills structure columns only where the cell is blank, so typed values survive.
• Loading, creating, importing and saving log their row and component counts; skipped paths are named.

Verification (latest)

• ./scripts/ci.sh — passed (fmt, clippy -D warnings, full workspace tests)