[bitbiter-dev]

Closes #250. Builds on #283, which landed separately as #288.

What changes

A file whose bytes change at a stored path (re-exported, rotated, metadata stripped) used to fail ingestion on every crawl. Its new hash was unknown, so it was inserted at a path the old row already held, which violated IX_MediaAssets_StorageConfigId_FilePath. Now the file is classified ContentChanged and updates that row in place:

• Kept: Id, interactions, BurstId, PairedAssetId. A Live Photo whose HEIC was edited stays paired to its MOV.
• New: ContentHash, capture date (DateCaptured, Month, Day, Year), MediaType, LastSeenOnNas, every Metadata field, and the proxy rows.
• Proxy rows are updated in place by type, so no row is deleted and re-inserted under ix_proxyfiles_assetid_proxytype. Missing types are added and types no longer generated are removed.
• Old proxy directory: left to the orphan sweeper (feat(worker): Orphaned proxy file sweeper (proxies ↔ DB reconciliation) #174) and documented in CLAUDE.md. It is not deleted after commit, because a copy of the old content elsewhere can be ingested into that same directory in the same crawl.

How

• The classifier gets a ContentChanged branch: no row has the hash, but an active row holds the path. A soft-deleted row at the path stays out of scope (Soft-delete assets whose file a complete crawl did not find #34).
• New ContentChangedWriter. It shares ProcessedContentWriter with NewContentWriter for the MOV block, the EXIF-to-Metadata mapping and the save. Capture date and proxy rows use refactor(worker): Capture date in one place, proxy rows added one way (#283) #288's single paths (DateCaptured setter, AddProxyFiles).
• New repository methods: LoadMetadataAsync and RemoveProxyFile.
• As the Update the row when a file's content changes at the same path #250 scope comment asked, IngestionContext.AttemptId is now separate from AssetId.
  • AttemptId names temporary files and scopes compensation.
  • AssetId is the row that owns the proxies, set by classification: the stored row for ContentChanged, the attempt id for New.
  • A failed attempt therefore deletes only its own proxies, never the row's previous ones. A runner test covers this.
• CLAUDE.md names ContentChanged wherever it lists the outcomes and writers, and documents the feat(worker): Orphaned proxy file sweeper (proxies ↔ DB reconciliation) #174 decision.

Deliberately not here

• An edited Live Photo MOV is classified ContentChanged on its own half, but the MOV row is not updated yet. It no longer fails either. Tracked in Update the MOV row when a Live Photo's video is edited at the same path #285.
• The row stays stale when the edited bytes are already an active row elsewhere, whether because a concurrent consumer won the race or because a copy was ingested first. The race is skipped rather than rethrown on purpose: rethrowing runs compensation over content-derived proxy paths shared with the winner. Tracked in Row at a path stays on old bytes when its edited content is already known elsewhere #286.
• FileName is not reassigned. The path is unchanged, so the name is too.

Verification

• dotnet test: all 1,327 tests pass.
• scripts/mutation-changed.sh (passed ClassificationMiddleware.cs and ImageProxyMiddleware.cs for the two test files that map to no single source): 100.00%, with 271 killed, 1 timeout and 0 survivors, after merging refactor(worker): Capture date in one place, proxy rows added one way (#283) #288. There are no new Stryker directives, so ADR 0005 has no new rows.
• EF delete and insert ordering was checked in a throwaway SQLite project before the in-place update replaced that approach.

🤖 Generated with Claude Code