[bitbiter-dev]

Closes #283.

Summary

Capture date in one place. Month, Day and Year are set only by MediaAsset.DateCaptured's setter and have private setters. The flashback index columns can no longer drift from the capture date, and #250's in-place update gets them right by assigning the date.

 public LocalDateTime DateCaptured
-{ get; set; }
-public int Month { get; set; }   // copied by hand in NewContentWriter and PairedMovResolver
+{
+    get;
+    set { field = value; Month = value.Month; Day = value.Day; Year = value.Year; }
+}
+public int Month { get; private set; }

The last-write-time fallback for a file without an EXIF date is one helper, CaptureDate.FromExifOrLastWrite, used by both builders.

Proxy rows added one way. Both writers hand their rows to IMediaAssetRepository.AddProxyFiles(context.ProxyFilesOf(ownerId)), which wraps AddRange:

 NewContentWriter.TryWriteAsync
   repository.Add(asset)                              // asset built with ProxyFiles = [...] before
+  repository.AddProxyFiles(context.ProxyFilesOf(asset.Id))
 PairedMovProxyWriter.WriteMissingAsync
-  foreach proxy in context.ProxyFiles.Where(AssetId == mov.Id)
-    repository.AddProxyFile(proxy)
+  repository.AddProxyFiles(context.ProxyFilesOf(mov.Id))

An explicit Add keeps a row whose Guid key is already set in the Added state. A row discovered through a navigation can be tracked as Modified instead.

The schema and the EF model are unchanged: docs/schema.sql is current and EF reports no pending model changes. Loading a row writes the backing fields directly, so the setter never runs on read.

Evidence

• Before: NewContentWriterTests.TryWriteAsync_Always_AddsTheContextsProxyFilesToTheRepositoryAsync and ..._AddsOnlyTheAssetsOwnProxyFilesAsync failed, because rows went through the navigation. MediaAssetTests.DateCaptured_Set_SetsTheFlashbackDateParts failed, because the parts stayed 0.
  After: all pass. Also new: TryWriteAsync_LivePhotoPairWithoutExifDate_DatesEachHalfByItsOwnFileAsync, which dates the HEIC from 2021-03-04 and the MOV from 2021-03-05. It pinned the fallback before the extraction.
• All test projects pass (1,304 tests), and dotnet format --verify-no-changes is clean.
• scripts/mutation-changed.sh: 100.00% (77 killed, 0 survived) across MediaAsset, CaptureDate, both builders, IngestionContext, PairedMovProxyWriter, MediaAssetRepository and the pipeline middlewares the end-to-end blurhash test covers.

Merge Danger

Door: two-way

Blast Radius: Worker ingestion writes

No schema change. The risk is EF's insert behaviour on real PostgreSQL. Proxy rows are now separate Added entries next to their asset, not children in its navigation, and EF still orders the inserts by foreign key. CI's unit tests don't run against PostgreSQL, so the first real crawl after merge is the end-to-end check.

🤖 Generated with Claude Code