feat(documents): wrap PlainText/PDF/DOCX as adapters and route DocumentParser through the registry

[mimeding]

Rebased onto origin/main after #1015. The diff is now the adapter wrappers plus the DocumentParser registry shim.

Business rationale

Osaurus today ingests every attached file through a single DocumentParser switch. To expand file-format fidelity (XLSX cells with formulas, DOCX tables, PDF tables, MT940/OFX bank exports), that switch has to move onto a pluggable adapter surface — and it has to do so without a user-visible behaviour change, because chat attachments are a core workflow. This PR is the bridge: it migrates the three file types already handled (plain text, PDF, DOCX-family) onto the adapter surface from #926 and proves the DocumentParser shim routes correctly. It deliberately adds zero fidelity; it's the groundwork every later adapter plugs into.

Coding rationale

Three design decisions worth calling out:

• DocumentParser.parseAll stays synchronous. The adapter protocol is async (parse(url:sizeLimit:) async throws -> StructuredDocument) because streaming adapters land in the next slice. Both existing call sites (FloatingInputCard, ClipboardService) already wrap parseAll in Task.detached, so the sync-to-async bridge inside the shim (a dispatch-semaphore + detached inner Task) blocks a cooperative thread rather than the main thread. Making parseAll itself async would touch every ingress call site and is out of scope for this PR.
• The PDF image-fallback stays in the legacy switch. PDFAdapter throws DocumentAdapterError.emptyContent when a PDF has no text layer; the shim catches that specific error and falls through to parsePDFWithFallback, which still renders each page as PNG. Moving the image path onto the adapter surface needs a typed PDFDocument representation that isn't in scope until a later fidelity slice.
• One RichDocumentAdapter covers DOCX, DOC, RTF, RTFD, HTML. They share the same NSAttributedString(url:options:documentAttributes:) call today, so a single adapter keeps the migration minimal. The high-fidelity DOCX reader can split this along format lines later; at that point this adapter becomes the RTF/HTML-only path.

What changed and why

• Packages/OsaurusCore/Services/Documents/PlainTextAdapter.swift: wraps the existing UTF-8-first / ISO-Latin-1-retry decoding path from DocumentParser.parsePlainText, including the 500K-character truncation marker, so the legacy behaviour stays byte-identical.
• Packages/OsaurusCore/Services/Documents/PDFAdapter.swift: wraps PDFDocument.page.string extraction; throws .emptyContent for text-layerless PDFs so the shim can fall through.
• Packages/OsaurusCore/Services/Documents/RichDocumentAdapter.swift: wraps NSAttributedString(url:options:documentAttributes:) across the rich-document formats; one adapter today, splits later.
• Packages/OsaurusCore/Managers/Documents/DocumentAdaptersBootstrap.swift: idempotent registration entry point.
• Packages/OsaurusCore/Models/Documents/PlainTextRepresentation.swift: neutral representation for adapters that don't yet publish a format-native shape.
• Packages/OsaurusCore/Utils/DocumentParser.swift: parseAll now tries the registry first, falls through to the existing switch on .emptyContent / .unsupportedFormat, and translates other DocumentAdapterError cases into the legacy ParseError surface. The translation is conservative — sizeLimitExceeded becomes fileTooLarge, readFailed preserves the underlying reason, anything else becomes a generic read error.
• Packages/OsaurusCore/AppDelegate.swift: calls DocumentAdaptersBootstrap.registerBuiltIns() near the start of applicationDidFinishLaunching so the shim sees adapters on the first file ingress.
• Packages/OsaurusCore/Tests/Documents/*: unit tests for each adapter (text decoding, size-limit refusal, empty-content handling, PDF text extraction, PDF without text layer, RTF/HTML extraction) plus integration tests for the shim (routing through a registered fixture adapter, fall-through on empty content, legacy path still works with no adapter registered, bootstrap registers all three built-ins on an isolated registry).

Validation

• swift build --package-path Packages/OsaurusCore — passed.
• swift build --package-path Packages/OsaurusCore -c release — passed.
• swift test --package-path Packages/OsaurusCore — passed: 1,454 tests in 196 suites. Sandbox integration skipped by its documented env gate.
• xcrun swift-format lint --strict on every touched Swift file — passed.
• swiftlint lint --strict on every touched Swift file — passed.
• git diff --check — passed.
• Packages/OsaurusCLI gate skipped; this PR does not touch CLI code.

Non-scope

• No new format support. CSV is already plain-text today; it stays that way until the streaming CSV/TSV adapter slice. XLSX is the first real fidelity win.
• No changes to Attachment.Kind. Chat attachments still decode as .document(filename, content, fileSize); the new .structuredDocument case lands after several real adapters prove the model.
• No plugin host changes. The plugin host API that lets external plugins register their own adapters is a later slice.
• No size-limit UI. The per-format ceilings in DocumentLimits are internal defaults; a user-facing "Maximum attached document size" preference is deferred.

Residual risks

• The sync-to-async bridge in routeThroughRegistry relies on parseAll always being called from a task context, not directly from the main actor. Both existing call sites satisfy this. A comment documents the invariant.
• DocumentAdaptersBootstrap.didRegisterShared uses the same nonisolated(unsafe) + NSLock pattern as OsaurusPaths.overrideRoot; a future migration to an actor or OSAllocatedUnfairLock would be cleaner but isn't in scope.
• The runBlocking helper allocates a detached Task per parseAll call. That's one task allocation per file ingress, which is negligible compared to the PDFKit / NSAttributedString work it wraps.

[mimeding]

Recovery-plan note: this branch is still the base of the document stack. The current test-core failure maps to slowToolReturnsTimeoutEnvelopeBeforeBudgetExpires, and I opened #976 as the narrow fix for the timeout race in ToolRegistry.runToolBody.

Recommended order: land #975 first for the shared CI cache failure, then land #976, then rerun/rebase #927 before continuing the document stack (#929 -> #936 -> #937 -> #939 -> #940 -> #941 -> #942).