suchi v0.1.0-beta.2

Added

• Screenshot paste previews with explicit upload confirmation and file-picker
  fallback, using the same upload path as drag-and-drop.
• Full extracted-text reading and copying, including manual copy fallback and
  the existing sensitive-document reveal gate.
• Local .ics downloads for reviewed exact-day Calendar entries, with explicit
  export disclosure and private source-document links.
• Browser-local QR codes for opening private documents on a phone and for
  existing document or selection share links, with ordinary copy fallbacks.
• One bounded rich query language across ranked search, Documents, the
  omnibox, and new saved views, with text prefixes, phrases, negation,
  filing metadata, dates, document state, qualifier suggestions, and
  apply-time validation.
• A scoped Archive research desk with reauthorized follow-ups, structured
  citation validation, Views made from retrieved documents, bounded model
  traffic, and explicit sensitive-evidence consent.
• A generic extracted-fact ledger with single-call date extraction,
  user-controlled confidence-based Calendar entry, optional bulk review, and
  rich-query date filters.
• A first-visit public-demo guide for rich queries and source-backed Calendar
  dates, plus a persistent help launcher. Archive research is clearly marked
  private-installation-only; public model access remains denied.
• A fixed 30-day Trash recovery window with automatic expiry, confirmed
  permanent deletion and Empty Trash actions, share-link revocation, minimal
  purge auditing. Original and derived blobs remain until offline GC; online
  cleanup cannot safely identify in-flight uploads reusing those bytes.

Changed

• Small document controls share one deferred bundle, reducing document-screen
  requests without loading administration or QR encoding on the initial page.
  Production-browser tests now exercise generated assets through local preview.
• Upload receipts distinguish new, duplicate, restored and failed files, and
  retain honest processing state when automatic status checks pause.
• Settings displays the release version in production. Development builds
  identify the source version line and revision, or explicitly unavailable metadata.
• OCR language documentation now covers installing packs in both Docker variants
  and bare-metal deployments, verifying model discovery, configuration precedence
  and rescanning existing documents.
• The web app imports screens directly and defers archive/mailbox configuration,
  removing six route wrappers. At the same dependency pins, Documents loads 48%
  less route JavaScript, Upload 83% less, and My account without mailbox access
  86% less. Finer chunks increase the all-routes compressed total by 8.3%.
• doctor drops a misleading shard-capacity sample whose threshold cannot
  occur in the hash-prefix layout. Explicit CAS integrity checks are unchanged.
• Reviewed tool and dependency pins advance to compatible previous-stable
  releases, including Go 1.27.0, Bun 1.4.1, AnyDoc 0.2.3, and SQLite 1.57.
  CI actions use immutable commits, and AnyDoc builds honor its Cargo lockfile.
• Release preflight now includes fresh code/UI checks and advisory scans.
  Artifacts-only branch builds use legal snapshot names and skip signing;
  image publication waits for the binary builds.
• Ready jobs and follow-up pipeline stages drain without a polling delay
  between batches. Document lists load correspondent names in one query.
• The web app reuses its date formatter and removes redundant search state
  updates and request wrappers.
• Beta.2 schema changes now ship as one migration, so beta.1 archives advance
  in a single transactional step.
• Rich-query lists now start from matching FTS rows, ranked Search bounds
  recency snippet work to its result page, and newest document pages use a
  stable partial index.
• Bulk document authorization is batched, and concurrent date reviewers now
  report and audit only the decision that actually changed each candidate.
• The web app cancels superseded list, search, Calendar, and completion reads;
  caches Calendar date formatters; loads route CSS lazily; and retains at most
  20 Archive research turns.
• Sampled rescans keep memory proportional to the requested sample, idle
  research rate entries expire, and date extraction holds SQLite's writer for
  less time.

Fixed

• Dismissed processing-update approvals no longer reappear after restart for
  the same pipeline revision. Startup closes already-recreated duplicates while
  preserving explicitly approved work and prompts for newer revisions.
• Public demo Calendar now exposes read-only, source-backed example dates
  across months and years. The tour links to usable Search and Calendar screens
  and clearly identifies Archive research as private-installation-only;
  demo model calls, date mutations, and non-demo facts remain denied.
• Demo navigation and first-visit tour detection now use the server's actual
  session identity field, including after upgrade to a scratch session.
• Documents and Inbox preserve pagination through browser Back, Forward, and
  reload, with numbered page links and URL-backed date filters. Invalid or
  vanished pages return to an available page without adding a history entry.
• Touchscreen document links no longer shift under taps when hover-only quick
  actions appear; filing and Trash remain available in document detail.
• Preset filing keywords match whole words and phrases: lease no longer
  matches please. Existing unchanged preset rules receive the same correction.
• Filing approvals close as superseded when a document has already been filed
  elsewhere, preserving its current category and recording the dismissal.
• Email backfills fetch small, checkpointed batches instead of one oversized
  IMAP command, preserving completed progress when a later batch fails.
• Successful mailbox connection tests no longer hide ingestion errors or
  overwrite the last successful sync timestamp.
• Successful confident classification clears only its own obsolete review tags.
  Existing, manually assigned and rule-maintained tags remain untouched.
• Photo OCR keeps its normal text and supplements it with confident lines from
  one bounded sparse-text pass, recovering isolated headings without replacing
  invoice text or changing ordinary PDF OCR. Low-confidence and short background
  noise is filtered; curved labels and perspective can still limit recognition.
• Difficult photographed QR codes now get a bounded local ZBar fallback when
  the Go reader cannot decode them; both runtime images include the tool.
• Lightweight OCR retries empty pages once using sparse-text segmentation for
  isolated labels. Rasterization, pages and retries share one real timeout, and
  cancellation/output-limit failures cannot be recorded as successful empty OCR.
  OCR revision 2 offers the existing controlled rescan path for older results.
• Camera photos retain their EXIF orientation in generated PDFs. Raster images
  use 300-DPI page density without changing pixel resolution, preventing the
  OCR stage from enlarging ordinary phone photos into hundreds of megapixels.
  Content remains at revision 2; selected rescans repair older results from
  immutable originals. Rescan prompts omit internal revision numbers.
• Approval cards omit internal workflow names, assignee IDs and step metadata.
  Decisions retain confidence, evidence and deadlines; rescan proposals list
  their affected documents under an explicit label.
• Created share links remain visible and selectable when clipboard access is
  unavailable. Copy can be retried without creating another link; delayed
  creation responses cannot copy links after account or document navigation.
• Startup rejects duplicate or invalid migration versions before changing any
  schema, identifying the conflicting migrations for developers.
• Dropping files onto the upload dialog submits each file once. Uploads dropped
  elsewhere refresh the visible document list immediately.
• Trash documents open in the existing viewer with read-only metadata and
  owner/admin preview, download, restore, and confirmed permanent deletion.
  Mobile rows keep document titles above their actions; sensitive reveal gates
  and public-share exclusions remain in place.
• Archive research opens its retrieved documents' Calendar dates in an all-dates
  agenda across months and years, with explicit scope and pagination. New date
  links clear stale Calendar filters; the ordinary Calendar stays month-based.
  Partial dates no longer display invented days in Calendar or Approvals;
  Calendar keeps labeled model confidence behind Automatic/Reviewed details.
• Calendar day links open paginated exact-day agendas, retain the month, View,
  and role in the URL, and return to the originating month. Server-side precision
  filtering excludes month/year placeholders before counting or pagination.
• Archive research allows a bounded 4,096-token generation budget so reasoning
  models can finish their JSON answer. Truncation has a distinct error; citation
  validation and the 6,000-character answer limit remain unchanged. Failure logs
  include fixed diagnostic reasons without question, evidence, or provider text.
• Settings identifies the running server version and source revision for bug
  reports. JSON API errors retain their codes even with mislabeled response
  headers, so Archive research can show the relevant failure message.
• Mail attachments with invalid or generic MIME headers are identified from
  their bytes. Explicit rescans repair existing PDFs labeled as binary files
  and restore preview and password handling without changing originals.
• Both container variants now produce real PDFs from images, including HEIC:
  standard includes the missing ImageMagick PDF encoder, and full permits PDF
  writing while keeping decoding restricted. The duplicate HEIC conversion
  path is gone; non-PDF converter outpu...

Suchi v0.1.0-beta.1

Added

• A local-first document archive in one Go binary, with an embedded web app,
  SQLite storage, full-text search, metadata, versions, and filing trees.
• Browser, API, watched-folder, and rules-based IMAP intake, including Microsoft
  sign-in, app-password providers, and Paperless-ngx bundle import.
• Ready-made filing trees, deterministic automations, optional local or hosted
  classification, approvals, and permanent Archive configuration.
• Source history, saved and shared views, secure email previews, document
  sharing, API tokens, MCP access, and multi-user permissions.

Changed

• Fresh installations start with only System/Inbox and require an explicit
  filing-tree or Blank choice; setup controls remain available afterward.
• Account and archive configuration have separate workspaces, with one home for
  mailbox, automation, taxonomy, and saved-view actions.
• Duplicate content records each distinct source, explicit configuration stays
  authoritative, and inactive interface work is deferred for faster navigation.

Fixed

• Mail intake advances ignored messages correctly, filters only ingestible
  attachments, and no longer leaves attachment-only source emails in Trash.
• Setup reminders retire after engagement, filing-tree selection, completion,
  or 48 hours instead of blocking Archive configuration indefinitely.
• Access checks now cover every document action; Confidential and Restricted
  thumbnails, previews, and extracted text share the same reveal protection.
• Filing-tree replacement, classification, polling, ingest jobs, taxonomy
  import, and configuration reloads preserve state and recover cleanly.
• Saved-view links, missing thumbnails, fast route changes, pipeline rescans,
  and password-protected documents no longer produce misleading interface state.

Security

• Session and token secrets are digested; provider keys, mailbox credentials,
  and saved document passwords are sealed at rest.
• Setup, demo isolation, error responses, and capability removal fail closed.