v0.5.0: Stable IDs and safer database handling

Stable IDs and conflict-safe writes

Elements, connections, categories, and statuses now have stable IDs that stay the
same when a name or a path changes. An AI client can address a record by its ID
instead of its current name, and pass the version it read along with the change.

If the record changed in the meantime, Home Memory reports a conflict instead of
overwriting the newer data.

• The detail tools and list_statuses report the oid and a version number. The
  create tools return the oid of the new record.
• Every update, delete, and move tool accepts oid as an alternative to the name
  or path, plus an optional expected_version. When a name and an ID are given
  together, they must identify the same record.

Breaking changes

• get_structure_overview: the parameter structuralAreasOnly is now called
  primaryAreasOnly. The old name is not rejected, it is ignored, so a client
  that still sends structuralAreasOnly=false quietly gets the primary-area view
  instead of the full content tree. Rename it wherever it is stored. The wording
  "primary area" is used consistently in tool descriptions and category output.
• HTTP mode requires HOME_MEMORY_API_KEY when HOME_MEMORY_BIND is not a
  loopback address. The server now refuses to start instead of printing a
  warning. Requests are additionally checked against the expected Host, and
  requests carrying an Origin header are rejected, because Home Memory has no
  browser client.
• create_connection and update_connection require length to be greater than
  zero, and parse it strictly. Use a decimal point, for example 4.5. Thousands
  separators are no longer accepted.

Safer database handling

Normal startup no longer changes the schema of an existing database. Recognized
Home Memory databases continue to open unchanged.

If Home Memory reports that a migration is required, stop every process using the
database and run:

HomeMemoryMCP.exe --migrate --yes

The command refuses to run while the database is in use and writes a verified
backup before it changes anything. The backup path and its checksum are printed.
Seed data is applied only to a database that Home Memory created itself.

New

• get_structure_overview takes metrics="counts" and then appends per-node
  rollups [b… c… x…]: elements below the node, connections touching it, and
  connections crossing its boundary. Rollups always cover the full subtree,
  independent of primaryAreasOnly, maxDepth, and under.
• list_categories and get_category_details count items across all
  subcategories, matching what get_by_category and get_connections return.
  Where a count covers both, the directly classified share is shown as
  ; N direct.
• The detail tools show who created a record and when, and the same for the last
  change. Records imported from other tools may not carry that data.
• All tools carry MCP behavior annotations, so clients can tell reads from writes
  without parsing descriptions.
• Writes report non-blocking advisories: an element whose status is in an earlier
  lifecycle phase than its parent's status, or a new connection that reuses an
  existing connection name.

Improvements

• The stdio server stays available when another process holds the database at
  startup. The tools remain listed and the check is repeated on every tool call
  until it succeeds, so the client does not have to be restarted.
• Faster count rollups on larger structures.
• Database commands time out after 30 seconds, so a blocked query returns an
  error instead of leaving the client waiting.
• get_structure_overview no longer shows misleading paths when an area sits
  below an element that is not a primary area. The ancestors needed to reach it
  are marked with [context] and counted separately in the total line.
• Tool failures are returned as MCP errors instead of plain result text, so
  clients can tell a failed call from a successful one.
• Blocked deletions are described as stop signals throughout. Attached documents,
  child categories, category usage, and status references are reported for a
  decision instead of being cleared by the assistant.
• Startup errors, including permission and IO problems, print a clear message and
  exit with a non-zero exit code.
• Upgraded to ModelContextProtocol 1.4.1.

Fixes

• Search terms containing % or _ are matched literally in find_element and
  get_connections, across all searched fields.
• get_by_category with an unknown category returns an explicit error naming the
  category instead of reporting an empty result.
• get_connections(under=...) now includes connections whose source or
  destination is exactly that element, not only those below it.
• maxDepth is applied to the primary-area overview as well, and the returned
  title states the depth that was used.
• Renaming an element to a name it already has is no longer rejected as a
  conflict with itself.
• The check for attached documents queries the correct column, so deletions are
  blocked as documented.

Updated

• Server instructions and tool descriptions were tightened: the model is stated
  once at the server level, and the details sit in the tool and parameter
  descriptions where they are read in context.
• Seed data: Fire Detector replaces Smoke Detector, a Flue System category
  was added under HVAC, cable and wastewater categories were refined, and seeded
  categories carry canonical paths and purposes. Seed data applies to newly
  created databases only.
• Setup guide: client registration examples and a section on updating the schema
  of an existing database.