fix(memory): stop remember and the memory tool from losing or reviving memories

[kevintseng]

Four ways a memory could be lost or come back without being asked:

• remember with replace: true and no content emptied the memory. A call that only meant to fix the title or tags (name + replace + title) cleared every observation, and the error text and the docs recommended exactly that call. It is now refused over MCP, HTTP and the CLI (an empty observations list too); the message says to send the observations the memory already has with the new title.
• MCP remember with a plain project:<name> tag skipped the session binding, so the memory was filed under a project no briefing reads. It is now stored under the bound project id and reported in retagged, as the tool description already said. A bare project: tag is refused on every transport.
• The memory tool showed and reactivated deleted memories. view, str_replace, insert and a second delete on a deleted path now say it does not exist; create over the path still brings it back on purpose.
• A memory tool edit that left the text unchanged raised the memory's confidence and reset every observation's time. It is now a no-op. (A real edit still does both; that needs a storage change and stays open in Editing a memory through the memory tool keeps raising its confidence #531.)

Verified locally: npm run verify green for the final tree; each fix has a test that fails when the fix is reverted.

Refs #531