feat: add chat projects with shared project memory (experiment) - #29242

Gives every chat in a project a durable, shared memory store modeled on
Claude Code's auto memory. Root chats in a project get a compact
<project-memory> index in the system prompt (guidance plus one line per
memory, capped at 200 lines / 25 KiB with an explicit truncation notice)
and read/save/delete tools that upsert by name. Memories carry a type
(user, feedback, project, reference), a one-line description, and a
markdown body, and are sanitized before storage.

After each completed root turn a detached, cursor-fenced extraction call
reviews the new transcript against the index and applies any upserts or
deletes the agent did not save itself. It never blocks the turn.

Adds the chat_project_memory RBAC resource (org members may read and
write), audited CRUD under /chats/projects/{project}/memories, and a
Memory section on the project page so users can inspect, edit, and
delete what the agent wrote. Gated by the chat-projects experiment.

…ation

Exercises extractProjectMemories end to end against a mocked store and an
intercepted model call: cursor fencing, transcript windowing by revision,
normalization of upserts and deletes, the per-project cap, and that model
failures leave the cursor untouched. Covers prepareGeneration placing the
project memory index and tools only on root chats in a project with the
experiment enabled.

Dogfooding with a small model surfaced three failure modes in the
background extractor: it deleted a correct memory because the assistant
said it did not know the fact, it re-recorded the assistant's restatement
of existing memories as new ones, and it split facts the agent had just
saved into duplicates.

The extractor is now upsert-only and reads only new user messages, skips
turns where the agent already used the memory tools, and the shared
guidance tells the agent to check the index before claiming something is
unknown and never to save unknowns.

A question-only turn led the extractor to rewrite a correct memory with
hallucinated content. Existing memories are now updated only by the main
agent's tool or the UI; extraction proposals that collide with an
existing name are dropped.

…as a subtle icon

…oject memories

The dialog stayed mounted across edits, so its form state never picked up
the selected memory. Remount it per memory.

Memories keep a name, one-line description, and body. The guidance still
tells the agent what kinds of facts are worth saving, without asking it to
categorize them.

Renumbers the chat projects migrations to 000595/000596 behind main's
000594_drop_task_tables and regenerates database and API artifacts.