fix(mental-models): let PATCH update content and stop misreporting empty updates as 404

[JoshFunnell]

Two small fixes to PATCH /v1/default/banks/{bank_id}/mental-models/{mental_model_id}.

1. content is settable on the engine but unreachable through the API

MemoryEngine.update_mental_model already accepts content= and does the right things with it — recomputes the embedding, records the previous content in mental_model_history, advances last_refreshed_at. The HTTP layer just never passes it: UpdateMentalModelRequest has no content field, so the capability exists and cannot be reached.

The concrete case is restoring a previous version from mental-model history. Today the only way to change a model's content over HTTP is to trigger a refresh and accept whatever the LLM produces, which is not a restore. /clear empties content for a clean rebuild; it cannot put a known-good version back.

This adds the field and forwards it. No engine change.

Worth being explicit about, since it is a product decision rather than a mechanical one: setting content goes through the existing engine path, which advances last_refreshed_at. A restore therefore counts as a refresh for staleness and scheduling purposes. That is inherited behaviour, not something this PR introduces, but it becomes reachable here for the first time — so if a restore should not reset that clock, that is an engine change and a separate discussion.

2. An empty update reports 404 for a model that exists

The handler passes every field through and then maps a None return to:

raise HTTPException(status_code=404, detail=f"Mental model '{mental_model_id}' not found")

update_mental_model returns None when there is nothing to update, so PATCH with an empty or all-null body answers 404 "not found" for a model that is present and fine. That sends the caller looking for a missing model instead of at their own request.

Now rejected up front with 422 when none of the fields this handler accepts are set (name, content, source_query, max_tokens, tags, trigger), with a message naming them. Empty and all-null bodies are rejected before the engine is called; a missing model still returns 404 whenever the body has at least one accepted field.

On field semantics, since it is the obvious question: omitting a field, or sending null, means "leave it alone". Any non-null value means "set it" — including "", which is therefore a second way to empty a model alongside /clear, differing in that it writes a history row and advances the clock. There are no length bounds on the field; validation, embedding and history are all the existing engine path.

Note the engine's return value stays overloaded — None still means both "not found" and "nothing to do" for any other caller. This PR peels the empty-body case off at the HTTP layer rather than changing that contract.

Tests

Three, in test_reflections.py, run locally against the embedded Postgres fixture:

• test_update_mental_model_content_via_api — content set through the API round-trips (PATCH echo and a fresh GET), a history row records the previous content, and last_refreshed_at advances. Embedding recomputation rides the same engine path but has no API-visible surface, so it is not asserted.
• test_update_mental_model_empty_body_returns_422_not_404 — pins the fix, and fails before it.
• test_update_mental_model_name_only_unchanged_behavior — a name-only PATCH still succeeds unchanged, so the new branch does not swallow ordinary single-field updates.

Generated artifacts

openapi.json, the bank-template schema, the docs skill, and the Go/Python clients are regenerated. The TypeScript entry in generated/types.gen.ts is the additive content field only, matching what #3332 and #3223 landed for their request-model fields — a full local TS regeneration rewrites ~7k lines across all 16 generated files even on a pristine main with the pinned generator and CI's Node 20, because hindsight-clients/typescript has no package-lock.json and the generator's transitive dependencies float. That drift is unrelated to this change and is not carried here.

Surfaces deliberately not updated

The repo is wider than the engine, so rather than leave these silent:

Surface	State
hindsight-api-slim handler + request model	Updated
openapi.json (+ docs skill copy)	Updated, regenerated
Go / Python / TypeScript clients	Updated, regenerated
hindsight-cli — mental-model update	Not updated. No --content flag. Setting content is a restore/repair operation and a whole document body is awkward as a CLI flag; mental-model refresh remains the CLI path for regenerating content. Note the coverage manifest needs no entry: cli-coverage-check already counts content as covered because it matches the identifier globally in main.rs, where it belongs to other commands. Happy to add the flag if you'd prefer it exposed.
hindsight-control-plane — mental-models-view.tsx	Not updated. No content editor added; the view is unchanged and unaffected.
Rust client	Not updated by hand — it is generated at build time from the spec, so it picks the field up automatically.
Helm / docker / monitoring	Not applicable — no configuration or deployment surface changes.

Scope

Nothing else changes. No new engine behaviour, no change to the refresh implementation, and the 422 fires only on a body with none of the accepted fields set.