[nicoloboschi]

Closes #4340.

A delta mental model can now be rewritten from scratch — once on demand, or on its own schedule — without stopping being a delta model. Delta keeps unchanged sections byte-for-byte; over months of small edits a document drifts from what a fresh write would say. This gives users the "edit incrementally, rewrite weekly" pattern the issue asks for.

What changes

One-off full refresh — POST /mental-models/{id}/refresh takes an optional body {"mode": "full"}.

• The stored trigger.mode is untouched; the next plain refresh is a delta again.
• A full refresh that fails keeps the previous content (the old advice, clear + refresh, emptied the model first).
• A full submit that folds into an already-queued plain refresh upgrades that one to full (_upgrade_queued_refresh_to_full), so "full" is never silently dropped by dedupe.

Scheduled full refresh — new trigger.full_refresh_cron (UTC, 5-field).

• Independent of refresh_after_consolidation and refresh_cron — combine with either.
• Runs inside the existing cron scan (MaintenanceLoop._run_scheduled_mm_refresh): same croniter due check, tenant routing, skip_if_in_flight, automatic=True (so the min-interval floor and the failure pause apply).
• Comes due against a new mental_models.last_full_refreshed_at, not last_refreshed_at (every delta moves that one, so a weekly full refresh would never fire on a model that deltas daily).
• Its gate is "a refresh ran since the last full one", not staleness: drift left by past edits is still there on a quiet bank, and with no refresh since the last full one there is nothing to fix.
• When both crons are due, one refresh runs, in full mode.

Migration d4f7a2c9e1b5 — the column (PG + Oracle), a partial index on the new predicate, and mental_models_with_cron() recreated to also return full-refresh-only models (guarded by execute_unless_owned; the loop reads the new columns optionally so an operator-owned routine keeps working).

Surfaces

• Control plane: "Scheduled full refresh" switch + cron presets (Sundays / Monthly) under Incremental in the trigger form; "Full refresh now" in both action menus with a token-cost confirm; "Last full refresh" in the detail view. 10 locales.
• CLI: mental-model refresh --full, --trigger-full-refresh-cron.
• Python/TS wrappers: mode="full" / { mode: "full" }. MCP trigger field. OpenAPI + all SDKs regenerated.
• Docs: "Scheduled Full Refresh" section, full refresh on demand with tested examples (py/mjs/sh/go), the clear+refresh tip replaced.

Tests

• api-slim: scheduler (due against own clock, skip with nothing to fix, not stale-gated, full wins over plain, routine discovery), fold upgrade, engine (full on a delta model keeps delta; empty scope still stamps), HTTP body.
• System story test_39_full_refresh.py: a delta model's full refresh through the client — a dropped override would hit the unscripted delta-merge step and fail.
• Control-plane form round-trip, CLI trigger update, wrapper mapping on both sides.
• Verified by hand in the UI against a real LLM: delta edit → scheduler full refresh at the next cron mark → no further runs while nothing changed; "Full refresh now" from list and detail.