docs(protocol): preserveAudit 的 readonly 白名单只在 UPDATE 半边成立 —— state-machine.mdx 的 create 半边旧说法 (#6827)

content/docs/protocol/objectql/state-machine.mdx

@@ -110,8 +110,9 @@ transitions: {
 - Only a rule with `severity: 'error'` (the default) blocks the write; `warning`/`info` are logged.
 - **Seed writes are exempt** (#3433). Curated seed data — package bootstrap fixtures, marketplace templates, per-org replay, all loaded by `SeedLoaderService` — is a snapshot of established facts, not a record walking its lifecycle, so it bypasses the `state_machine` rule entirely: a seed may be born mid-lifecycle (a `completed` project, a `closed_won` opportunity) and neither `initialStates` (insert) nor `transitions` (update) is enforced. Every *other* validation still runs, so a seed must still satisfy field shape, `format`, `script`, and the rest. `os lint` warns when a seeded value is not a state the machine declares, so a typo is still caught before boot.
 - **A "historical" data import is exempt too** (#3479). Migrating established facts — a batch of already-`closed` tickets, `closed_won` deals — is the same "snapshot, not a lifecycle event" situation. Set `treatAsHistorical: true` on the import request (default **off**) and the runner puts `skipStateMachine` on the write context, so `initialStates` doesn't reject those mid-lifecycle rows. A normal import leaves it off and still walks the FSM — the strict behavior is the default, so the exemption is always an explicit opt-in.
-- **`treatAsHistorical` also preserves the original audit timeline** (#3493). Skipping the FSM is only half of migrating established facts; the other half is keeping *when* they happened and *who* did them. Under the same flag the write context also carries `preserveAudit`, which (1) makes `updated_at` / `updated_by` **client-preferred** — a supplied historical last-modified survives instead of being stamped with the import instant, symmetric with how `created_at` / `created_by` already behave on insert — and (2) admits a **whitelist** through the static-`readonly` write strip: the audit/timestamp family plus author-declared business `readonly` fields (`closed_at`, `resolved_by`, …), so an `upsert` refresh no longer silently drops them. Platform-managed `system` columns outside that family (`organization_id` and other tenancy/generated columns) stay stripped — a historical import reinstates facts, it does not forge tenancy. Like the FSM exemption this is opt-in: a normal write still auto-stamps `updated_at`/`updated_by` and strips `readonly` exactly as before, and permissions / RLS / field-level security are unchanged.
-- **Undoing a historical import is symmetric** (#3549 / #3556). The import undo (`POST /api/v1/data/import/jobs/:jobId/undo`) logically rolls back a finished job — deleting the rows it created and restoring the captured pre-import snapshot on the rows it updated. That restore write now carries `preserveAudit` too, but **only** when the job was flagged `treatAsHistorical`, so the snapshotted `updated_at` / `updated_by` and business `readonly` fields (`closed_at`, …) are reinstated verbatim instead of being re-stamped to the undo instant. Without it the undo would silently overwrite the very timeline the historical import preserved; a normal (non-historical) import's undo keeps the default stamp/strip.
+- **`treatAsHistorical` also preserves the original audit timeline** (#3493) — **on the rows an import UPDATES** (#6640). Skipping the FSM is only half of migrating established facts; the other half is keeping *when* they happened and *who* did them. Under the same flag the write context also carries `preserveAudit`, which (1) makes `updated_at` / `updated_by` **client-preferred** — a supplied historical last-modified survives instead of being stamped with the import instant — and (2) admits a **whitelist** through the static-`readonly` write strip: the audit/timestamp family plus author-declared business `readonly` fields (`closed_at`, `resolved_by`, …). Platform-managed `system` columns outside that family (`organization_id` and other tenancy/generated columns) stay stripped — a historical import reinstates facts, it does not forge tenancy. Like the FSM exemption this is opt-in: a normal write still auto-stamps `updated_at`/`updated_by` and strips `readonly` exactly as before, and permissions / RLS / field-level security are unchanged.
+- **…but a historical `upsert` still drops those columns from the rows it CREATES** (#6640). `preserveAudit` is an **UPDATE-path exemption and nothing else reads it**, because the two write paths run two different strips: UPDATE is stripped inside the engine (`stripReadonlyFields`), which consults `preserveAudit`; CREATE is stripped earlier, at the DataProtocol ingress (`stripReadonlyForInsert`, #3043) that every REST-import create travels, and that one's only exemption is `context.isSystem`. So a single `treatAsHistorical` upsert keeps `closed_at` on the rows it **matches** and strips it — together with a supplied `created_at` / `updated_at`, which the injected audit columns also declare `readonly` — from the rows it **inserts**. The asymmetry is deliberate, not an oversight: `treatAsHistorical` arrives on an ordinary (non-system) import request, so honouring it on create would let any caller seed the approval/status columns that create-side strip exists to protect. The ignored request is at least no longer silent — the server logs a `WARN` naming the object, the stripped fields and this UPDATE-only rule — but the strip still applies. **To replay archival read-only facts on the rows an import creates, write from a system context** (`isSystem`). Full rule and rationale: [Security & Access Control](/docs/protocol/objectql/security).
+- **Undoing a historical import is symmetric** (#3549 / #3556). The import undo (`POST /api/v1/data/import/jobs/:jobId/undo`) logically rolls back a finished job — deleting the rows it created and restoring the captured pre-import snapshot on the rows it updated. That restore write now carries `preserveAudit` too, but **only** when the job was flagged `treatAsHistorical`, so the snapshotted `updated_at` / `updated_by` and business `readonly` fields (`closed_at`, …) are reinstated verbatim instead of being re-stamped to the undo instant. The undo is unaffected by the create-side carve-out above: it only ever *deletes* the rows the import created and *updates* the rows it touched, so every write it makes is on the path where the exemption is real. Without it the undo would silently overwrite the very timeline the historical import preserved; a normal (non-historical) import's undo keeps the default stamp/strip.
 
 ### Conditional transitions