feat(perception): publish immutable lab session instances

docs/09_OBSERVATION_SESSIONS.md

@@ -102,6 +102,38 @@ the obsolete browser poll; it does not cancel the process-owned conversion.
 A queued job may legitimately wait behind another export and therefore uses the
 overall preparation deadline rather than the active-export heartbeat timeout.
 
+## Immutable LAB instances
+
+An accepted experiment is published as a new saved-session identity instead of
+changing the AI result selected for its source recording. A LAB instance binds:
+
+- a stable `LAB E…` marker and opaque session id;
+- the immutable source observation session;
+- the exact result, source-result and configuration SHA-256 identities;
+- run and publication timestamps plus JSON provenance;
+- a content-addressed compute job, LiDAR pack and integrated visual result.
+
+The source evidence remains the only raw sensor master. A LAB replay cache and
+unchanged camera inputs use hard links on the same filesystem; no second copy of
+the source RRD, video or native K1 capture is created. Derived arrays are stored
+only when the run did not already persist a viewer-ready form. Deleting a LAB
+catalog entry cannot delete the referenced source session, and a source session
+with published LAB instances fails closed on deletion.
+
+The visual result is resolved by the LAB session id, so `LAB E19` and `LAB E21`
+can coexist over `RAVNOVES00` without “latest accepted result” replacing an
+earlier experiment. The operator can open either row in **Сохранённые сессии**
+and use the same **Объекты 2D / Сегментация / Кубы 3D** controls. First
+publication must also warm the result-specific Rerun overlay; later opens reuse
+that verified cache and do not rerun AI inference.
+
+E21 is a bounded special case: its worker persisted 121 semantic mask hashes,
+not duplicate mask pixels. Its visual publication materializes a mask only
+after an exact SHA-256 match against the immutable accepted E19 semantic
+reference. The seven detector frames replaced by the bounded latest-wins queue
+remain explicit empty frames. This preserves the measured near-real-time
+behavior rather than presenting a fabricated gap-free run.
+
 The first preparation of a long capture can take time because every decodable
 point/pose message and every valid K1 `ModelingReport` is projected into RRD.
 Later openings reuse a verified, digest-bound cache. Derived RRD cache v9 is