diff --git a/harness/experiments/state-label-layout-20260924/README.md b/harness/experiments/state-label-layout-20260924/README.md
index 8c7e9e27..71707bfb 100644
--- a/harness/experiments/state-label-layout-20260924/README.md
+++ b/harness/experiments/state-label-layout-20260924/README.md
@@ -6,7 +6,7 @@ Issue [#36](http://217.177.34.108:3000/sorokin.oleg.edu/ARCADIA-Studio/issues/36
 
 The original `Ready\nfor review` title no longer collides with its summary. Routing prepares a shared, measured hard-line plan used by canvas paint, hit boxes, previews, focus/Fit and SVG. Opaque State labels receive hits before covered node/row pixels, in reverse paint order; exposed action rows retain their identity. Authored geometry and all semantic/source data remain unchanged. See the [contract](../../state-label-layout-contract.md) for the retained excerpt budgets, separators, measurement and scope.
 
-The bounded Windows implementation passed its focused core gates, full desktop suite, formatting, strict workspace/all-target Clippy, product builds, and maximized-window native B/W/U Save/reopen checks. Native hit acceptance is limited to the observed maximized window: an unsuccessful restored-window attempt and subsequent Home-button input discrepancy remain explicit in [observations](observations.md) and [independent native review](native-review.md). No all-window input or universal Unicode rendering claim follows. This is a follow-up contribution, not a replacement release or complete application acceptance.
+The bounded Windows implementation passed its focused core gates, full desktop suite, formatting, strict workspace/all-target Clippy, product builds, and maximized-window native B/W/U Save/reopen checks. A later isolated trial also passed Home Open and all three transition text-row hits at restored1442×932/displayed79%, preserving all nine source files and the entire Project. The earlier unsuccessful restored1442×842 attempt and Home-button discrepancy remain explicit in [observations](observations.md) and [independent native review](native-review.md); the later success does not establish their cause. No all-window input or universal Unicode rendering claim follows. This is a follow-up contribution, not a replacement release or complete application acceptance.
 
 ## Automated checks
 
@@ -38,6 +38,34 @@ Preserved GUI:119440896 bytes, SHA256`0401341388025544185455ad5520aa2c1b79ca1387
 
 ## Reusable evidence
 
+### Supplemental restored-window trial, September 25
+
+The [supplemental observations](restored-native-observations.md) record the same
+preserved70bdcdf8 executable, rechecked before launch, with a fresh isolated copy
+of B and a private recent-model profile. At1442×932/displayed79%, Home Open loaded
+the fixture. First title, second title, and summary clicks each selected stable
+transition alias `between`; the latter two followed independently observed
+selection of `state_b`. Pure Save preserved all nine source files byte-for-byte
+and the complete expected Project. Alt+F4 termination was observed and both GUI
+logs were empty, but the watcher returned a null exit code; this is not exit0
+evidence. The initial incorrect direct snapshot invocation rejected includes;
+the established converter-then-snapshot verification subsequently passed.
+That invocation error is preserved narratively; its original stderr was replaced
+by the successful rerun's empty log. The [independent supplemental review](restored-native-review.md)
+checks all ten original images, archive members, source/model comparisons and
+tool hashes, and retains this evidence limitation.
+
+[Supplemental ZIP](restored-native-evidence.zip):888553 bytes,
+SHA256`f120509338fb0ac55480ce11929381fb1a46fecf7d3a1c8869c395335f0eba0a`.
+Its45 evidence members plus manifest include original captures, source copies,
+verification scripts/results and launch/close records. Executables/DLLs are
+excluded. The [integrity receipt](restored-native-integrity.json) records successful
+CRC and every member length/hash checks. Historical builder paths require local
+relocation. Earlier failures remain unexplained; this bounded success does not
+replace the previous evidence or establish acceptance at other window sizes.
+
+### Original trial and automated gates
+
 - [Native evidence ZIP](native-evidence.zip):2886914 bytes, SHA256`039add524eb99ce855c2dc5680eec01c1d5b7f23024025c428125b9771a8e687`.
 - [Native archive receipt](native-archive.json) and [member/relocation index](native-collection.json) preserve original paths and hashes; these paths describe the builder machine and are not peer-accessible dependencies. The ZIP carries original captures, metadata, source fixtures/oracles, checkpoint/process/helper receipts and hash-bound scripts, plus unchanged parent observations and review. Binary/cache payloads are excluded with their hashes retained.
 - [Automated evidence ZIP](automated-evidence.zip):6244436 bytes, SHA256`7706302ef7ee5d160a23ff2d562bedb8911b4e6d7ad4c941544088974b539b6e`,113 verified members. Its [index](automated-index.json), [integrity receipt](automated-integrity.json), and [scope note](automated-README.md) retain both successful and failed command records, frozen build environment, exact source inventories, compiler/dependency receipts and reproducible helper sources. Consult original per-command results rather than adding counts from different scopes.
diff --git a/harness/experiments/state-label-layout-20260924/restored-native-evidence.zip b/harness/experiments/state-label-layout-20260924/restored-native-evidence.zip
new file mode 100644
index 00000000..0d12b1b8
Binary files /dev/null and b/harness/experiments/state-label-layout-20260924/restored-native-evidence.zip differ
diff --git a/harness/experiments/state-label-layout-20260924/restored-native-integrity.json b/harness/experiments/state-label-layout-20260924/restored-native-integrity.json
new file mode 100644
index 00000000..72fbcd54
--- /dev/null
+++ b/harness/experiments/state-label-layout-20260924/restored-native-integrity.json
@@ -0,0 +1,10 @@
+{
+  "archive": "C:\\Users\\User\\Documents\\Git\\ARCADIA-Studio\\.local\\state-restored-native-20260925.zip",
+  "bytes": 888553,
+  "sha256": "f120509338fb0ac55480ce11929381fb1a46fecf7d3a1c8869c395335f0eba0a",
+  "crc_ok": true,
+  "all_member_lengths_and_sha256_verified": true,
+  "payload_member_count": 45,
+  "total_member_count": 46,
+  "manifest_sha256": "f9a6997aa4e3b6358814d57b0da60f5d235ce81ac638e139a25bc7664b341624"
+}
diff --git a/harness/experiments/state-label-layout-20260924/restored-native-observations.md b/harness/experiments/state-label-layout-20260924/restored-native-observations.md
new file mode 100644
index 00000000..d890902d
--- /dev/null
+++ b/harness/experiments/state-label-layout-20260924/restored-native-observations.md
@@ -0,0 +1,37 @@
+# #36 restored-window native observation
+
+Local-only evidence, 2026-09-25 UTC. Parent owns publication/integration. No source code edits, commits, Gitea posts, model content edits, movement, deletion, or renaming were performed.
+
+## Identity and isolation
+
+Preserved source commit `70bdcdf8fec1728e971cd90ad7f79a1df99777ff`, earlier evidence commit `57b3fe4bff43a7df602f1567b8867bb1d31051d0`. Before launch rechecked GUI SHA256 `0401341388025544185455ad5520aa2c1b79ca1387d1a57c2bb0bddb8b23754a`; confirmed adjacent libunwind.dll. `launch.json` records executable, child-only LOCALAPPDATA profile, unrelated working directory and PID2892. Copied original immutable baseline B into this evidence directory's `working/B`; original inputs were never opened in the GUI. Seeded private recent-model list with only this copy. `original-source-hashes.json` was recorded before launch.
+
+All UI input used `node_repl` plus `@oai/sky`, with one action then refresh. Raw original screenshot payloads were saved once under `raw/`, accompanied by metadata without payload duplication. Screenshots were inspected through the tool's original display, not re-emitted or transformed. No capture/input helper errors occurred.
+
+## Observed sequence
+
+The restored window remained 1442×932 throughout (screenshot dimensions), with displayed State zoom79%. No maximize, resize, pan or zoom was performed. The fixture opens directly to Behavior & scenarios, Request lifecycle, State multiselect - main.
+
+| Capture | Action and observed outcome |
+| --- | --- |
+| 01-home | Home, private B recent-model card visible. |
+| 02-open-attempt | Current observed Open clicked at(1061,574); title became State group native and Working status appeared. |
+| 03-opened | Refreshed without another input; fixture loaded, 0 selected. Ready and for review have separate title rows followed by External · Completion [request_ready]. No observed self-overlap or clipping within this title/summary. |
+| 04-title-hit | First title row clicked(652,515); inspector State transition, stable alias between. |
+| 05-clear | Blank canvas click(990,630) did **not** clear selection. Retained as an observation; not counted as reset or transition hit. |
+| 06-state-b | Click(798,596) selected State, stable alias state_b. |
+| 07-second-title | Second title row clicked(656,525); inspector State transition, stable alias between. |
+| 08-state-b | Click(798,596) again selected State, alias state_b. |
+| 09-summary | Summary clicked(661,534); inspector State transition, stable alias between. |
+| 10-save | Ctrl+S; bottom status All source files saved. |
+| 11-close-windows.json | Alt+F4, then fresh list_windows returned no Arcadia windows. |
+
+Text is small at this zoom. The authored diagram remains crowded, and this is not global collision-avoidance evidence. This new run establishes bounded successful restored-window Home Open and first-title/second-title/summary hits at1442×932/79%, not all-window acceptance. Earlier1442×842/79% and restored Home failures remain recorded in the prior review; their cause is not established or reinterpreted by this success. No exact-world-space movement claim is made.
+
+## Save and close verification
+
+`verification.json` and reproducible `verify.py` record all nine source members byte-identical to the immutable original baseline and full Project JSON byte/value-identical to pinned expected B, SHA256 `bf04d72167e8779940dec040f47c7ac0ca69cb7feea35272f2a382c3e73dceb7`. The preserved bundle_converter.exe and attachment_snapshot.exe both returned0 with their previously recorded hashes; their stderr logs are empty. GUI stdout/stderr are both zero bytes.
+
+Preserved failed verification attempt: attachment_snapshot.exe was first invoked directly on project.arcadia and rejected includes with `Error: Text("4:5: An include requires a source bundle; single-string parsing never reads files.")`; no actual.json was created then. The correct established converter-then-snapshot flow above subsequently passed. This was a verification invocation error, not a GUI/model defect.
+
+Normal Alt+F4 close was directly performed and disappearance observed. The attached PowerShell process watcher records has_exited=true in `close.json`, but ExitCode is null. Therefore this run claims observed process termination, **not** observed exit0. No warning or crash appeared; GUI logs remain empty.
diff --git a/harness/experiments/state-label-layout-20260924/restored-native-review.md b/harness/experiments/state-label-layout-20260924/restored-native-review.md
new file mode 100644
index 00000000..74f8bfcd
--- /dev/null
+++ b/harness/experiments/state-label-layout-20260924/restored-native-review.md
@@ -0,0 +1,27 @@
+# Independent review: #36 restored-window observation
+
+Accept the bounded successful observation at 1442 x 932 / 79% as additional evidence. It does not resolve the earlier 1442 x 842 failures, establish their cause, certify arbitrary restored sizes, or establish GUI exit code 0. No material contradiction found in the bounded outcome.
+
+Reviewed `.local/state-restored-native-20260925` and its sibling ZIP against original `.local/state-label-trial-70bdcdf8-20260924/baseline/B` and `expected/B.json`. Read REPORT, verification/launch/close receipts, package manifest, verification/package scripts, and raw screenshot metadata. No UI input, builds, model/source edits, publication or Gitea actions. This review file is the only write.
+
+## Independently checked
+
+- ZIP CRC test passed. Exact member set: 45 payload files plus PACKAGE-MANIFEST.json. Every payload length/SHA256 matched the manifest and current loose file bytes. Manifest bytes matched the ZIP copy and external receipt hash. ZIP is 888553 bytes; SHA256 `f120509338fb0ac55480ce11929381fb1a46fecf7d3a1c8869c395335f0eba0a`. Manifest SHA256 `f9a6997aa4e3b6358814d57b0da60f5d235ce81ac638e139a25bc7664b341624`.
+- All nine saved source members, including their relative-path set, are byte-identical to the original trial's immutable baseline B. Captured actual.json is byte-identical and JSON-value-identical to original expected/B.json; SHA256 `bf04d72167e8779940dec040f47c7ac0ca69cb7feea35272f2a382c3e73dceb7`.
+- Preserved converter and snapshot executable hashes match verification.json. GUI executable independently rehashed as `0401341388025544185455ad5520aa2c1b79ca1387d1a57c2bb0bddb8b23754a`, matching launch.json. No helpers were rerun; converter/snapshot exit results remain recorded receipts. Source-byte identity and full snapshot identity were independently recomputed.
+- All ten raw JPEGs were independently viewed without editing/cropping. Their decoded dimensions are 1442 x 932, consistent with metadata. State captures display 79% and the restored-window titlebar control. Home Open progresses from the copied B recent entry, through Working, to the loaded fixture.
+- Capture03 has no selection. Capture04 shows State transition / stable alias `between` after the first-title action. Capture05 retains that selection following the blank-canvas action, as the report explicitly states. Captures06/08 show State / `state_b`; captures07/09 return to State transition / `between`, making the second-title and summary observations nonvacuous. Capture10 shows All source files saved.
+- The transition's two title rows and following summary are vertically distinct. Text is small and the diagram remains crowded; this is not a global readability or collision-avoidance pass. Screenshot state supports reported action outcomes, while exact input coordinates/action chronology come from the operator report, not independent replay.
+- close.json says has_exited=true with exit_code=null; final window inventory is empty. GUI stdout/stderr are empty. This supports observed termination only. Alt+F4 cause is the operator's recorded action; empty logs do not prove exit0.
+
+## Scope and handoff qualifications
+
+The original source pin is reported as `70bdcdf8fec1728e971cd90ad7f79a1df99777ff`; executable-to-source provenance inherits the original trial's build/staging evidence. This review independently binds the executable bytes, not a new build.
+
+This success at 1442 x 932 cannot falsify a failure at 1442 x 842 or explain prior restored Home failures. Preserve both observations and keep #36's wider acceptance open. Source save with no model content changes proves preservation in this run, not changed-content persistence or a fresh-process reopen.
+
+The failed direct snapshot invocation is preserved narratively in REPORT; the packaged snapshot stderr is the successful rerun's empty log. Do not describe the package as containing an independently reviewed raw log of that failed attempt.
+
+The package omits executables and the original baseline/expected inputs. Publication should retain explicit links to the original trial evidence/source pin for reproducible comparison. This review makes no publication or integration claim.
+
+Audit method: read-only Python standard-library ZIP CRC/member checks, SHA256 and complete byte/JSON equality against original trial files; Pillow image dimension decoding; view_image inspection of all ten raw images; PowerShell file hashing. No production decoder or GUI rerun.