fix #485: isolate console capture and align Gitea portable test lanes

docs/ISSUES.md

@@ -39,6 +39,20 @@ confirmed closed by the owner, 11 need a focused live gate, and 43 are safe to
 remain closed. See
 [`docs/research/2026-08-28-owner-closed-issue-validity-audit.md`](research/2026-08-28-owner-closed-issue-validity-audit.md).
 
+## #485 — Gitea portable CI races console capture and selects Vulkan-only tests
+
+**Status:** DONE — correction implemented and locally verified, 2026-09-06;
+new remote run to be verified after push. Run209 on3ebb120dd built cleanly,
+but Windows App tests failed: one StringBuilder capture race, four
+ErrorIncompatibleDriver failures. Linux passed; release was skipped.
+The camera diagnostics collection now excludes other parallel collections;
+Gitea filters match the canonical portable filter (including Lane!=Vulkan).
+Two regression guards added. Release test assembly build0W/0E and corrected
+App lane7,156/0/0 PASS; no production code changed. The dedicated Vulkan lane
+remains available and strict, not a runtime skip.
+Evidence and separate alternate-output build limitation:
+`research/2026-09-01-overhaul/2026-09-06-ci-209-correction.md`.
+
 ## #484 — Distant mountain silhouettes flicker and change with viewpoint
 
 **Status:** OPEN — OWNER-DEFERRED, 2026-09-06, after report on #483 candidate.

docs/ci-and-releases.md

@@ -22,6 +22,14 @@ test can fail on them and a run costs ~7 minutes plus a 121 MB release. A red ga
 cannot publish: `release` uses `needs:`, not a `workflow_run` trigger, whose
 Forgejo support is unreliable.
 
+The streaming Windows filter must exactly match the default in
+`tools/run-release-gate.ps1`; Linux differs only by enabling `Lane=Linux`.
+`VulkanLaneOwnershipContractTests` enforces both copies. These ordinary Gitea
+runners do not promise Vulkan capability, so `Lane=Vulkan` is excluded here,
+not silently skipped inside a test. The explicit hardware lane remains in
+the GitHub `linux-vulkan` workflow and can be run locally on a capable device
+(see `release-gate.md`); a green Gitea run alone does not certify GPU pixels.
+
 ## Why Gitea and not GitHub
 
 GitHub Actions is **billing-blocked** on this account ("recent account payments

docs/research/2026-09-01-overhaul/2026-09-06-ci-209-correction.md

@@ -0,0 +1,65 @@
+# Post-closeout Gitea CI correction — #485, 2026-09-06
+
+## Evidence and cause
+
+[Run209](https://git.snakedesert.se/erik/acdream/actions/runs/209) tested
+`3ebb120dd271c873476cd071b2a33f09d17ce710`. Linux passed. Windows built
+successfully (0 warnings/errors), then App.Tests failed five cases:
+
+- `Collect_WeatherTurnFiredMatchesThePrintedOcLineExactly(true)`:
+  `StringBuilder.ToString` threw ArgumentOutOfRangeException (chunkLength).
+  The camera diagnostics collection serialized its members but still ran
+  beside other collections. Other tests could write to the process-wide
+  captured Console.Out while the test read its underlying StringWriter.
+  They could also observe the temporary global walk-transcript flag.
+- The four `Lane=Vulkan` pixel cases (ordinary shared-index, both
+  world-shader light strides, atmospheric receivers) failed at
+  `vkCreateInstance: ErrorIncompatibleDriver`. They were already correctly
+  tagged, but Gitea's copied portable filter lacked the Vulkan exclusion
+  present in `tools/run-release-gate.ps1`.
+
+Windows App result: 7,153 passed / 5 failed / 0 skipped. No other Windows
+test assembly failed. Release was skipped because it depends on both gates.
+These are test isolation and CI lane-routing defects, not a new distance
+algorithm failure.
+
+## Bounded correction
+
+Make the existing camera diagnostics collection exclusive with
+`DisableParallelization=true`. Other ordinary collections remain parallel.
+No timer, retry, assertion weakening or production change. A deterministic
+metadata regression protects exclusivity.
+
+Align the Windows Gitea filter with the canonical portable filter; both Linux
+filters retain only their intentional difference (enable Lane=Linux).
+A contract regression compares all three filters to that canonical source.
+The Vulkan tests are unchanged, still fail on missing capability when selected,
+and remain explicitly owned by the separate hardware lane. Gitea's ordinary
+runners do not certify GPU pixels; this distinction is documented.
+
+## Verification before push
+
+- Changed App test assembly Release build, reusing unchanged production
+  references (`BuildProjectReferences=false`): **PASS, 0 warnings/errors**.
+- Full App suite under the exact corrected Gitea Windows filter:
+  **7,156 passed / 0 failed / 0 skipped**. The total is run209's7,158 minus
+  four hardware-only cases plus two new regression tests, not lost coverage.
+  TRX: `artifacts/ci-209-fix-tests/ci-209-app-corrected.trx`.
+- Lead reviewed the console/global-isolation fix; independent review of
+  CI filters, ownership contract and documentation: **PASS**.
+- Running client binary SHA256 remained
+  `869BA14CB2C610E006C80BDFFCBD75629339A3486B8B98D7708B23050700FF34`.
+  Neither ACDream nor retail was closed or controlled.
+
+An attempted full solution build with alternate
+`--artifacts-path artifacts/ci-209-fix-build` failed in existing packaging
+path assumptions: smoke-plugin GetTargetPath resolves against App's absolute
+output path (MSB3030), and nested Bake publication misses its win-x64 assets
+target (NETSDK1047). That attempt is **FAIL**, not the passing test-only build
+above. No packaging workaround was added to this CI correction. The ordinary
+CI build already passed on the original commit; the subsequent remote run
+must verify the correction in CI's actual clean-build layout.
+
+This note records pre-push evidence. Remote success is reported only after
+observing the new commit's Gitea results. Campaign closeout and deferred #484
+are unchanged; no retail-behavior deviation is introduced.