diff --git a/docs/ISSUES.md b/docs/ISSUES.md
index 1bfb34ee..9da6482f 100644
--- a/docs/ISSUES.md
+++ b/docs/ISSUES.md
@@ -53,6 +53,77 @@ parameters in the viewport render setup), port the roll drift, express
 the exit as the view-plane distance animation. Retail-faithfulness fix —
 no tuning knobs.
 
+**TWO FIX ATTEMPTS DISCARDED BY USER DIRECTION (2026-08-17 evening;
+commits removed from the branch, patches preserved in the session
+scratchpad `tunnel-419-patches/`). Read this record before the next
+attempt.**
+
+*Attempt 1 (was `9e83a49b`) — scoped winding mirror.* Theory: retail
+draws the tunnel's `sides_type 0` polys with `D3DCULL_CW` through an
+improper det−1 view basis; our proper basis flips screen winding, so the
+inherited cull mapping keeps the mirror-image face set. Changed:
+`WbDrawDispatcher.cs` (+`PushMirroredAuthoredWinding` scope +
+`MirrorAuthoredCullMode` + one `EffectiveCullMode` chokepoint),
+`WbDrawDispatcher.PackedOracle.cs`, `PortalTunnelPresentation.cs` (draw
+inside the scope), `TeleportViewPlaneController.cs` (citations),
+`PortalTunnelInteriorTests.cs` (new). USER GATE RESULT: exit flash gone
+BUT the tunnel showed as a dark tube seen from OUTSIDE — the mirror
+culled the interior; the "fixed" exit was a FALSE POSITIVE (nothing
+drawn cannot flash).
+
+*Attempt 2 (was `213345bd`) — revert attempt 1 + Surface.Luminosity
+emissive port.* The revert's justification (verified, durable):
+`MeshExtractor.BuildPolygonIndices` emits positive-surface fans in
+REVERSED index order — WB's baked D3D↔GL compensation — so
+(bake reversal) ∘ (proper det+1 look-at) ≡ retail's
+(authored order) ∘ (improper det−1 + `D3DCULL_CW`): the DEFAULT cull
+mapping already keeps retail's faces and attempt 1 double-mirrored it.
+Theory 2: the rim/flash = ambient-only dark facets because the swirl
+surface `0x08000C31` authors `Surface.Luminosity = 1.0` (retail's
+material emissive — `FUN_0059DA60` writes `Material.Emissive.rgb =
+Luminosity`; the sky shader ports this law, the OBJECT path never did).
+Changed: luminosity plumbing extractor→`WbMeshAdapter` startup
+luminous-surface map→`ObjectRenderBatch`/`InstanceGroup`/`CachedBatch`/
+`PackedClassifiedBatch`→`BatchData.Reserved` float bits→`mesh_modern`
+shader term (recompiled SPIR-V, 13 .spv files), plus
+`PortalTunnelPresentation.cs`. USER GATE RESULT: **tunnel looked
+IDENTICAL to baseline** — the rim and exit flash are back (expected,
+attempt 1 reverted) and the luminosity change produced NO VISIBLE
+DIFFERENCE, meaning either theory 2 is wrong or the plumbing never
+reached the tunnel's draw (UNRESOLVED — never verified live; the
+per-batch value could be defaulting to 0 anywhere along the six-hop
+chain).
+
+*Durable byte-decoded facts (PDB-paired binary, keep):* roll drift
+`duration = RandDouble(0.6, 1.8)`, `endAngle = RandDouble(0.0, 360.0)`,
+interp `start + (end−start)·GetAnimLevel/1024`, applied
+`SetCameraDirection_Degrees(0, angle, 0)` @0x004D6FD6–704E;
+`TRANSITION_VIEW_PLANE_DISTANCE @0x007BD260 = 0.001` (double),
+`TELEPORT_ANIM_FADE_TIME @0x007BD278 = 1.0`, MIN/MAX_CONTINUE
+@0x007BD268/70 = 2.0/5.0, `TELEPORT_ANIM_FPS @0x007BD280 = 40.0f`;
+fade `dist = (TVPD − gameVDist)·level/1024 + gameVDist`;
+`set_vdst @0x0054B240`: `fov = 2·atan(1/d)` clamp (0.001, π),
+`znear = d<0.4 ? 0.1 : d·0.25`. CreatureMode ambient (0.3,0.3,0.3)
+@0x004543CF (ours matches), distant light 2.0 dir (0.3,−1.9,0.65)
+(matches). The original filing's "own fixed projection" premise is
+REFUTED: `gmSmartBoxUI::PostInit @0x004D6DB3` calls
+`CreatureMode::UseSmartboxFOV` — the tunnel FOLLOWING the Config FOV is
+retail. Camera IS at the authored eye (three-way enclosure proof: 0
+escape rays at all 120 frames, live-pak batches identical to fresh
+extraction). Tunnel = Setup `0x02000306`, two half-torus parts of GfxObj
+`0x0100080B` (360 polys `sides_type 0` + 4 unstippled pos-surface quads
+— the purple floaters are authored content).
+
+**MANDATORY NEXT PROTOCOL (agreed with the user after two failed
+rounds): apparatus BEFORE any further fix.** (1) a probe-gated tunnel
+FREEZE (`ACDREAM_PROBE_*` family) holding TAS_TUNNEL indefinitely so
+the scene is statically inspectable; (2) ONE RenderDoc capture of the
+frozen frame — which triangles draw, cull state, material constants
+actually bound; (3) the ACViewer oracle on Setup `0x02000306` (clean
+interior there ⇒ our pipeline, rings there ⇒ shared dat
+interpretation); (4) a retail side-by-side screenshot for ground truth
+(brightness included). Fix only against that evidence.
+
 ## #418 — Login world load takes ~27 s: publication advances at a flat 32 blocks/s
 
 **Status:** IN-PROGRESS 2026-08-17 — producer half landed (this commit's