docs(overhaul): contract S5 c3 fix round 2

docs/research/2026-09-01-overhaul/s5-consumers-material-closeout-packet.md

@@ -1,8 +1,8 @@
 # Campaign OVERHAUL v2 — S5 consumers, material, and closeout packet
 
 **Status:** S5-c1 and S5-c2 LANDED + lead graphical gates PROVISIONAL PASS;
-S5-c3 implemented in scratch, retail review 1 FAIL, bounded fix round 1 in
-progress under §18.
+S5-c3 implemented in scratch; retail review passes 1 and 2 FAIL, and bounded
+fix round 2 is contracted under §19.
 **Branch:** `claude/campaign-w-retail-frame-walk`.
 **Gate:** G4 remains unpassed. Nothing merges to `main` before G4.
 
@@ -1335,3 +1335,57 @@ first. Only its PASS unlocks the sequential OpenAI architecture/production/
 gate-honesty review over the entire S5-c3 stack. Each further code finding
 receives its own bounded contract; S5-c3 stops before an eleventh review pass.
 Documentation-only findings may be corrected directly by the lead.
+
+## 19. S5-c3 fix round 2 — CalcDegLevel's four intermediate stores
+
+OpenAI retail review pass 2/10 returned **FAIL** on one remaining MAJOR F1
+defect. Fix round 1 otherwise passed: the FPS witness is `0x43A6AAAB`, the
+14-FPS candidate is `0x3BB2BC0A`, the selector boundary stays at level zero,
+the qword `0.01` edge and physical FPS slot direction pass, and AP-198 now
+states the exact five-row residual while AD-78 remains 17/27. The reviewer
+also re-confirmed the paired executable/PDB identity, clean allowed scope,
+the exact installed census, and the official evidence identities.
+
+### 19.1 R2-F1 — reproduce every executable-visible numerator store
+
+The remaining defect is only
+`BuildingDegradeController.CalculateCandidate`. Its combined
+`float weightedNumerator` narrows once after all five terms. The paired
+`Render::CalcDegLevel @0x0054CAF0` bytes instead store the evolving numerator
+to binary32 four times: after w0 at `0x0054CC14`, after the negative w1 term
+at `0x0054CC61`, after the zero-weight w2 term at `0x0054CCA6`, and after the
+positive w3 term at `0x0054CCEB`. Retail then promotes that fourth stored
+value, adds the w4 term wide at `0x0054CD3F`, divides wide, adds/clamps wide,
+and narrows only at the final candidate store.
+
+Implement that literal schedule allocation-free. Each of the first four
+stages must be an explicit `float` store followed by promotion for the next
+stage; do not algebraically combine, reassociate, or eliminate the `+0*w2`
+store. Feed every term with the existing exact binary32 coefficient. Keep the
+five membership weights, denominator, history/stability behavior, clamp,
+exceptional-value behavior, and every other fix-round-1 change untouched.
+
+Add an independent bit-exact assertion:
+
+- `CalculateCandidate(16.25f, 0f)` must publish `0x3CA3D70A`; the current
+  combined store publishes `0x3CA3D70B`.
+
+Sabotage by recombining the five terms into one store and record this witness
+as the first failure, then restore it. Re-run the three §18 witnesses, qword
+edge, FPS slot-direction pin, complete controller/selector lane, original
+S5-c3 focused surface, allocation lane, official hermetic lane, and inclusive
+InstalledDat lane with the same documented nonpassing identities.
+
+### 19.2 Scope and review order
+
+Production scope is exactly
+`src/AcDream.App/Rendering/BuildingDegradeController.cs`; test scope is exactly
+`tests/AcDream.App.Tests/Rendering/BuildingDegradeControllerTests.cs`. No
+other production, test, register, packet, plan, UI, renderer, RHI, shader, DAT,
+or lifecycle behavior changes. No graphical client and no commit; the lead
+owns landing.
+
+Return the exact two-file diff, `git diff --check`, Release solution build
+0W/0E, every lane above, and the mutation's first failure. The same OpenAI
+retail lens performs narrow review pass 3/10 over §19 first. Only its PASS
+unlocks the sequential full architecture/production/gate-honesty review.