docs(overhaul): contract S5 c3 fix round 1

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

@@ -1234,3 +1234,99 @@ lifecycle identity, allocation, allowed scope, build/lane provenance, and at
 least three reproduced sabotage claims. Reviews run sequentially. A failed
 lens gets one bounded fix contract; a chunk needing a third fix round stops
 and is written up.
+
+## 18. S5-c3 fix round 1 — x87 store schedule and register truth
+
+The first sequential OpenAI retail/deviation review returned **FAIL** on the
+uncommitted scratch implementation: one MAJOR exact-arithmetic defect and one
+MINOR pre-existing register-truth defect. Everything else in §17 passed. This
+is the first bounded fix round; it does not reopen selection ownership,
+building traversal, residency, Config wiring, or any S5-c1/S5-c2 behavior.
+
+The reviewer independently re-confirmed the executable SHA, CodeView GUID/age,
+all seven §17.1 byte-range hashes, the named-decomp semantics, the complete
+§17.2 census, the complete-body order, Setup part zero, selected-shell identity
+and surface preservation, no base fallback, scope, lifecycle wiring, the other
+register edits, and every required mutation. Its clean focused return was
+64/64, and the scratch state was restored exactly.
+
+### 18.1 F1 — preserve retail's x87/store boundaries
+
+The candidate currently rounds arithmetic too early in three places. Repair
+only these exact boundaries, allocation-free:
+
+1. `SceneTool::UpdateFPSCounter @0x0043E510` sums the 20 stored single-precision
+   frame durations in x87 precision and divides before narrowing once at the
+   published FPS store. Accumulating into a C# `float` is not equivalent. Use
+   a wider accumulator that exactly sums the 20 binary32 inputs, perform the
+   division wide, then narrow once to `float`. Restore retail's physical
+   history direction: shift old slots `0..18` to `1..19`, then store the
+   just-finished binary32 duration at slot 0. The prior-20-before-insert order
+   remains unchanged.
+2. `Render::CalcDegLevel @0x0054CAF0` evaluates its five membership weights,
+   weighted numerator/denominator, addition, and clamp in x87 precision except
+   at the explicit binary32 stores visible in the named function. Re-express
+   the pure calculation with wide intermediates fed by the exact binary32
+   constants (`-0.150000006f`, `-0.02f`, `0.01f`, `0.1f`) and narrow only at
+   retail's candidate store. The stability comparison promotes each binary32
+   history slot and candidate and compares against the executable's qword
+   `0.01`, not `0.01f`. Keep the existing shift-before-evaluate and final
+   current-multiplier store order.
+3. `GfxObjDegradeInfo::get_degrade @0x0051E4B0` stores the effective distance
+   after `abs(distance)-DegradeDistance`, but evaluates each positive/negative
+   threshold in x87 precision and compares the stored binary32 effective value
+   against that wide threshold. Do not round the threshold to `float` before
+   the strict comparison. Disable/force precedence, strict `<`, equality
+   advance, final slot, modes, and NaN/infinity behavior stay unchanged.
+
+Add bit-exact independent assertions for all three proven witnesses:
+
+- FPS history `[0.001f, 0.005f] × 10` publishes bits `0x43A6AAAB`
+  (`333.333344f`), not the candidate's `0x43A6AAAC`;
+- `CalculateCandidate(14f, 0f)` publishes bits `0x3BB2BC0A`, not
+  `0x3BB2BC0B`;
+- with ideal/max `24f/48f`, multiplier `0.0020020019728690386f`, and
+  effective distance `24.04804801940918f` (`0x41C06267`), retail's wide
+  threshold remains strictly greater and selects the first level.
+
+Also pin the qword stability edge directly and the slot direction without
+reflection-only self-comparison. Each of these changes needs a named mutation
+and actual first failing assertion: restore float FPS accumulation; restore
+float CalcDeg intermediates or `0.01f`; restore a rounded selector threshold;
+reverse the FPS slot direction. Restore every mutation before return.
+
+### 18.2 F2 — correct AP-198's already-live landscape radius
+
+`AP-198` is stale independently of this chunk: `Render_LandscapeDrawDistance`
+already feeds `RuntimeSettingsController`'s live far radius, and the Config
+controller maps its persisted default 8 to the selected Medium entry. Remove
+that control from AP-198, change every six-row statement to five, and delete
+the obsolete no-highlight sub-note. Keep the actual Config dim set at 17/27;
+S5-c3 still changes only Automatic Degrades, Graphics Performance, and Degrade
+Distance from store-only to live. No row is added or retired, so active table
+counts do not change. Add or update a truth pin that fails on the old six-row
+wording or the obsolete no-highlight statement.
+
+### 18.3 Scope, return, and review order
+
+Production edits are limited to `BuildingDegradeController.cs` and
+`Walk/WalkBuilding.cs`. Test edits are limited to their existing focused test
+files plus the existing Config/register truth pin if required. The only
+documentation edit in the implementation remains
+`docs/architecture/retail-divergence-register.md`. Do not change controller
+ownership/tick placement, selection/result shapes, building factory or
+transforms, portal walk/order, shell classification, residency, materials,
+surface overrides, UI wiring, RHI/shaders, DAT readers, or any other register
+row. No graphical client.
+
+Return the exact changed-file list; `git diff --check`; Release solution build
+0W/0E; the bit-exact math/selector/controller and Config/register focused
+lanes; the original S5-c3 focused lane; official hermetic lane; inclusive
+InstalledDat lane with exactly the documented global identities; and the new
+mutation ledger. Do not commit—the lead owns the reviewed landing commit under
+the owner's OpenAI-only process override.
+
+The same OpenAI retail lens narrowly re-reviews F1/F2 and regression evidence
+first. Only its PASS unlocks the sequential OpenAI architecture/production/
+gate-honesty review over the entire S5-c3 stack. A failed re-review receives
+the one remaining bounded fix round; another failure after that stops S5-c3.