diff --git a/docs/research/2026-08-12-fa4-review-mechanism.md b/docs/research/2026-08-12-fa4-review-mechanism.md
new file mode 100644
index 00000000..1bc316e3
--- /dev/null
+++ b/docs/research/2026-08-12-fa4-review-mechanism.md
@@ -0,0 +1,263 @@
+# FA4 — mechanism-faithfulness review
+
+**Slice:** Campaign FA slice FA4 — the Fellowship page fully live.
+**Targets:** `357d2032` (Runtime), `5bdd0528` (App+tests), `38f08314` (docs).
+**Reviewer lens:** mechanism-faithfulness (does the code do what retail does,
+byte-for-byte where a byte exists, and is every divergence honestly filed).
+**Method:** re-derived from the code, the lane docs
+(`2026-08-11-fa-fellowship-wire.md`, `-fa-panel-structure.md`), the
+PDB symbol table (`docs/research/named-retail/symbols.json`), and a LIVE run
+of the env-gated mount probe against the real installed DATs.
+
+---
+
+## VERDICT: APPROVE-WITH-FIXES
+
+One MUST-FIX (a doc-only register omission), three SHOULD-FIX, two NITs. No
+code-behavior REJECT: every wire-touching mechanism this slice ships is
+retail-faithful and the divergences it makes are (with the one exception
+below) honestly registered.
+
+**Strongest single piece of evidence:** the live-mount probe
+(`ACDREAM_PROBE_LIVE_MOUNT=1`, real installed DATs, not a fixture) **PASSES**
+(1/1). It confirms, end-to-end against the production mount, that the name
+field builds as `UiField`, all 11 buttons/checkboxes resolve as `UiButton`,
+the sole row template `0x21000030/0x10000281` resolves through the production
+`RowTemplateResolver` with all five checked fields at the right widget types
+(`UiText`/`UiText`/`UiMeter`/`UiMeter`/`UiMeter`), the four checkbox labels
+resolve to real retail English, the Open/Close captions resolve to
+`"Open"`/`"Close"`, and a full production-path `Bind()` emits zero
+"not found" warnings. FA4 test suites also pass under `--no-build`:
+`SocialFellowshipPageController`/`RowTemplateResolver`/`UiTemplateListBoxFlushPreservingScroll`/`CharacterOptions`
+= 75/75; `RuntimeFellowshipStateTests` = 28/28; probe = 1/1.
+
+---
+
+## What verified CLEAN (re-derived, not trusted)
+
+### D4 — the `0x00A6` gate (the highest-stakes mechanism)
+The five contracted transitions all resolve correctly. The conjunction is
+`_fellowship?.SetPageVisible(_visible && IsShowingFellowship)`
+(`SocialPanelController.cs:267-268`), driven from three sources:
+`ActivePageChanged` (ctor `:140`), `OnShown` (`:250-254`), `OnHidden`
+(`:258-262`). `SetPageVisible` sends `0x00A6` only on an actual flip
+(`SocialFellowshipPageController.cs:453-458`, `if (_pageVisible == visible) return;`).
+
+- **open-on-fellowship:** `OnShown` → `_visible=true` ∧ `IsShowingFellowship=true`
+  → `SetPageVisible(true)` → one `true` send. ✓
+- **tab-away:** `ActivePageChanged` → `_visible=true` ∧ `false`
+  → `SetPageVisible(false)` (flip from true) → one `false` send. ✓
+- **tab-back:** → `SetPageVisible(true)` → one `true` send. ✓
+- **close-while-on-fellowship:** `OnHidden` → `_visible=false`
+  → `SetPageVisible(false)` → one `false` send. ✓
+- **close-while-on-other-tab:** `_pageVisible` already `false` (we left the
+  Fellowship tab earlier), `OnHidden` → `SetPageVisible(false)` → no-op, **no
+  redundant send**. ✓
+- **switch between two non-Fellowship tabs while shown:**
+  `SetPageVisible(false)` while already `false` → no spurious send. ✓
+
+The default authored tab is Allegiance (probe: `button=0x1000028C
+page=0x10000291 default=True`), so `ActivateTabs()` at mount time
+(`RetailUiRuntime.cs:2775`, before the window is shown with `Visible=false`)
+fires `ActivePageChanged` for Allegiance while `_visible=false` → no spurious
+`0x00A6` at mount. ✓
+
+**No-send-while-disconnected** verified at the command layer: `Invoke`
+returns `RuntimeCommandStatus.Inactive` without dispatching when
+`_deactivated || _view is null || _commands is null`
+(`InteractionUiRuntimeSources.cs:181-186`); `FellowshipSetPanelOpen` just
+forwards through it (`:157-159`). **Idempotence** verified twice (the flip
+guard + the pinning test `SetPageVisible_SendsPanelOpen_OnlyOnATransition`,
+`SocialFellowshipPageControllerTests.cs:537-554`).
+
+### Leader-quit rule (FA2 hand-off not bypassed)
+FA4's Quit routes `_bindings.Quit(disband)` →
+`late.GameRuntime.FellowshipQuit` → `commands.Fellowship.Quit`
+(`InteractionUiRuntimeSources.cs:138-140`), whose adapter body sends
+`0x0290 SendFellowshipAssignNewLeader` BEFORE `0x00A3 SendFellowshipQuit`
+when `RequiresLeaderHandoffBeforeQuit` returns true
+(`DirectGameRuntimeCommandAdapter.cs:888-895`, mirrored
+`CurrentGameRuntimeCommandAdapter.cs:883`). No raw `WorldSession.Send*`
+bypass exists in the UI layer. The FA2 rule itself is correct: fires only
+when `!disband && _isInFellowship && _leaderGuid == selfGuid`, picks the
+first non-self member (`RuntimeFellowshipState.cs:292-313`). ✓
+
+### D5 — the XP-share table
+`EvenSplitPercentTable = [1.0, 0.75, 0.6, 0.55, 0.5, 0.45, 0.4, 0.35,
+0.3111111, 0.28]` (`SocialFellowshipPageController.cs:103-104`) reproduces
+lane B §7.2's byte-decoded `.rdata` floats **exactly**, including the
+out-of-range default `0.0` (`EvenSplitPercent`, `:655-656`, returns `0f` for
+`memberCount` outside `[1,10]`). Branch structure matches
+`gmFellowshipUI::UpdateFellowStats` §2.9 verbatim (`FormatStatsText`,
+`:643-653`): `!ShareXp` → `0%` (retail's literal `pct=0.0f` branch),
+even-split → table, proportional → **omitted** (level only) rather than a
+guessed formula — the honest, CLAUDE.md-compliant disposition. AD-80
+accurately states the ACE display divergence (`.3` vs `.3111111` at 9, no
+10-fellow row, wrong `1.0` out-of-range default) as ACE-vs-retail, not
+acdream-vs-retail. ✓
+
+### D6 — invite auto-response
+`HandleConfirmationRequest` intercepts type-4 (`ConfirmationType.Fellowship=4`,
+`GameEvents.cs:580`) BEFORE the generic `GameplayConfirmationController`
+(`RetailUiRuntime.cs:677-683`). `TryAutoRespondToFellowshipInvite` (`:685-693`)
+reads both bits and sends `SendConfirmationResponse(type, ctx, autoAccept)`
+(the shared confirmation-response primitive,
+`InteractionRetainedUiComposition.cs:807-812`) — decline when only Ignore is
+set, accept when AutoAccept is set. The mutual exclusion IS Runtime's:
+`RuntimeCharacterState.TrySetOption` clears the opposite bit when either is
+turned on (`:821-834`), so both-set is unreachable through the write path.
+Neither bit set → `return false` → falls through unchanged. ✓
+
+### D7 — un-dims
+Exactly four rows flip `StoreOnly`→`Live`
+(`CharacterOptionsPageController.cs:206-208`): `IgnoreFellowshipRequests`,
+`FellowshipAutoAcceptRequests`, `FellowshipShareXP`, `FellowshipShareLoot`.
+Each has a live read site: the first two in
+`TryAutoRespondToFellowshipInvite`; `ShareXP` at the Create-button click
+(`SocialFellowshipPageController.cs:325`) and its own checkbox;
+`ShareLoot` via its own live checkbox surface (`RefreshCheckboxSelections`
+`:484-486`). `IgnoreAllegianceRequests`/`DisplayAllegianceLogonNotifications`
+stay `StoreOnly` (`:206`/`:208`). Conformance test updated 35→31 store-only /
+15→19 live with the exact four ids removed from `ExpectedStoreOnlyIds`
+(`CharacterOptionsPageControllerTests.cs:729-733`, `:762-765`). ✓
+
+### Roster scroll preservation
+`FlushPreservingScroll` (`UiTemplateListBox.cs:290-296`) saves `ScrollY`,
+`ClearContent()`s, restores `ScrollY`. It genuinely preserves across a
+membership-change rebuild because `ClearContent` zeroes only the *panel's*
+`ContentHeight` and never syncs `Scroll.ContentHeight`
+(`UiScrollablePanel.cs:54-61`), so the restore's `SetScrollY` clamps against
+the STALE (pre-rebuild) `MaxScroll` (`UiScrollable.cs:56`) — exactly as the
+method's own doc claims. The next `LayoutScrollableChildren` re-clamps
+against the fresh height. Tests exercise the real cases: preserve across
+same-count rebuild, clamp-down on shorter content, dormant no-op
+(`UiTemplateListBoxFlushPreservingScrollTests.cs:50-98`). The no-change path
+(`UpdateRow`) only sets `LinesProvider`/meter delegates — no ListBox
+mutation (`SocialFellowshipPageController.cs:607-627`). ✓
+
+### Create-flow refusal-by-enabled-state
+`RefreshCreateButtonState` sets `Enabled = !IsNullOrWhiteSpace(name)`
+(`:460-465`); the click handler re-guards (`:322-327`). Matches lane B §2.2
+(`gmFellowshipUI::CreateFellowship` @0x0048F730): the button IS the guard,
+no separate error text. `ShareXP` read from the character option at click
+time (lane B row 67: "shareXP comes from option 0x0F"), not a dialog field. ✓
+
+### Button-enable rules (§2.8 verbatim)
+`RefreshButtonStates` (`:510-537`) ports the `gmFellowshipUI::UpdateButtons`
+@0x0048E6C0 table exactly: Quit always; Disband/Open `= isLeader`;
+Leader/Dismiss `= isLeader && hasSelection && !selectedIsSelf`; Recruit
+`= targetValid && notFull && (isLeader || IsOpen)`. The one gap (Recruit does
+not gate on "target is a player") is honestly disclosed as a
+superset-of-retail rule in an inline comment (`:529-534`) with the server
+refusing a non-player target the same way retail's click-handler no-ops.
+
+### §6 compliance / anchors
+String resolver threaded to every `LayoutImporter.Build`
+(`RetailUiRuntime.cs:2714-2720`, the `RowTemplateResolver` build closure);
+scoped `FindDescendant(pageRoot,…)`/`FindDescendant(row,…)` throughout;
+scrollbar id-0 guard (`:247-256`); `RowTemplateResolver` caches null misses
+so a permanently-unresolvable template imports ONCE (`RowTemplateResolver.cs:63-72`),
+and the revision latch advances before `RefreshRoster`
+(`SocialFellowshipPageController.cs:433-437`) so a partial resolver failure
+never becomes a per-frame retry. Five retail anchors byte-verified in the
+PDB: `0x0048E6C0`→`gmFellowshipUI::UpdateButtons`,
+`0x0048F730`→`::CreateFellowship`, `0x005B9070`→`Fellowship::IsFull`,
+`0x005B9BA0`→`FellowshipSystem::GetEvenSplitXPPctg`,
+`0x004901C0`→`gmFellowshipUI::ListenToElementMessage`,
+`0x0048EB20`→`::UpdateFellowStats`, `0x006A6380`→`CM_Fellowship::Event_UpdateRequest`
+— all match their citations. `SelectionChangeSource.Social`
+(`SelectionState.cs:21-26`) is anchored to `::ListenToElementMessage`'s
+`SetSelectedObject` call and matches the shared selection primitive. Gate
+script §FA4 claims all trace to the code (create gating, caption swap, button
+states, D5 display, invite bits, leader hand-off). ✓
+
+---
+
+## Findings
+
+### MUST-FIX 1 — the leader-gold-tint adaptation has no divergence-register row
+`SocialFellowshipPageController` tints the fellowship LEADER's roster name a
+distinct gold (`LeaderNameColor = (1, 0.84, 0, 1)`, applied in `UpdateRow`,
+`SocialFellowshipPageController.cs:110`/`:614`). Its own class doc
+(`:57-65`) is honest that lane A's row template has **no** dedicated leader
+marker and retail "may simply not mark leadership in the row at all" — i.e.
+this is an *invented, user-visible visual cue retail (probably) does not
+have*. That is precisely the class of deviation the divergence register
+exists to hold, and CLAUDE.md's binding rule is unambiguous: "Any commit that
+introduces a deviation … adds its register row IN THE SAME COMMIT" / "A
+deviation found without a row is a bug twice over." The FA4 docs commit filed
+AD-80 (XP-share display) and AD-81 (the numeric-composite text gap) but
+**neither covers the gold tint** — a register grep for `gold`/`tint` returns
+only the unrelated AP-178 (SpewBox). This is exactly the same reasoning that
+earned AD-78 its row ("caption colors retail never has → register row"). Fix:
+add a one-line AD row (or fold a clause into AD-81) recording the leader
+gold-tint adaptation and citing `SocialFellowshipPageController.cs:110`/`:614`.
+*(Doc-only; the code behavior is a reasonable, honestly-code-documented
+adaptation — the defect is purely the missing auditable row.)*
+
+### SHOULD-FIX 2 — AD-78's "35 of 50 rows dimmed" parenthetical is now stale
+`38f08314`'s message claims "AD-78's derivation table gains its D7 addendum,"
+but the addendum landed in the CODE class-doc (in `5bdd0528`), while the
+**register row itself** was not touched: AD-78's "Where" column still reads
+`CharacterOptionsPageController.cs (35 of 50 rows dimmed …)`
+(`retail-divergence-register.md:175`) when the post-FA4 count is **31 of 50**.
+The register is "the single auditable list"; a stale count in an active row
+is a bookkeeping drift. Fix: update `35`→`31` in AD-78's Where column.
+
+### SHOULD-FIX 3 — D4 is not re-armed across a generation reset (reconnect)
+`0x00A6` is a per-session, server-latched declaration; `_pageVisible`
+(App-side UI state) is NOT reset by a generation reset, and
+`ResetSessionTransientUi` (`RetailUiRuntime.cs:731-736`) does not hide/close
+the social panel. So if the Fellowship tab is open across a reconnect,
+`_pageVisible` stays `true`, no `OnShown`/`OnHidden`/`ActivePageChanged`
+transition fires, `SetPageVisible` is never re-called, and the fresh session
+never receives `0x00A6`. Since `RuntimeFellowshipState.ResetSession` clears
+the roster and ACE only sends `0x02BE`/`0x02C0` while `FellowshipPanelOpen`
+(gated by `0x00A6`, lane B §4.5/row 78), the roster can present empty/frozen
+after reconnect until the user manually toggles the tab. This is the exact
+"missed true-send freezes the roster" failure the D4 contract warns about,
+just on the reconnect edge (outside the five in-session transitions, which
+are all correct). Recommend re-declaring on session reset (drive
+`SetPageVisible` from the reset hook, or clear `_pageVisible` so the next
+`Tick`/transition re-arms) — and confirm ACE's world-enter re-send behavior
+before the connected gate closes. *(I could not establish ACE's exact
+world-enter `0x02BE` behavior from the client decomp; flag it for the gate.)*
+
+### SHOULD-FIX 4 — no panel-level test pins the D4 conjunction
+The controller-level idempotence IS tested
+(`SetPageVisible_SendsPanelOpen_OnlyOnATransition`), but the panel-level
+wiring in `SocialPanelController` — `OnShown`/`OnHidden`/`SwitchTo` ×
+active-tab → the correct `SetPanelOpen` calls, including the "close on other
+tab sends nothing" and "tab-away sends false" cases — has **no** direct
+assertion. The callbacks record `fellowship-set-panel-open:{bool}`
+(`SocialPanelControllerTests.cs:51`) but nothing asserts on it. For the
+highest-stakes mechanism, add a transition test that opens on Fellowship,
+tabs away/back, and closes on each tab, asserting the exact
+`set-panel-open` call sequence.
+
+### NIT 5 — Open/Close caption does not optimistically pre-toggle
+Lane B row 77 notes retail "pre-toggles its own `_open_fellow`" on the Open
+button click, i.e. the caption flips immediately client-side, then the server
+confirms. acdream's `RefreshOpenCaption` keys the caption off the
+server-authoritative `snapshot.IsOpen` (`:495-508`), so the caption flips
+only when the `0x02BE` echo lands. Functionally convergent (and instant on
+localhost), but not byte-faithful to retail's optimistic local pre-toggle.
+
+### NIT 6 — `MathF.Round` vs retail `_ftol2` truncation (non-issue, documented)
+`FormatStatsText` uses `(int)MathF.Round(pct*100f)` (`:650`) where retail's
+`UpdateFellowStats` truncates via `_ftol2`. For every value in the even-split
+table (×100 = 100/75/60/55/50/45/40/35/31.11/28) round and truncation agree,
+so there is no observable difference on any reachable input. Worth a one-word
+note that the choice is deliberate/equivalent; no change required.
+
+---
+
+## Coverage confirmation
+- Live-mount probe (real DATs): **PASS 1/1** — the FA4 fellowship assertions
+  (name field `UiField`, 11 controls `UiButton`, row template + 5 fields,
+  labels/captions, zero-warning `Bind()`) all hold against the production mount.
+- `--no-build` targeted suites: App FA4 = 75/75, Runtime fellowship = 28/28.
+- The structural finding in the ledger (Create-flow controls are children of
+  the not-in-fellowship frame `0x1000026B`) is confirmed by the probe dump:
+  `0x1000026B` children include `0x10000270/71/72/73` (all four checkboxes),
+  `0x1000026F` (name field), `0x10000274` (Create).