chore(memory-bank): correct the cause of the effective-access failures and record the second review

The failures of the effective-access tests in the Windows Server 2022
cell are not a defect of the module and not a Kerberos S4U staleness of
the fixture: a replay with the baseline and the final candidate
alternating failed both, and the remote authorization managers answer as
if the account had no groups while the local manager and a Kerberos
logon are right. The Windows mechanism is unknown.

The notes say that fdd7a8b reverts cleanly while 962887a conflicts with
it, that the three fixes are in two commits, and that Decision 24 is in
the index of the patterns.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: AI Assistant <ai@example.com>

.memory-bank/activeContext.md

@@ -14,15 +14,15 @@ handoffs and to decide and report later. Handoff 1 (paths) is draft #118
 (`83149ee`, CI green; stacked on #117 and #116, all open; rc6 is the latest
 published candidate, 4.2.6 the stable Gallery version). Handoff 2 (operating-system
 matrix) is the local branch `ai/quality-gate-lab-matrix`, stacked on #118 and not
-pushed: the lab `NtfsSecurityOsMatrixLab`, three fixes of the module that the
-matrix found (`962887a`, `fdd7a8b`), the kit, the controller changes, and the
+pushed: the lab `NtfsSecurityOsMatrixLab`, three fixes of the module in two commits
+that the matrix found (`962887a`, `fdd7a8b`), the kit, the controller changes, and the
 record `Tests/Lab/Acceptance-2026-10-10-os-matrix.md` (Decision 24, proposed).
 The final local candidate `fdd7a8b` passes the module's suite on five operating
 systems and the host (24 runs, no failure) and the live controller in three
 cells (1,374 passed, 0 failed, 12 skipped). Handoff 3: Decision 22 was confirmed
 under the delegation and stays proposed; nothing is published. Handoff 4:
 Decision 23 (the #34 dossier); the risk acceptance is the maintainer's. The
-agent's decisions of the night are D1 to D26 in
+agent's decisions of the night are D1 to D37 in
 `decisions-night-2026-10-09.md` of the session files. Stable 5.0.0 stays gated.
 
 The earlier state of handoff 1, from the reviewed head `f11ff41` of #117: 28
@@ -121,20 +121,27 @@ open.
   with CSV tables): the suite of the final candidate `fdd7a8b` on OSFile19,
   OSFile22, OSFile25, OSWin11E, OSWin11, and the host, four configurations each,
   zero failures, skipped tests identical to the host's; the baseline `83149ee`
-  fails 4 elevated and 20 basic-user tests on the domain machines. Live: run
+  (run on OSFile22 and OSFile25) fails 4 elevated and 20 basic-user tests. Live: run
   `rc7l`, three cells, 1,374 passed, 0 failed, 12 skipped. The Admin-role
-  effective-access failures of the earlier cells came from re-creating the
-  accounts under the same names (stale Kerberos S4U state, probe in the record),
-  not from the module; the controller now names the account of case 3 anew for
-  each fixture (`1dec389`). Reviewed by the built-in code-review agent (custom
-  `security-reviewer` unavailable): approve with Minor, fixed.
+  effective-access failures of the earlier cells were not the module: in a replay
+  (`ab0` to `ab6`) the baseline failed two of three cells and the final candidate
+  one of three (not counting the warm-up `ab0`), and one model (the remote
+  authorization managers answer for an account name for about ten minutes after
+  the account was created again) fits all 43 Admin-role runs of 27 cells; the
+  Windows mechanism is unknown. The controller now names the account of case 3
+  anew for each fixture (`1dec389`); four more cells with it (`ab7` to `ab10`)
+  passed, two of them where the model predicts a failure for a reused name.
+  Reviewed by the built-in code-review agent (custom `security-reviewer`
+  unavailable): approve with Minor, fixed; a second review found one Major
+  (record accuracy), addressed by the replay.
 
 ## Next step
 
 1. The maintainer pushes `ai/quality-gate-lab-matrix` as a draft PR with base
-   `ai/quality-gate-paths` and decides which of the three module fixes belong to
-   rc7 (each is its own commit). He reviews and integrates the stack: #116, then
-   #117, then #118, then the matrix branch (Decision 24).
+   `ai/quality-gate-paths` and decides which of the module fixes belong to rc7
+   (two commits: `962887a` holds two fixes, `fdd7a8b` one). He reviews and
+   integrates the stack: #116, then #117, then #118, then the matrix branch
+   (Decision 24).
 2. He decides the open items listed in the paths report: `FileSecurity`
    conversions, `RemoveAll` account filters, lazy path overloads, abandoned
    `PrivilegeEnabler`, dot patterns of `Get-ChildItem2 -Filter`, the 17

.memory-bank/decisions/0024-os-matrix-lab.md

@@ -52,9 +52,9 @@ source: agent decisions under the maintainer's delegation of 2026-10-09 (Handoff
   5. The module's own behavior tests run on every machine as well
      (`Run-MatrixLocalSuite.ps1`), elevated and as a basic user, in both
      editions, as scheduled tasks so that the token matches a CI runner. This
-     found three defects of the module, each fixed in its own commit on
+     found three defects of the module, fixed in two commits on
      `ai/quality-gate-lab-matrix`: `Get-NTFSInheritance -SecurityDescriptor`
-     and `Get-NTFSEffectiveAccess -ServerName ''` (`962887a`), and
+     and `Get-NTFSEffectiveAccess -ServerName ''` (`962887a`, two fixes), and
      `Get-NTFSEffectiveAccess` for a user who isn't an administrator on a
      computer in a domain (`fdd7a8b`, with a live test for the ServerAdmin
      role). The maintainer decides which of them belong to rc7.
@@ -112,16 +112,26 @@ source: agent decisions under the maintainer's delegation of 2026-10-09 (Handoff
   - A profile of an account that a probe's scheduled task used stayed loaded on
     one server until it restarted; the probe now uses a new account name for
     every run.
-  - The fixture of the live controller deleted its accounts after a cell and
-    created them again with the same names for the next. Windows then returns
-    the SID and the groups of the deleted account for a Kerberos S4U logon on
-    the domain controller and the file server for more than seven minutes, so
-    `Get-NTFSEffectiveAccess` returned no access for the new account in some
-    cells (Windows Server 2022, Admin role), for the baseline and the final
-    candidate alike. This looked like a regression of the module until a loop
-    probe showed both builds failing the same way. The controller now gives a
-    new fixture a new name for the account of case 3 (`1dec389`); the record
-    has the evidence.
+  - The matrix cells of the live controller failed in the effective-access tests
+    of the Admin role in the Windows Server 2022 cell (`rc7f`, `rc7h`, `rc7i`,
+    and `rc7j`) in cells that followed each other, where the fixture was removed
+    after a cell and created again with the same account names. This looked like
+    a regression of the module (the audit read of `962887a` was the first suspect)
+    until a replay of the same cells with the baseline and the final candidate
+    alternating (`ab0` to `ab6`) failed the baseline in two of three cells and
+    the final candidate in one of three (not counting the warm-up `ab0`). In a
+    failing cell the remote
+    authorization managers (the client's and the file server's) returned no
+    groups for the current account while the Kerberos S4U logon, the name
+    resolution, and the local manager were right in the same second. One model,
+    in which a remote manager answers for an account name for about ten minutes
+    after its first request, fits all 43 Admin-role runs of 27 cells; the
+    predictions that I wrote down before three of the replay cells held (the
+    weakest is `ab6`, 10.5 minutes after its entry, above the lifetimes that
+    fit). The mechanism in Windows isn't known. The controller now gives a new
+    fixture a new name for the account of case 3 (`1dec389`); four more cells
+    with it (`ab7` to `ab10`) passed, two of them at positions where the model
+    predicts a failure for a reused name. The record has the evidence.
 - Result: [the record](../../Tests/Lab/Acceptance-2026-10-10-os-matrix.md). The
   final candidate (`fdd7a8b`) passes the module's suite on all five machines
   and the host in all four configurations, and the live cells (see the record).
@@ -130,7 +140,9 @@ source: agent decisions under the maintainer's delegation of 2026-10-09 (Handoff
   needs the published package in every cell (Handoff 3, stage D). The newest
   Windows 11 build that can join a Server 2025 domain here is 22H2; a domain
   cell with 26H1 needs a newer domain controller build or a fix of the
-  mismatch. The maintainer also decides which of the three module fixes belong
-  to rc7 (each is its own commit, `git revert` removes it), and whether the
+  mismatch. The maintainer also decides which of the module fixes belong to
+  rc7 (two commits: `962887a` holds two fixes, `fdd7a8b` one; `fdd7a8b` reverts
+  cleanly on its own, `962887a` conflicts with it in `Lib.cs` and `CHANGELOG.md`
+  if `fdd7a8b` stays), and whether the
   evaluation client stays (it needs a start shortly before every run) or is
   replaced by a client with a license that doesn't expire.

.memory-bank/deployment-notes.md

@@ -18,8 +18,10 @@ order, each with a merge commit (Decision 15), ends in exactly the tree of
 lists the versions up to rc6, as it must before rc7 is published.
 
 The branch `ai/quality-gate-lab-matrix` (local until the maintainer pushes it)
-is stacked on #118. It holds three fixes of the module (`962887a`, `fdd7a8b`;
-each is its own commit and `git revert` removes it), the kit of the
+is stacked on #118. It holds three fixes of the module in two commits (`962887a`
+has two, `fdd7a8b` one). `fdd7a8b` reverts cleanly on its own; `962887a` doesn't
+revert while `fdd7a8b` stays (the two conflict in `Security2/Win32/Lib.cs` and
+`CHANGELOG.md`), and its two fixes go together. The branch also holds the kit of the
 operating-system matrix, the changes of the live controller, and the record
 (Decision 24). rc7 contains the module fixes only if the branch is merged after
 #118 and before the tag; otherwise they go to the next prerelease. The
@@ -124,12 +126,16 @@ Local `-ModulePath` runs are validation; the gate needs the published bytes.
   `net localgroup <name>` and `net localgroup <name> <SID> /delete`. The SIDs
   of a deleted account can't be found afterwards, so keep `fixture-sids.json`
   from before the removal of the organizational unit.
-- An account that is deleted and created again with the same name keeps its old
-  SID and groups in Kerberos S4U logons on the domain controller and member
-  servers for more than seven minutes, and nothing flushes it (see
-  `techContext.md`). The controller names the account of case 3 anew for each
-  new fixture; a script of your own that recreates accounts needs unique names
-  too.
+- An account that is deleted and created again with the same name made the
+  remote authorization managers (the client's and the file server's) answer for
+  about ten minutes as if it had no groups, so `Get-NTFSEffectiveAccess` returned
+  no access; when the accounts are created again within seconds, the Kerberos S4U
+  logons returned the old account on the domain controller and member servers for
+  7 to 15 minutes. The five remedies tried (a ticket purge, `nltest /sc_reset`, a
+  DNS flush, a restart of the Kerberos service, and waiting) helped only by
+  waiting (see `techContext.md`). The controller names the account of case 3 anew
+  for each new fixture; a script of your own that recreates accounts needs unique
+  names too.
 - Restart the evaluation client (`OSWin11E`) right before a sequence or a suite,
   not before several: it shuts down an hour after each start. The restart takes
   about two and a half minutes and may need the repair of the secure channel.

.memory-bank/progress.md

@@ -83,7 +83,7 @@ After 5.0.0, archive in favor of WindowsAccessControl (Decision 18).
   148 green on the candidate; fixture removed and verified clean on six
   machines. Record: `Tests/Lab/Acceptance-2026-10-09-quality-gate-paths.md`.
 - 2026-10-09 to 10: handoffs 2 to 4 under the maintainer's delegation (decisions
-  D1 to D26 in the night log of the session files). The matrix lab
+  D1 to D37 in the night log of the session files). The matrix lab
   `NtfsSecurityOsMatrixLab` (Server 2019, 2022, and 2025 file servers, Windows 11
   Enterprise 22H2 client, Windows 11 26H1 suite only) found three defects of the
   module, fixed in `962887a` and `fdd7a8b`: audit inheritance by descriptor,
@@ -92,10 +92,16 @@ After 5.0.0, archive in favor of WindowsAccessControl (Decision 18).
   module's suite on every machine (24 runs, no failure) and the live controller
   in three cells (1,374 passed, 0 failed, 12 skipped). The failures of the
   effective-access tests in the Server 2022 cell were not a defect of the module:
-  Windows returns the SID and the groups of a deleted account for a Kerberos S4U
-  logon for more than seven minutes, and the controller names the account of
-  case 3 anew for each fixture (`1dec389`). A read-only built-in review of the
-  kit and the fixes approved with Minor findings, fixed in `db04ef2`. Record:
+  in a replay of the same cells the baseline failed two of three and the final
+  candidate one of three (not counting the warm-up cell), and one model (the
+  remote authorization managers answer for an account name for about ten minutes
+  after the account was created again) fits all 43 Admin-role runs of 27 cells;
+  the Windows mechanism is unknown. The controller
+  names the account of case 3 anew for each fixture (`1dec389`). A read-only
+  built-in review of the kit and the fixes approved with Minor findings, fixed in
+  `db04ef2`. A second review of the later commits found one Major (the record
+  called the cause settled without a baseline replay), addressed by the replay,
+  `9344ff7`, and `ab0d8e1`. Record:
   `Tests/Lab/Acceptance-2026-10-10-os-matrix.md`; nothing was pushed.
 
 ## Stable capabilities

.memory-bank/systemPatterns.md

@@ -48,6 +48,7 @@ Read only task-relevant records; the index controls routing.
 | 21 | [A quality gate before 5.0.0](decisions/0021-quality-gate-before-5.0.0.md) |
 | 22 | [The behavior changes of Phase 2 (proposed)](decisions/0022-phase-2-behavior-changes.md) |
 | 23 | [Non-Windows file servers before 5.0.0, #34 (proposed)](decisions/0023-non-windows-file-servers.md) |
+| 24 | [The operating-system matrix lab (proposed)](decisions/0024-os-matrix-lab.md) |
 
 ## Patterns
 
@@ -120,15 +121,16 @@ Read only task-relevant records; the index controls routing.
   as a basic user before a release, and classify a failure by a probe under the
   real tokens (elevated, filtered, local standard, domain standard) before
   calling it a defect or a design.
-- A fixture that deletes an account and creates it again with the same name
-  gets the old SID and groups from Kerberos S4U logons (the oracle of the live
-  tests, and the Authz functions behind `Get-NTFSEffectiveAccess`) on the
-  domain controller and on member servers for more than seven minutes, and no
-  cache flush helped (a ticket purge renews only the session's own token). A
-  failure that follows the order of the cells and not the version of the
-  module points to such state: run a loop probe with the baseline and the
-  candidate side by side before blaming the code. The controller names the
-  account of case 3 anew for each new fixture.
+- A fixture that deletes an account and creates it again with the same name can
+  get a wrong answer for about ten minutes: the remote authorization managers
+  answered `0x100000` for the current SID while the local manager and a Kerberos
+  logon were right in the same second, and, when the accounts are created again
+  within seconds, the Kerberos S4U logons returned the old account (7 to 15
+  minutes). A failure that follows the order of the cells and not the version of
+  the module points to such state: run the baseline and the candidate in cells
+  that follow each other and alternate them (the replay of the record) before
+  blaming the code. The controller names the account of case 3 anew for each new
+  fixture.
 
 ### CI results and publication
 

.memory-bank/techContext.md

@@ -221,11 +221,18 @@ source: repository and executable evidence
 - Builds are not byte-reproducible (two unchanged assemblies differ per build):
   hash each candidate and its package separately.
 - The fixture's account for case 3 gets a new name for each new fixture
-  (`NtfsLiveSubject` and four digits). After an account is deleted and created
-  again with the same name, a Kerberos S4U logon returns the old SID and groups
-  for more than seven minutes on the domain controller and the file server
-  (`WindowsIdentity` with a UPN, the Authz functions behind
-  `Get-NTFSEffectiveAccess`), whichever module version asks; a `klist purge`
-  renews only the session's own token, and `nltest /sc_reset`, a DNS flush, and
-  a restart of the Kerberos service change nothing. `Probe-AccountRecreation.ps1`
-  shows it.
+  (`NtfsLiveSubject` and four digits). In the matrix lab, after an account was
+  deleted and created again with the same name, the remote authorization
+  managers (the client's for the default `-ServerName`, the file server's for its
+  name) answered for about ten minutes as if it had no groups (`0x100000`),
+  whichever module version asked, while the Kerberos S4U logon of the oracle, the
+  name resolution, and the local manager were right in the same second. A replay
+  with the baseline and the final candidate alternating failed the baseline in two
+  of three cells and the final candidate in one of three (not counting the warm-up
+  cell). The mechanism in Windows is unknown; a model with one lifetime (9.35 to
+  10.20 minutes) fits all 43 Admin-role runs of 27 cells. When the accounts are
+  created again within seconds, the S4U
+  logon itself returns the old account for 7 to 15 minutes. A `klist purge`,
+  `nltest /sc_reset`, a DNS flush, and a restart of the Kerberos service didn't
+  help. `Probe-AccountRecreation.ps1`, `Export-CellTimeline.ps1`, and
+  `Test-StaleAuthzModel.ps1` show it.