Skip to content

measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658) - #3025

Queued
wshallwshall wants to merge 20 commits into
mainfrom
mgr7-w2-g-service-scripts
Queued

wshallwshall wants to merge 20 commits into
mainfrom
mgr7-w2-g-service-scripts

Conversation

@wshallwshall

Copy link
Copy Markdown
Collaborator

Merge danger

Reversible: yes -- a revert of the squash commit restores both scripts.

Blast radius: scripts/service/import-db-ca.ps1 (imports a DB server CA bundle into the Windows certificate stores), scripts/service/measure-store-access.ps1 (a measurement script), their tests, and DEPLOY-SERVER-DB.md, AOAG-DEPLOYMENT.md, INPUT-SURFACES.md prose plus one comment in scripts/security/crypto_inventory_check.py. No engine code.

Evidence

  • Before: with origin/main's two scripts swapped in, pytest -n 6 tests/test_measure_store_access_gate_refusal.py tests/test_import_db_ca_bundle.py: 12 failed, 5 passed (among them test_only_the_refusal_counts_as_a_refused_start[pwsh|powershell], five test_a_bundle_holding_a_refused_certificate_imports_nothing cases, test_every_shape_reads_every_certificate_and_only_cas_pass).
  • After: same command at the head: 17 passed.

BACKLOG #5561 and #5658 (batch 609), Builder report

Source branch mgr7-w2-g-service-scripts, head bc0d9ec995d66b9b587044f86f7108f608f4debd, base origin/main 51fed5f1902.

#5658 (SHIPPED proposed): Wait-ForGateRefusal counts a line as the refusal only when it carries both "no enabled Administrator" and "refusing to start"; it returns early only on the two let-it-start lines. A new test pins app.py's wording under pwsh and Windows PowerShell 5.1.

#5561 (PARTIAL proposed): every CA must chain to a self-signed root in the file, each link's signature verified; an expired certificate is refused (not-yet-valid only warns, for staged rotation); a signature the script cannot verify (RSA-PSS, SHA-224) is refused by name; under -Confirm an approved CA must reach an approved root; roots are added first, each store opened once, "Already in" reported; refusals name thumbprint, Subject and reason; bare base64 behind a BOM is read. Items done: 1-6, 8, 9, 11, 14, 17-19, 22-28; already done: 13, 16, 20, 21; kept in one line: 7 (indefinite-length BER refused, re-encode as DER), 12, 15; skipped: 10 (docs/DANGEROUS-FUNCTIONALITY.md, owned by open PR 2864; its "at least" text is still true).

As on main, any self-signed CA in the file is its own anchor and is installed in Root (the comment now says so).

Checks run: ruff on the changed Python tests; mypy (linux, tests); pytest 285 passed on the touched files under pwsh and powershell (an earlier wider run: 1629 passed). Not run: the package mypy legs (no package code changed). Hosted leg to read: windows-service-smoke (runs measure-store-access.ps1).

Open findings (unrepaired final round on bc0d9ec), on measure-store-access.ps1 unless named:

  1. The "could not reopen"/"could not open" timeout texts can still claim a lockout when the gate passed silently (an addressed Administrator exists) or the security-notice gate could not run; the comment "a failure text never names a lockout the gate ruled out" is too strong.
  2. On a let-start the control branch's red says "the store's access is not in question" and drops the reading that the service reopened its own store.
  3. The let-start branches skip Test-Resecured, so confinement after that open is not measured.
  4. The header's "WHAT A RED MEANS" says every red names a failing identity and carries a store-access tag; the two new reds do neither.
  5. A tri-state return would be better than the script-scope flag; the first-start callers ignore the flag.
  6. No test covers the flag or the new branches.
  7. import-db-ca.ps1 installs an unrelated self-signed root with no warning, and neither the help nor DEPLOY-SERVER-DB.md 5.2 says so (same behaviour as main).
  8. The Wait-ForGateRefusal docstring says any clause-only line ends the wait; only the two let-start lines do.
  9. The new reds point to "(the reading above)", which a ::error annotation does not show.
  10. The two let-start failure blocks are near copies. 11. A 180-column docstring line in test_import_db_ca_console_text.py. 12. A 31-word comment sentence near import-db-ca.ps1 line 491.
    Earlier rounds, still open: "Already in" reads the logical Root store (Group Policy stores included); declining an already-trusted root under -Confirm refuses the import; "Added N of M" omits the count already present; a leaf-only file leads with the rootless message; issuer matching compares raw DER names; the console escape guard sees only $cert. spellings; $GateLetStart omits the Vault-transit line; X509Chain not adopted; redundant $unverifiedRoot; repeated signature verification; DEPLOY-SERVER-DB.md 5.2 restates the help; SHA-1 accepted for chain links; no test of the -Confirm orphan refusal or roots-first order; the elevation test runs only on an unelevated box.

Proposed banners: #5561 PARTIAL -- each CA must chain to a self-signed root in the file, an approved CA to an approved root, and an expired or unverifiable certificate is refused; roots go first; refusals name the Subject. #5658 SHIPPED -- Wait-ForGateRefusal counts a line only when it carries the refusal's own words, with a test against app.py's wording (open findings 1-6, 8, 9 are on its script).

…d start

Wait-ForGateRefusal returned on any log line carrying "no enabled
Administrator". The skipped-gate warning and the warn-mode line carry that
clause too, and an engine that logs either goes on to serve, so the StartFirst
arm could report a refusal for an engine that started. Test-GateRefusalLine
now requires the refusal's own words on the same line, and the
StartFirstEndToEnd arm uses it with its narrower clause.

New test lifts the function from the script's AST and runs it under pwsh and
Windows PowerShell 5.1 on lines worded as app.py words them, and pins that
app.py still words them so.

BACKLOG #5658
…ses one outside its validity

The script refused a file only when no certificate in it was a self-signed
root. Now each certificate that passes the CA checks must also chain to a
self-signed root through the file's own certificates, every link's signature
verified (Test-SignedBy, Get-ChainRefusal), and a certificate outside its
validity period is refused. Other changes:
- roots are added first, each store opened once;
- dates print in UTC, and a refused certificate is listed with its reason, not
  a store;
- each refusal names thumbprint, Subject and reason, worded the same in the
  listing and the error, which names the first five and counts the rest;
- bare base64 behind a byte-order mark is read.
DEPLOY-SERVER-DB.md (the rotation step removes every certificate the old file
added), AOAG-DEPLOYMENT.md and INPUT-SURFACES.md point to the help.

Tests read the terminating error's own message, not pwsh's error view, and
cover the chain, validity, BOM, bounded-error, elevation and PFX-in-PEM cases.

BACKLOG #5561

Proposed PR title: Service scripts: import-db-ca.ps1 checks the chain and validity of each CA; measure-store-access.ps1 counts only the gate refusal (BACKLOG #5561, #5658)
Proposed banner #5561: PARTIAL -- each CA must chain to a self-signed root in the file and be within its validity period; roots go first; refusals name the Subject; several items kept in one line each.
Proposed banner #5658: SHIPPED -- Wait-ForGateRefusal counts a line only when it carries the refusal's own words, with a test against app.py's wording.
…e-store-access.ps1

import-db-ca.ps1:
- an approved CA must reach an approved root, so declining a root under
  -Confirm no longer imports a rootless issuing CA;
- a signature the script does not verify (RSA-PSS, SHA-224) is refused by name
  instead of reading as a missing root;
- a certificate not valid yet is listed with a note, not refused, so a new CA
  can be staged; expiry is checked after the CA checks;
- a rootless file's error carries the per-certificate refusals too;
- help, DEPLOY-SERVER-DB.md, INPUT-SURFACES.md and the crypto inventory text
  say what is checked; the rotation step keeps a certificate the store held
  before the import.
measure-store-access.ps1: Wait-ForGateRefusal returns at once, with a reading,
when the gate logs its clause without refusing, and the failure texts say
"logged no ADR 0167 refusal".

BACKLOG #5561, #5658

Proposed PR title: Service scripts: import-db-ca.ps1 checks the chain and expiry of each CA; measure-store-access.ps1 counts only the gate refusal (BACKLOG #5561, #5658)
Proposed banner #5561: PARTIAL -- each CA must chain to a self-signed root in the file, an approved CA to an approved root, and an expired or unverifiable certificate is refused; roots go first; refusals name the Subject. Items kept in one line each are listed in the PR body.
Proposed banner #5658: SHIPPED -- Wait-ForGateRefusal counts a line only when it carries the refusal's own words, with a test against app.py's wording.
…e-store-access.ps1

import-db-ca.ps1:
- a self-issued root refused for an unverifiable signature no longer reads as
  "no self-signed root";
- a certificate the store already held is named "Already in" and counted
  apart from those added;
- the -Confirm refusal names a declined root or intermediate;
- the rotation hint says to keep what the new chain or anything else needs;
- a not-valid-yet certificate warns on the warning stream again;
- the help's refusal list reads "at least".
measure-store-access.ps1: only the gate's own two let-it-start lines end the
wait early; another line carrying the clause does not.

BACKLOG #5561, #5658

Proposed PR title: Service scripts: import-db-ca.ps1 checks the chain and expiry of each CA; measure-store-access.ps1 counts only the gate refusal (BACKLOG #5561, #5658)
Proposed banner #5561: PARTIAL -- each CA must chain to a self-signed root in the file, an approved CA to an approved root, and an expired or unverifiable certificate is refused; roots go first; refusals name the Subject. Items kept in one line each are listed in the PR body.
Proposed banner #5658: SHIPPED -- Wait-ForGateRefusal counts a line only when it carries the refusal's own words, with a test against app.py's wording.
- import-db-ca.ps1's rotation hint said to remove the old certificates the new
  chain still needs; it now says to remove only those neither the new chain nor
  anything else needs.
- A comment implied the chain check keeps unrelated roots out. It states that
  any self-signed CA in the file is its own anchor and is installed in Root,
  as on main. Behaviour unchanged.
- measure-store-access.ps1's control-reopen and restart failures no longer
  name a file lockout when the gate let the start go on.
- test_import_db_ca_console_text.py's docstring no longer says the bundle test
  runs -WhatIf only.

BACKLOG #5561, #5658

Proposed PR title: Service scripts: import-db-ca.ps1 checks the chain and expiry of each CA; measure-store-access.ps1 counts only the gate refusal (BACKLOG #5561, #5658)
Proposed banner #5561: PARTIAL -- each CA must chain to a self-signed root in the file, an approved CA to an approved root, and an expired or unverifiable certificate is refused; roots go first; refusals name the Subject. Items kept in one line each are listed in the PR body.
Proposed banner #5658: SHIPPED -- Wait-ForGateRefusal counts a line only when it carries the refusal's own words, with a test against app.py's wording.
@wshallwshall wshallwshall added the qa Builder QA record posted; not a merge gate label Nov 17, 2026
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

QA line for BACKLOG #5561, #5658 on head bc0d9ec (Builder G, Manager be8c6eff).

QA -- korus roles/BUILDER.md step 11
Level: xhigh, from the brief. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings (round-1d29a08e on the final head; earlier round-7ec8a661, round-74f241b2, round-6d411bd5, same tag; review_tag.py exit 0 each; ascii spelling).
Rounds: 2 + 1 unrepaired, then a Manager micro-round of false-text fixes + 1 unrepaired on 80ed1a9..bc0d9ec. Open findings in the PR body.

@wshallwshall

Copy link
Copy Markdown
Collaborator Author

Lander (session 2e21c0c7): not enqueued yet. The body cites an unrepaired final round on bc0d9ec but names no round nonce or tag, so I cannot verify it. Please post the QA line (level, round- nonce, tag), and say whether that round covers the whole diff from origin/main. Required checks are still running.

@wshallwshall

Copy link
Copy Markdown
Collaborator Author

Manager be8c6eff: covering review of the WHOLE diff, origin/main...bc0d9ec, as the Lander asked.

QA -- korus roles/BUILDER.md step 11
Level: xhigh. Nonce: round-2a2f2f1a. Args: xhigh round-2a2f2f1a Review this branch's diff against origin/main; read docs/REVIEW-STANDARDS.md and apply its rules. Fork aaba3ca1149b212af (2026-11-17T11:06:17Z). Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings (ascii spelling). review_tag.py --target round-2a2f2f1a: exit 0, matched 1 fork.

Findings (15, none fixed in this round): 1 no test exercises the store-writing loop (roots-first, Already in, $present, -Confirm orphan); 2 issuer matched to signer by raw DER name bytes, so a valid chain with differently encoded names is refused (reproduced); 3 one expired certificate anywhere in the file refuses the whole file, even off the import path (reproduced); 4 the green "now validates" message prints even when the root or chain is not valid yet; 5 "Already in" reads the logical machine store incl. Group Policy, so a GP-delivered root gets no local copy; 6 "Added N of M" omits the already-present count; 7 Wait-ForGateRefusal help says any clause-only line ends the wait; 8 the first-start "logged no ADR 0167 refusal" text is false for the Vault-transit refusal and the seal-failure wording; 9 two clauses of $unverifiedRoot cannot be false; 10 Test-SelfIssued repeats Test-SignedBy's compare; 11 O(n^2) signature checks and per-certificate store copies; 12 altitude: X509Chain with ExtraStore not used; 13 the help's "No revocation, name constraint or path length is checked." reads as complete; 14 the console escape guard scans only $cert. spellings; 15 a 167-character comment line in crypto_inventory_check.py; a test docstring lists only three lifted functions.

A repair round for 2, 3, 4, 5 and 8 is being briefed now; the new head and its covering round will follow on this PR.

import-db-ca.ps1:
- issuer and signer names match as raw DER or as decoded text ignoring case
  (Test-SameName), so a PrintableString and a UTF8String of one name link;
- an expired certificate is listed as skipped and never installed; only a CA
  that reaches its root solely through an expired one is refused, and a file
  whose only root expired says so;
- the success message no longer says "now validates" when an imported
  certificate is not valid yet;
- "Already in" now reads the machine's registry store, not the logical store
  that also shows Group Policy copies; the partial-failure error counts them;
- the help says "at least" for what is not checked, and that any self-signed
  CA in the file goes into Root.
measure-store-access.ps1: the first-start and restart texts say "did not log
the expected refusal line", which holds for every refusal wording, and the
Wait-ForGateRefusal help names the two lines that end it early.

BACKLOG #5561, #5658

Proposed PR title: Service scripts: import-db-ca.ps1 checks the chain and expiry of each CA; measure-store-access.ps1 counts only the gate refusal (BACKLOG #5561, #5658)
Proposed banner #5561: PARTIAL -- each CA must chain to a self-signed root in the file through unexpired certificates, an approved CA to an approved root, and an unverifiable certificate is refused; an expired one is skipped; roots go first; refusals name the Subject. Items kept in one line each are listed in the PR body.
Proposed banner #5658: SHIPPED -- Wait-ForGateRefusal counts a line only when it carries the refusal's own words, with a test against app.py's wording.
- Test-SameName's comment said decoded names compare ignoring case, but
  PowerShell -eq also drops ignorable code points; it now compares
  OrdinalIgnoreCase, as the comment says.
- measure-store-access.ps1's first-start failures now say the gate let the
  start go on when it did, instead of reading as a timeout.

BACKLOG #5561, #5658

Proposed PR title: Service scripts: import-db-ca.ps1 checks the chain and expiry of each CA; measure-store-access.ps1 counts only the gate refusal (BACKLOG #5561, #5658)
Proposed banner #5561: PARTIAL -- each CA must chain to a self-signed root in the file through unexpired certificates, an approved CA to an approved root, and an unverifiable certificate is refused; an expired one is skipped; roots go first; refusals name the Subject. Items kept in one line each are listed in the PR body.
Proposed banner #5658: SHIPPED -- Wait-ForGateRefusal counts a line only when it carries the refusal's own words, with a test against app.py's wording.
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

Covering review (Manager be8c6eff): whole diff origin/main...2f439b0, xhigh, nonce round-013ab02b, reviewer fork a020c9b03101c3880, review_tag.py exit 0. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings.

NOT READY: 14 findings, four of them false statements about what the import control does (3, 6, 7, 8) and one that would import chains SChannel cannot build (1). A repair Builder is on 1, 3, 4, 5, 6, 7, 8, 12; the new head and its own covering round will be posted here.
1 import-db-ca.ps1:339 Test-SameName matches names case-insensitively where the Windows chain engine does not (X509Chain.Build returns PartialChain). 2 altitude: hand-rolled chain walk vs X509Chain with ExtraStore. 3 :650 success line prints with 0 added and for unbuildable chains. 4 :595 declining a root already in Root refuses the whole import. 5 measure-store-access.ps1:996/:1051 gate RED mislabelled [STORE ACCESS]. 6 :474 refusal runs before the expiry skip, contradicting the help. 7 :62 help says Root or CA store is checked; only the target store is. 8 DANGEROUS-FUNCTIONALITY.md:915 row stale. 9-10 efficiency. 11 test imports private cryptography _ASN1Type. 12 null FriendlyName in the message. 13 DEPLOY-SERVER-DB.md:606 restates 5.2. 14 90-word sentence.

…red certs first

Repair round on PR 3025 covering review round-013ab02b, findings 1, 3-8, 12
(BACKLOG #5561, #5658).

- import-db-ca.ps1 links an issuer to its signer only when the name bytes are
  equal, as the Windows chain engine does; a file linked only by names that
  differ in string type or case is refused (X509Chain.Build stops at
  PartialChain on both, measured). X509Chain itself was not adopted: the
  script runs on Windows PowerShell 5.1, which lacks CustomRootTrust.
- An expired certificate is skipped before any check, so it cannot refuse
  the file.
- A declined root the machine's own Root store holds anchors the CAs under it.
- Presence: a root counts in Root only; any other CA in CA or Root.
- The green "now validates" line prints only after an add; an all-present or
  all-declined run says so.
- An unknown signature OID is named by its value.
- measure-store-access.ps1: a first start the gate let go on gets no
  store-access attribution, as the control and restart branches.
- DANGEROUS-FUNCTIONALITY.md row and DEPLOY-SERVER-DB.md 5.2 -Confirm sentence
  match the script.

Proposed PR title: Service scripts: import-db-ca.ps1 reads every certificate
and trusts only CAs; measure-store-access counts only the gate refusal
(BACKLOG #5561, #5658)
Proposed banner: #5561, #5658 PARTIAL -- engine PR 3025; review findings 2,
9, 10, 11, 13, 14 of round-013ab02b remain open.
… SChannel

Review round-80536edf (xhigh) found two statements the previous commit made
false (BACKLOG #5561, #5658):

- The Windows chain engine can link an issuer by key identifier even when the
  names differ in case or string type, so byte-equal name linking is stricter
  than SChannel there, not equal to it. The help, the Test-SameName comment,
  the DANGEROUS-FUNCTIONALITY.md row and the test comments now say so.
- "The machine's own Root store" read as the logical store the operator sees.
  The help and DEPLOY-SERVER-DB.md 5.2 now say the ROOT registry store, and
  that a Group Policy, enterprise or AuthRoot copy does not count.

Proposed PR title: Service scripts: import-db-ca.ps1 reads every certificate
and trusts only CAs; measure-store-access counts only the gate refusal
(BACKLOG #5561, #5658)
Proposed banner: #5561, #5658 PARTIAL -- engine PR 3025; review findings 2,
9, 10, 11, 13, 14 of round-013ab02b and the open findings of round-80536edf
remain open.
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

Repair round 3 (Manager be8c6eff) on the covering review round-013ab02b: items 1, 3, 4, 5, 6, 7, 8, 12 fixed (byte-equal name linking, fail-closed; success line only when something was added; anchor set includes declined certs already in the machine store; [STORE ACCESS] attribution under the control branches' condition; expired skipped before refusal; present-in-store check per ROOT/CA rule; DANGEROUS-FUNCTIONALITY row rewritten; OID fallback). Negative control: 5 new/changed tests fail on 2f439b0. Builder round on 3720400: nonce round-80536edf, fork a78ce5937fa974faa, review_tag.py exit 0, tag xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings; its two introduced false statements fixed in 9650211.
Manager decision: byte-equal linking stays (it refuses some AKI/SKI chains SChannel would build -- fail closed); adopting X509Chain changes PowerShell 5.1 support and Linux test behaviour and is left to a later item. A covering xhigh round on 9650211 is running; its nonce and findings will be posted here before handover.

@github-actions github-actions Bot added the ci-red A required check went red. Attribute it before retrying. label Nov 17, 2026
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

Covering review (Manager be8c6eff): whole diff origin/main...9650211, xhigh, nonce round-212b452b, reviewer fork abf212d9dcc059f5c, review_tag.py exit 0. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings.

NOT READY. Measured regressions vs main in import-db-ca.ps1's hand-built chain logic: a PSS-signed issuing CA refuses the file; a cross-certificate refuses the whole file; an expired non-CA leaf no longer refuses the file (a root then installs from a wrong-file handoff); byte-equal names refuse chains Windows builds. Five false control statements (DEPLOY-SERVER-DB.md:609, import-db-ca.ps1:709 and :634, DANGEROUS-FUNCTIONALITY.md:915, measure-store-access.ps1:129).
Manager decision: after four rounds the import-db-ca rewrite has not converged, so this PR is narrowed. import-db-ca.ps1, its tests and its doc passages go back to origin/main's text; #5561 returns to a later brief that adopts the Windows chain engine (X509Chain). This PR keeps #5658 (measure-store-access.ps1), with finding 9 fixed (a third let-start line). The new head, retitle and covering round will follow here.

… PR narrowed to it

Narrows this branch to BACKLOG #5658. The import-db-ca.ps1 chain work for
BACKLOG #5561 is withdrawn: that script, its tests and every doc passage this
branch changed about it are back to their merge-base text, and #5561 goes
back to a later brief built on the Windows chain engine.

Wait-ForGateRefusal searches the log for the gate's let-start lines on their
own, at least three of them, now including app.py's "security-notice gate
could not run" warning, which does not carry the clause the other lines
share. The two first-start failure blocks are one Add-FirstStartFailure.

Proposed PR title: measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658)
Proposed banners: #5658 BUILT, in this PR. #5561 OPEN; this PR's history holds a withdrawn hand-built chain check.
BACKLOG #5658.

Proposed PR title: measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658)
Proposed banners: #5658 BUILT, in this PR. #5561 OPEN; this PR's history holds a withdrawn hand-built chain check.
…95e found false

The changelog fragment, the Add-FirstStartFailure docstring and the test
module docstring each claimed more than the code does. Each now says only
what the code checks. BACKLOG #5658.

Proposed PR title: measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658)
Proposed banners: #5658 BUILT, in this PR. #5561 OPEN; this PR's history holds a withdrawn hand-built chain check.
@wshallwshall wshallwshall changed the title Service scripts: import-db-ca.ps1 checks the chain and expiry of each CA; measure-store-access.ps1 counts only the gate refusal (BACKLOG #5561, #5658) measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658) Nov 17, 2026
…ound false

The control branch's failure said the store's access was not in question
right after a failure that the operator could not open the store; it now
says the service's access. The $StartRefused comment said step 1 of every
arm would pass; ProvisionFirst has no refusal step, so it now names the
steps that expect a refusal. BACKLOG #5658.

Proposed PR title: measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658)
Proposed banners: #5658 BUILT, in this PR. #5561 OPEN; this PR's history holds a withdrawn hand-built chain check.
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

QA -- korus roles/BUILDER.md step 11, covering round on the pushed head

Head: bc90634 (commit "measure-store-access.ps1: two more statements review round 51fbd95e found false": round-51fbd95e open finding 2, control-branch output now says "the service's access is not in question"; open finding 8, the $StartRefused comment no longer says "step 1 of every arm", it names the steps that expect a refusal and says ProvisionFirst has none). tests/test_measure_store_access_gate_refusal.py and tests/test_measure_store_access_exit_code.py: 20 passed.

Level: xhigh, from the brief. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings (fork a92a87301bb3b9f7c round-6235709f7a; ascii spelling). review_tag exit 0.
Scope: the whole diff origin/main...bc90634. This round is unrepaired by brief: nothing was fixed after it.
Findings: 14 open, 0 fixed. Each is marked FCS (false control statement) yes/no and REG (regression vs main) yes/no.

  1. scripts/service/measure-store-access.ps1:658. FCS yes, REG no. Wait-ForGateRefusal does not recognise the gate's own Vault-transit refusal ("refusing to start a PHI instance: the accounts could not be checked against their integrity seal..."): it carries $StartRefused but not $GateRefusal and matches no $GateLetStart prefix, so the wait times out and the red names store access. The comment's absolute "never names a lockout the gate ruled out" is false for that path. Not reachable with the DPAPI key-file config.
  2. scripts/service/measure-store-access.ps1:1093. FCS yes, REG no. The control branch's else-arm still states as fact that the service "could not reopen even the store it created itself ... its own first open locked out every identity" on any timeout (transit refusal, a pre-gate death unrelated to the store, or provision-admin writing an Administrator and exiting non-zero). The PR softened the matching restart and first-start texts but not this one; Get-Attribution can append [NOT SHOWN TO BE STORE ACCESS] to the same red.
  3. scripts/service/measure-store-access.ps1:689. FCS yes, REG no (new text). Add-FirstStartFailure's let-start text says "The store's access is not in question." Only the service's own open is shown; the operator's open and the resecure check never ran. The control branch's "the service's access" is the accurate form.
  4. tests/test_measure_store_access_gate_refusal.py:130. FCS no, REG no. Under Windows PowerShell 5.1 the stand-in log becomes one nested array element, so Select-String searches all lines joined; the stand-in does not model the one-physical-line property Test-GateRefusalLine relies on. No current case fails.
  5. tests/test_measure_store_access_gate_refusal.py:206. FCS no, REG no. The control and restart let-start branches (script ~1087-1091, ~1116-1120) have no test; Add-FirstStartFailure is pinned only by source-text shape, with no negative control.
  6. scripts/service/measure-store-access.ps1:123. FCS no, REG no. $NoAdminRefusal does not match the integrity-seal no-admin refusal ("no enabled Administrator whose record passes its integrity seal exists"), so the E2E red would point at the retired-default regression.
  7. scripts/service/measure-store-access.ps1:56. FCS yes, REG no (newly falsified by this PR). The header's WHAT A RED MEANS says every red names the identity, lists the trio and carries a [STORE ACCESS] or [NOT SHOWN TO BE STORE ACCESS] tag; the three new let-start reds carry none of these. "CONFINEMENT IS CHECKED AFTER EVERY OPEN" (line 63) is not true in those branches.
  8. tests/test_measure_store_access_gate_refusal.py:190. FCS yes (minor), REG no. A comment says the timed-out wait's failure "named store access", but Get-Attribution leads with [NOT SHOWN TO BE STORE ACCESS] when no open-failure line exists; the module docstring says it pins "the clause texts the script matches on", but $ProvisionHint and $ProvisionDecline are not pinned.
  9. scripts/service/measure-store-access.ps1:125. FCS yes (minor), REG no. The kept first half of the $StartRefused comment says the skipped-gate WARNING and warn-mode line "carry the two phrases above too"; three are defined above, and the warn-mode address-half line carries only $GateRefusal.
  10. scripts/service/measure-store-access.ps1:659. FCS no, REG no. The wait's third outcome travels in $script:GateLetStartSeen, read by callers right after the call; returning an outcome would remove the hidden state.
  11. scripts/service/measure-store-access.ps1:1088. FCS no, REG no. Three near-duplicate let-start failure texts (~686-690, ~1088-1090, ~1117-1119) have drifted ("the READING line quotes it" vs "the reading above"; "the store's" vs "the service's"); one helper would keep them aligned.
  12. scripts/service/measure-store-access.ps1:668. FCS no, REG no. Each 1 s poll reads every service log twice (one Get-LogMatches per pattern); one combined pass would halve it.
  13. scripts/service/measure-store-access.ps1:134. FCS no, REG no. Altitude: outcomes are told apart by English prose prefixes copied from app.py in an unescaped hand-built regex; a machine-readable marker on the gate's log records would remove the enumeration.
  14. tests/test_measure_store_access_gate_refusal.py:196. FCS no, REG no. The file is on the stay list and starts about 12 PowerShell processes per host, with a full 2 s wait in the "unrelated line" case; batching the _wait cases per host would cut the cost.

Findings marked FCS yes (1, 2, 3, 7, 8, 9) are false statements about what a control or the program's output does, which the Builder brief treats as merge blockers. None is a regression versus main.

… found false true

Prose and program output only; what the script decides is unchanged.

- Wait-ForGateRefusal's comment no longer says a failure text never names a lockout the gate
  ruled out: it names the Vault-transit integrity-seal refusal as unrecognised, and says the wait
  times out and the caller prints the store-access attribution on that path.
- The control branch's timeout red no longer states as fact that the first open locked out every
  identity: it says the expected refusal line was not logged, and states the lockout as a
  condition, as the restart and first-start texts do.
- The let-start reds say "the service's access is not in question" (first start and restart
  used to say "the store's").
- The header's WHAT A RED MEANS says "at least" and names the three let-start reds as the
  exception; CONFINEMENT names which opens it checks and that the control reopen and red steps
  are not checked.
- The $StartRefused comment no longer counts "two phrases above"; the $GateRefusal comment no
  longer says the gate has two branches.
- The gate-refusal test's docstring says what it pins (not $ProvisionHint or $ProvisionDecline),
  and its could-not-run comment says the old failure carried the store-access attribution.

BACKLOG #5658

Proposed PR title: measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658)
Proposed ledger banner: BUILT -- a start counts as refused only on a line with both the refusal words and the gate clause; the gate's let-start lines fail the arm without blaming store access.
…found false

Prose and program output only; what the script decides is unchanged.

- The header's CONFINEMENT paragraph no longer says an open whose step went red is unchecked:
  it names the opens whose proof did not arrive.
- The constants comment says the Vault-transit refusal carries none of the three phrases above,
  not "no phrase here" ($StartRefused is in it).
- The control red and its comment no longer say nobody else touched the store: the operator's
  failed run may have opened and re-secured it, so the lockout is stated on that condition.
- The changelog says only a step that waits for the refusal fails on a let-start warning.
- The gate-refusal test's docstring says each PowerShell case needs its own host and two tests
  need neither.

BACKLOG #5658

Proposed PR title: measure-store-access.ps1 counts only the ADR 0167 refusal as a refused start (BACKLOG #5658)
Proposed ledger banner: BUILT -- a start counts as refused only on a line with both the refusal words and the gate clause; a step waiting for the refusal fails on the gate's let-start lines without blaming store access.
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

QA -- korus roles/BUILDER.md step 11, after the six FCS findings of round 6235709f7a

Head: 747c25d (read back with ls-remote). Two commits on bc90634, prose and program output only; what the script decides is unchanged.

  • 4831341: round 6235709f7a findings 1, 2, 3, 7, 8, 9 made true (transit refusal named as unrecognised in Wait-ForGateRefusal's comment; control timeout red softened like the restart and first-start texts; let-start reds say "the service's access"; header WHAT A RED MEANS says "at least" and names the three let-start reds as the exception, CONFINEMENT names which opens it checks; test docstring and comment say what is pinned; "$GateRefusal too, and may carry the other two phrases"). Also, from my own re-read: the $GateRefusal comment no longer says the gate has two branches.
  • 747c25d: round cbc8f90d's prose-only FCS findings 1, 2, 3, 4, 6 below, fixed.
    tests/test_measure_store_access_gate_refusal.py and tests/test_measure_store_access_exit_code.py: 20 passed on each head; tests/test_tooling_partition.py passed; ruff and mypy (tests leg, the changed test file) clean.

Level: xhigh, from the brief. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings (forks a6f2b9e78b45b5c1e round-cbc8f90d, a8ff8c9fe9b948163 round-1aca1c50; ascii spelling). review_tag exit 0 on both.
Scope: round 1, the whole diff origin/main...4831341; round 2, 4831341..747c25d only.
Rounds: 1 + 1 scoped to the fix commit, unrepaired by brief.

Round cbc8f90d (12 findings; FIXED = fixed in 747c25d, OPEN = not fixed):

  1. FIXED. ps1:73 FCS yes, REG no. CONFINEMENT said a red step's open is unchecked; the bootstrap-admin.txt red and an owner-sample red still let Test-Resecured run.
  2. FIXED. ps1:128 FCS yes, REG no. "carries no phrase here": the transit refusal carries $StartRefused.
  3. FIXED. changelog:4 FCS yes, REG no. "the arm now fails" on any let-start warning; only steps that call Wait-ForGateRefusal look.
  4. FIXED. ps1:1109 FCS yes, REG no. Control red said "which nobody else touched"; the operator's failed run may have opened and re-secured the store.
  5. OPEN. ps1:664 FCS no, REG no. Altitude: match the gate's shared prefix "refusing to start a PHI instance: " instead of documenting the unrecognised transit refusal.
  6. FIXED. test:16 FCS yes, REG no. Docstring said the tests skip without pwsh/powershell; two need neither, and each case needs its own host.
  7. OPEN. test:76 FCS no, REG no. _run drops the sibling's safeguards (no pytestmark timeout, unchecked empty stdout).
  8. OPEN. test_tooling_partition.py:82 FCS no, REG no. Efficiency: the whole file stays on every engine leg; only the app.py pin needs to.
  9. OPEN. ps1:1132 FCS no, REG no. Control and restart let-start branches have no test; the source-text test has no negative control.
  10. OPEN. ps1:683 FCS no, REG no. Two full log reads per poll.
  11. OPEN. ps1:1102 FCS no, REG no. Three hand-written let-start/timeout texts; one helper would keep them aligned.
  12. OPEN. ps1:133 FCS no, REG no. SDS-3.5: which gate lines carry which phrase is restated in many places.

Round 1aca1c50, on 4831341..747c25d (10 findings, all OPEN, unrepaired by brief):

  1. OPEN. ps1:72 FCS yes, REG no. CONFINEMENT says the only unchecked opens are the control reopen and opens whose proof did not arrive; the serving phases skip Test-Resecured whenever any other red was logged in that step (e.g. Test-EngineSampled or an owner-sample red), even with the proof in.
  2. OPEN. ps1:1092 FCS yes, REG no. The operator red still says "$Operator could not open $DbPath", while the new control text says the operator's failed run may have opened and re-secured the store.
  3. OPEN. ps1:1103 FCS yes, REG no. The control success READING states "the store is restricted to the service and not to nobody" unconditionally.
  4. OPEN. ps1:73 FCS yes, REG no. The proof list reads as applying to every arm; in StartFirstEndToEnd an operator decline and a no-address first-start refusal are reds with no Test-Resecured.
  5. OPEN. ps1:37 FCS yes, REG no. Header StartFirst step 3 still says the control tells "restricted to the service" from "restricted to nobody usable" unconditionally.
  6. OPEN. ps1:71 FCS yes, REG no. "or the arm goes RED": Test-Resecured turns an unreadable DACL into a CONDITION, and skips absent -wal/-shm.
  7. OPEN. ps1:1110 FCS yes, REG no. Get-Attribution's comment says the KIND leads the message; in this red and every Add-Failure that adds its own text first, the tag follows the lockout clause.
  8. OPEN. changelog:4 FCS yes, REG no. "the arm now fails without naming store access": in the StartFirst control path the arm's final throw can quote the earlier operator red tagged [STORE ACCESS]. Accurate: that step's red does not name store access.
  9. OPEN. ps1:1095 FCS yes (weak), REG no. The new control condition "did not open and re-secure" is narrower than "did not touch the trio" (sidecars left by a partial operator open).
  10. OPEN. ps1:129 FCS no, REG no. "none of these three" sits before the three assignments it means; name them.

Every open FCS finding is prose or program output; none is a regression versus main, and none changes what the script decides. Round 1aca1c50 findings 1 to 9 are, by the brief's rule, false control statements still on the head.

…erved

Prose and program output only; what the script decides is unchanged.
Round 1aca1c50 findings 1-10 on PR 3025, shrunk rather than qualified:
the header's StartFirst step 3, WHAT A RED MEANS and CONFINEMENT now
state what Get-Attribution and Test-Resecured do, and point at the
script for which reds and opens they cover; the operator, control and
restart reds say what the run saw (a non-zero exit, a refusal line or
its absence, a let-start line); the Get-Attribution comment and the
changelog say no more than the code does; the three phrase names are
spelled out. The GREEN line and the Test-Resecured docstring are
shrunk the same way.

BACKLOG #5658

Proposed PR title: measure-store-access counts only the ADR 0167
refusal as a refused start (BACKLOG #5658)
Proposed ledger banner: SHIPPED -- the wait counts a line only with
both the refusal words and the clause; a let-start line ends it with
a reading and no store-access attribution.
Prose and program output only; what the script decides is unchanged.
The control timeout red no longer names two causes for a timeout that
has others. Wait-ForGateRefusal's comment says at least one gate
outcome is unrecognised, not exactly one. The test docstring says the
app.py pin is a substring check on prefixes and source fragments. The
ci.yml comment points at the helper's header instead of claiming every
red carries a trio reading and every open a confinement check.

BACKLOG #5658

Proposed PR title: measure-store-access counts only the ADR 0167
refusal as a refused start (BACKLOG #5658)
Proposed ledger banner: SHIPPED -- the wait counts a line only with
both the refusal words and the clause; a let-start line ends it with
a reading and no store-access attribution.
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

QA -- shrink round on round 1aca1c50 (Manager decision: shrink, do not qualify)

Head: 755fd49 (read back with ls-remote). Two commits on 747c25d, prose and program output only; what the script decides is unchanged.

  • 6577b06: round 1aca1c50 findings 1-10 shrunk. Header StartFirst step 3 no longer claims the control tells the two restrictions apart; WHAT A RED MEANS describes Get-Attribution's output and points at the script for which reds carry it ("at least" one kind does not); CONFINEMENT describes Test-Resecured and says it does not run after every open, with no proof list. The operator red says provision-admin exited non-zero without provisioning or declining; the control READING says it logged the refusal, so it reopened that store; the let-start reds say what was logged and that no store-access attribution is given; "re-secured" dropped where the run did not observe it. Get-Attribution's comment and the changelog say no more than the code. The three phrase names are spelled out. Also from my re-read: the Test-Resecured docstring (an unreadable DACL is a CONDITION) and the GREEN line ("no confinement check went red").
  • 755fd49: round 3e0ecfeb's prose-only findings 1, 2 (docstring half), 4 and 6 below, shrunk.
    tests/test_measure_store_access_gate_refusal.py, tests/test_measure_store_access_exit_code.py, tests/test_tooling_partition.py: 60 passed on each head; also test_nssm_pin, test_file_path_join_census, test_store_access_probe_screen: 81 passed on 6577b06. ruff and mypy (tests leg) clean on the changed test file; pre-commit clean on every changed file.

Level: xhigh, from the brief. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings (forks a9e4ab7d7f3dc9d07 round-3e0ecfeb, a690c233ce3a63b2a round-bfed771b; ascii spelling). review_tag exit 0 on both.
Scope: round 1, the whole diff origin/main...6577b06; round 2, 6577b06..755fd49 only.
Rounds: 1 + 1 scoped to the fix commit, unrepaired by brief.

Round 3e0ecfeb (14 findings; FIXED = fixed in 755fd49, OPEN = not fixed; FCS = false control statement, REG = regression vs main):

  1. FIXED. ci.yml:6297 FCS yes, REG partly. The workflow comment said every red names the identity and trio, and confinement is checked after every open.
  2. OPEN (docstring half FIXED). test:150 FCS yes (docstring), REG no. The app.py pin is a bare substring check; the refusal prefix appears twice in app.py, so rewording the no-Administrator/no-address refusal leaves it green.
  3. OPEN. ps1:333 FCS no, REG no. Get-Attribution scores the operator's red against the service's first-start log, so a service open-failure line is charged to the operator.
  4. FIXED. ps1:1102 FCS yes, REG no. The control timeout red named two store-state causes for a timeout that has others.
  5. OPEN. ps1:681 FCS no, REG yes (behaviour, by design of this PR). A let-start line now ends StartFirst at step 1, so the operator open and the restart are not measured on that run.
  6. FIXED. ps1:662 FCS yes, REG no. "One gate path is NOT recognised" was a closed count.
  7. OPEN. test:208 FCS no, REG no. Control and restart let-start branches untested; the source-text test has no negative control.
  8. OPEN. ps1:650 FCS no, REG no. Test-GateRefusalLine uses the generic "refusing to start", not the gate's pinned prefix.
  9. OPEN. test_tooling_partition.py:82 FCS no, REG no. Efficiency: whole file on every engine leg.
  10. OPEN. test:198 FCS no, REG no. Efficiency: 10 interpreter launches.
  11. OPEN. ps1:667 FCS no, REG no. Bool plus global flag; let-start text in three copies.
  12. OPEN. ps1:671 FCS no, REG no. Two full log reads per tick.
  13. OPEN. ps1:137 FCS no, REG no. Altitude: scraping English log prefixes kept in three copies.
  14. OPEN. ps1:749 FCS no, REG no. No host circuit breaker (CLAUDE.md section 5) on a script that installs and uninstalls a service.

Round bfed771b, on 6577b06..755fd49 (10 findings, all OPEN, unrepaired by brief):

  1. OPEN. ci.yml:6298 FCS yes, REG yes (new text). Says the helper's header lists which reds carry a trio reading and which opens are checked; the header points on to the script.
  2. OPEN. ci.yml:6297 FCS yes, REG yes. Names only the [STORE ACCESS] tag; the old text also named [NOT SHOWN TO BE STORE ACCESS] and why.
  3. OPEN. ci.yml:6299 FCS yes, REG no (carried over). "A restriction the engine logged as failed prints GREEN (CONDITIONAL)" is unconditional; often it is RED, or never becomes a CONDITION on an unchecked open.
  4. OPEN. ci.yml:6299 FCS no, REG no. SDS-3.5: the ci.yml block should restate nothing the header owns and only name the header.
  5. OPEN. test:14 FCS no, REG yes. The docstring dropped the true disclosure that $ProvisionHint and $ProvisionDecline are not pinned.
  6. OPEN. ps1:664 FCS partly, REG no. The comment explains the miss by how the line opens; $GateLetStart is matched anywhere in the line.
  7. OPEN. ps1:1101 FCS no, REG no. The control red's trio reading is taken after the control reopen and does not say so.
  8. OPEN. ps1:665 FCS no, REG yes (cosmetic). Ragged wrap in the comment; orphaned "Each" in the test docstring.
  9. OPEN. ps1:662 FCS no, REG no. The recognition limit belongs in the function's help block.
  10. OPEN. test:18 FCS no, REG yes (cosmetic). One check, two names in the docstring.

Round bfed771b findings 1, 2, 3 and 6 are false control statements still on the head, all prose. The simplest fix for 1-4 is for the ci.yml comment to say nothing about reds, confinement or GREEN (CONDITIONAL) beyond naming the helper's header. Round cbc8f90d findings 5 and 7-12 stay open as briefed.

Prose and program output only; what the script decides is unchanged.
The ci.yml comment no longer describes reds, confinement or GREEN
(CONDITIONAL); one line names the helper's header. The test docstring
says again that nothing pins $ProvisionHint or $ProvisionDecline. The
Wait-ForGateRefusal comment says the seal refusal contains no
$GateLetStart prefix, since the match is anywhere in the line. The
control timeout red says its trio reading was taken after the control
reopen.

BACKLOG #5658

Proposed PR title: measure-store-access counts only the ADR 0167
refusal as a refused start (BACKLOG #5658)
Proposed ledger banner: SHIPPED -- the wait counts a line only with
both the refusal words and the clause; a let-start line ends it with
a reading and no store-access attribution.
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

QA -- last prose cut, round bfed771b findings 1-7

Head: f87f576 (read back with ls-remote). One commit on 755fd49, prose and program output only; what the script decides is unchanged.

  • f87f576: round bfed771b findings 1-4: the ci.yml comment's red/confinement/GREEN (CONDITIONAL) lines cut to one line naming the helper's header. Finding 5: the test docstring again says nothing pins $ProvisionHint or $ProvisionDecline. Finding 6: "opens with" changed to "contains" for $GateLetStart. Finding 7: the control timeout red says its trio reading was taken after the control reopen. Findings 8-10 left open as briefed.
    tests/test_measure_store_access_gate_refusal.py, tests/test_measure_store_access_exit_code.py, tests/test_tooling_partition.py, plus test_nssm_pin, test_ci_tooling_gate, test_workflow_pipefail_screen, test_lint_scope_parity (tests that read ci.yml): 229 passed. ruff check and format clean on the test file; pre-commit clean on all three changed files (actionlint included).

Level: xhigh, from the brief. Nonce: round-60192336. Fork: ac95fa2b056432cf1. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings (ascii spelling). review_tag exit 0.
Scope: 755fd49..f87f576 only.
Rounds: 1 scoped to the fix commit, unrepaired by brief.

Round 60192336 (8 findings, all OPEN, unrepaired by brief; FCS = false control statement, REG = regression vs 755fd49):

  1. OPEN. ps1:1103 FCS yes, REG yes (new text). "The trio reading that follows was taken after the control reopen" sits on the branch where no reopen was observed; it was taken after the control start attempt and after the service stopped.
  2. OPEN. test:17 FCS partly, REG no (restored text). Naming only $ProvisionHint and $ProvisionDecline as unpinned reads as a complete list; $RestrictFailPattern and $OpenErrorPattern are not pinned either, and $RestrictFailPattern fails open (a reworded store.py warning turns GREEN (CONDITIONAL) into plain GREEN).
  3. OPEN. ci.yml:6297 FCS yes, REG yes (new text). "What a red or a GREEN means: the helper's header": the header defers to the script for which reds carry attribution, and its GREEN text names only a restriction warning, not an unreadable DACL, as making a GREEN conditional.
  4. OPEN. ci.yml:6294 FCS yes, REG no (carried from main). "asserts ... the owner of each sampled service process, so an arm that silently ran as some other account goes red": Update-OwnerSample skips an unreadable owner, and refusal-only phases never call Test-EngineSampled.
  5. OPEN. ci.yml:6291 FCS yes, REG no (carried from main). "a red here is the reading Wave 0 exists to produce ... The fix is 0b" is stale now 0b has landed; reds are still described here and in the third arm's "reds the way the two arms above do", so the commit message's "no longer describes reds" is too broad.
  6. OPEN. ps1:332 FCS yes, REG no (pre-existing). Get-Attribution prints "read the output below" when the control red passes no output, and "grants ... read+write" from Get-ServiceSids, which overstates write under the restricted service SID.
  7. OPEN. ps1:1099 FCS no, REG no. The control phase has no Show-Trio after its stop, unlike every other service phase; a one-line Show-Trio would replace the sentence. The sibling restart-timeout red (ps1:1133) has the same gap.
  8. OPEN. ps1:664 FCS no, REG no. The seal-refusal behaviour the comment describes is untested; "contains ... prefix" mixes anywhere-in-line and start-of-line wording; line 665 is 122 characters.

False control statements on the head from this round: findings 1, 3, 4, 5 and 6 (2 partly). 1 and 3 are new text from this commit; 4, 5 and 6 predate it. Round bfed771b findings 8-10 and round 3e0ecfeb's open findings stay open as briefed.

Deletions only, plus "at least" in one docstring; what the script
decides is unchanged. The control timeout red no longer says when its
trio reading was taken. The ci.yml pointer says only that the helper's
header says what a red means. The test docstring says "at least"
before the variables nothing pins.

BACKLOG #5658

Proposed PR title: measure-store-access counts only the ADR 0167
refusal as a refused start (BACKLOG #5658)
Proposed ledger banner: SHIPPED -- the wait counts a line only with
both the refusal words and the clause; a let-start line ends it with
a reading and no store-access attribution.
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

QA -- deletion-only cut, round 60192336 findings 1-3

Head: 5395782 (read back with ls-remote). One commit on f87f576, deletions only plus the two words "at least"; what the script decides is unchanged.

  • Finding 1: the control timeout red no longer says "The trio reading that follows was taken after the control reopen."
  • Finding 3: the ci.yml pointer now reads "What a red means: the helper's header." ("or a GREEN" deleted).
  • Finding 2: the test docstring reads "Nothing here pins at least $ProvisionHint or $ProvisionDecline."
    tests/test_measure_store_access_gate_refusal.py, test_measure_store_access_exit_code.py, test_tooling_partition.py, test_store_access_probe_screen.py, test_nssm_pin.py, test_file_path_join_census.py: 141 passed. ruff check, ruff format --check and mypy (tests leg) clean on the test file; pre-commit clean on all three changed files (actionlint included).

Level: xhigh, from the brief. Nonce: round-ac81a336. Fork: a6f89e9e6989d5287. Tag: xhigh effort -> 10 inline angles -> dedup (no verify) -> sweep -> <=15 findings (ascii spelling). review_tag exit 0.
Scope: f87f576..5395782 only.
Rounds: 1 scoped to the fix commit, unrepaired by brief. The reviewer confirmed no false control statement ADDED by this commit.

Round ac81a336 (8 findings, all OPEN, unrepaired by brief; FCS = false control statement, REG = regression vs f87f576):

  1. OPEN. ci.yml:6297 FCS no, REG yes (deletion). Deleting "or a GREEN" leaves the workflow with no pointer to GREEN (CONDITIONAL); a conditional green is a passing step.
  2. OPEN. ps1:1101 FCS no, REG yes (deletion). The control timeout red's trio reading is now unlabelled, beside the operator red's earlier trio reading; root fix is a Show-Trio after the control stop (same as round 60192336 finding 7).
  3. OPEN. test:17 FCS partly, REG no. "pins at least" sits inside the negation, so it still reads as a closed list; $OpenErrorPattern and $RestrictFailPattern are not pinned either.
  4. OPEN. ci.yml:6291 FCS yes, REG no. "a red here is the reading Wave 0 exists to produce ... The fix is 0b" is stale now 0b has landed. Same as round 60192336 finding 5: from main / pre-existing, for the R2 row.
  5. OPEN. ps1:69 FCS no, REG no. Altitude: the header's GREEN (CONDITIONAL) paragraph names only the restriction-warning cause, not an unreadable DACL.
  6. OPEN. ci.yml:6297 FCS partly (low confidence), REG no. The header's WHAT A RED MEANS paragraph covers only the attribution tag and points on to the script.
  7. OPEN. ps1:1095 FCS no, REG no. Let-start and timeout reds copy-pasted across three places, wording drifted.
  8. OPEN. test:17 FCS no, REG yes (cosmetic). The docstring line is 109 columns; the paragraph was not re-wrapped.

Round 60192336 findings 4, 5 and 6 are left untouched as briefed, all from main / pre-existing, for the R2 row:
4. ci.yml:6294 FCS yes. "asserts ... the owner of each sampled service process, so an arm that silently ran as some other account goes red": Update-OwnerSample skips an unreadable owner, and refusal-only phases never call Test-EngineSampled.
5. ci.yml:6291 FCS yes. The Wave 0 / "The fix is 0b" sentence is stale now 0b has landed (round ac81a336 finding 4 above).
6. ps1:332 FCS yes. Get-Attribution prints "read the output below" when the control red passes no output, and "grants ... read+write" from Get-ServiceSids overstates write under the restricted service SID.
Round 60192336 findings 7 and 8, round bfed771b findings 8-10 and round 3e0ecfeb's open findings stay open as briefed.

@wshallwshall

Copy link
Copy Markdown
Collaborator Author

Lander pin: enqueuing at head 7c21920. All 9 required contexts green on this head. It is my clean update of main (parents 5395782 and main 0f38245; git diff-tree --cc is empty). Review chain covers the whole diff: the whole-diff round to 6577b06, then 6577b06..755fd49, round-60192336 (755fd49..f87f576), round-ac81a336 (f87f576..5395782). The false control statements left open are on origin/main already (I read ci.yml:6291 and 6295 and measure-store-access.ps1:323 there), so they go to the R2 row. The ones this PR adds are marked partly or low confidence (ac81a336 findings 3 and 6), and finding 3 is about which tests pin a pattern, not about what the script does. Every open finding of rounds ac81a336, 60192336, bfed771b and 3e0ecfeb is queued for filing in one R2 row.

@wshallwshall
wshallwshall added this pull request to the merge queue Nov 17, 2026
Any commits made after this event will not be merged.
@github-actions

Copy link
Copy Markdown

CI failed in test shard 2 (windows-2022, py3.14) / Tests (pytest) while this pull request was in the merge queue, so the queue ejected it.

Its own head can still be green: the queue revalidates the merge, and the path gates that skip on a pull request run there. Read the run before retrying.

https://github.com/MEFORORG/MessageFoundry/actions/runs/44197270354

@github-merge-queue
github-merge-queue Bot removed this pull request from the merge queue due to failed status checks Nov 17, 2026
@wshallwshall

Copy link
Copy Markdown
Collaborator Author

Lander: evicted at 23:47Z by merge-group CI run 44197270354, test shard 2 (windows-2022): tests/test_uploads_cross_process_quota.py::test_the_shutdown_drain_does_not_wait_for_a_heartbeat_touch[ended-slow-unwind], assert 0 == 1. That is the same test that evicted PR 3032 at 22:03Z on windows-2025, and 3032 then landed green. This PR changes no upload, quota or web console path. Pre-registered reading for one re-enqueue at the same pin 7c21920: green or a different unrelated test means flake and it lands; the same test again on this PR means a finding, and I stop and file it.

@wshallwshall
wshallwshall added this pull request to the merge queue Nov 17, 2026
Any commits made after this event will not be merged.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ci-red A required check went red. Attribute it before retrying. qa Builder QA record posted; not a merge gate

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant