Owner decision docket: 2026-08-04 (refreshed 2026-08-04, late)
Every open owner decision, one page each: what needs deciding, the verified evidence, the options, and a recommendation. The goal is that each ruling takes minutes, not a re-derivation. Items that closed since they were queued are at the bottom with their closing sha, not deleted.
This is a full re-verification, not an edit. The docket landed at
62f6e3673; 77 commits have landed on shc and 23 on ce since. Every item
below was re-measured against today’s origin/main and, where it touches an
estate, against both live estates; nothing here is carried on report.
What that re-measurement found:
- Seventeen of the eighteen items are still open, and the eighteenth split:
its netbird half closed outright (
8d0743e86), its jellyfin half survives below as item 22. So the docket did not shrink by ruling. It shrank by one. - What closed en masse is the mechanism queue the items cited as context: nineteen separate defects and decisions, all listed at the bottom with shas. That is why several items’ evidence moved even though the items themselves did not, and it is why a re-derivation was cheaper than an edit.
- Three items came back materially different from how they were filed (items 7, 13 and 14), and two more (8, 11) came back with smaller corrections. All five are flagged inline and summarised at the end.
- Four items are new (18–21), all surfaced by tonight’s work: what pinning a
system app should even mean, whether repo
priorityshould break the ambiguity tie, the six corp stacks with no safe pin, and the catalog design’s remaining questions.
Baselines at refresh: shc origin/main 894d2a186, ce origin/main
c99b51c, shc’s ce pin c99b51c2865f (425bc61b1), i.e. the pin is at
ce’s head, nothing landed in ce is waiting outside it. Both estates are on
v0.0.19 (shc version on both leaders: promanager 204.168.129.34, bitspur
37.27.215.210; the pm deb reads 1:0.0.19). v0.0.20’s changelog is drafted
(894d2a186) in a concurrent lane.
Cost-of-delay legend: blocks-work (lanes or landings wait on it), money (spend or refund windows), security (exposure or data loss), cosmetic (naming/posture, no mechanism).
1. DefectDojo licence shape, and whether edge-gating moots the purchase
Section titled “1. DefectDojo licence shape, and whether edge-gating moots the purchase”Decision. The three-way shape is landed: apps/defectdojo at 3.2 with no
SSO, apps/defectdojo-classic frozen at 2.58.4 with SSO behind a loud EOL
block. What remains: (a) ratify the shape, (b) is one untested env flip worth an
hour before any Pro purchase.
Evidence (verified).
apps/defectdojo/app.yaml:2:version: "3.2";:14-23now states the measurement in the shipped package itself: “THE GATE IS A WALL, NOT A SIGN-IN … 3.2.0’sAUTHENTICATION_BACKENDSis("django.contrib.auth. backends.ModelBackend",)with no RemoteUserBackend and no RemoteUserMiddleware, so it cannot consume theX-SHC-Userheader the gate sets.” A gated deployment asks for two credentials.:25-28: the gate is opt-in (access:deliberately unset;shc install defectdojo --gated), so nobody inherits that second credential.- The API carve-out is DONE and in the pin: ce
68e0c4b(ungated_paths:), consumed by shc6eb8d56bcfor/api/v2/. - Still untested, the only thing that could revive free SSO:
DD_AUTH_REMOTEUSER_ENABLEDon a 3.x install behind the gate (docs/handoff/2026-08-03-session-close-still-to-do.md§3).
Options. (A) Ship as landed, buy nothing. (B) Test the one flip first, then A. (C) Buy a Pro licence for native SSO.
Recommendation. B then A. The measured default wiring says free SSO is probably dead, but the flip costs an hour and a Pro licence costs money every year. Disprove it before any purchase. C only if a customer demands current-version SSO. The classic EOL block is adequate as written.
Cost of delay: blocks-work (dd worktrees parked); no money at risk until someone buys prematurely.
2. MinIO: is minio-classic simply THE minio?
Section titled “2. MinIO: is minio-classic simply THE minio?”Decision. Both apps shipped. Does apps/minio (console-less, newest)
deserve the default name, or should the catalogue steer everyone to
minio-classic?
Evidence (verified, and now sharper than at filing).
apps/minio/app.yaml:2:version: "2025.9.7";apps/minio-classic/app.yaml:2:version: "2025.4.22".- The un-freeze is live on the estates.
shc search minioon the bitspur leader returnsminio 2025.4.3+0,2025.4.3+1,2025.4.22+0,2025.9.7+0andminio-classic 2025.4.22+0.2025.9.7+0is the build that permanently removes the admin Console, and it is now what an unpinnedshc update minioresolves to. - Both estates are safe today, by pin, not by luck:
minioandnc-s3on bitspur andminioon promanager all readapp_version 2025.4.3+0(shc ls -t corp -o json). Nothing there is unpinned. - A useful consequence of the un-freeze:
minio 2025.4.22+0is also still installable, so “the console line” is reachable under the plain name too, though it is the Docker Hub artefact, whereminio-classicpins the newer quay 2026-03-31 hotfix.
Options. (A) Keep both as landed. (B) Make classic the ONLY minio. (C) Pin
apps/minio to the classic image (two names, one image).
Recommendation. A, as landed. S3-API-only consumers are the majority and
deserve a maintained server; B orphans them on a 2025-04 build forever, and C
destroys the distinction. One follow-up that is not an owner decision: the
catalogue description for minio should name the console loss and point at
minio-classic, because the tile is what a UI user reads.
Cost of delay: cosmetic; both apps ship; only the steering text is open.
3. shc_node’s 8 silent-drop attributes
Section titled “3. shc_node’s 8 silent-drop attributes”Decision. Eight attributes (adopt, private_ip, nebula, fabric,
fabric_cidr, host_domain, modules, acme) accept an edit at plan time,
show “update in-place”, persist to state, and send nothing to the estate. State
then lies. Per-attribute ruling wanted.
Evidence (verified, unchanged at HEAD).
terraform/provider/internal/provider/node_resource.go:907-912: theupdateBootstrapdoc comment: a changedvault_passwordis refused loud,packagerolls the deb,public_addressrepairs metadata,hostre-derives the endpoint; “All other genesis fields are persist-only (init-time).”:920-948: thevault_passwordrefusal is the in-tree precedent, and its text is the model to copy: it names the real verb (shc vault rotate), and it explicitly removes the old “taint the node” advice because that is cluster re-genesis for a password change.docs/reference/legacy-and-breaking.md:173: no attribute has ever silently become a node destroy.
Options per attribute. (A) RequiresReplace: honest but destructive.
(B) refuse-at-plan like vault_password: honest, non-destructive, names the
real mechanism. (C) wire into Update: only where a daemon verb exists.
Recommendation. Do not take A anywhere. B for the six
genesis-consumed fields (nebula, fabric, fabric_cidr, host_domain,
modules, acme); the fabric is cluster-scoped and read back from the genesis
record, so a post-genesis change is a daemon operation. C for private_ip
(a verification-gate input; re-verifying on update is cheap and safe). B for
adopt with a message saying to taint only if the management mechanism truly
changed. One small PR once ruled.
Cost of delay: blocks-work; every estate plan touching these records fiction, and this class has already cost live debugging time.
4. dockerVolumeAbsent matches a bare "not found"
Section titled “4. dockerVolumeAbsent matches a bare "not found"”Decision. A docker volume inspect failure is classified “volume absent” if
stderr contains bare "not found"; absent ⇒ the volume is skipped from the
backup. A real docker fault whose stderr happens to contain “not found”
(plugin, cluster-volume driver) fails open into data omission.
Evidence (verified, unchanged at ce HEAD).
daemon/backup_swarm_capture.go:335-337: the three measured wordings ("no such volume","volume <v> not found","No such volume");:349-353:dockerVolumeAbsentreturns true on"no such volume"or bare"not found".- The two adjacent fixes do not cover it: ce
66e3abb(“a capture that read nothing stops rotating the last good archive away”) and15c3c81(“a capture that read no data stops being a successful backup”) both act on the read result, and an absent-classified volume is skipped before capture, so neither refusal ever sees it.
Options. (A) Keep the bare match, add a loud per-skip log + manifest entry.
(B) Tighten to the name-anchored wordings and treat everything else as a hard
X500447SWMVCP. (C) Both.
Recommendation. C. B restores fail-closed for genuinely ambiguous
stderr; a false hard-fail costs a retry, a false skip costs a backup that
quietly lacks a volume, which is the exact asymmetry 66e3abb/15c3c81 were
landed to kill. A makes even legitimate absence auditable. The table-driven
tests already hold both directions.
Cost of delay: security (data loss).
5. The backup pgdata converter snapshots with the PACKAGE’s image
Section titled “5. The backup pgdata converter snapshots with the PACKAGE’s image”Decision. The postgres builtin’s backup converter boots a server on a snapshot of the data dir using the image the PACKAGE pins at render time, not the binary that wrote the dir. Cross-major that is the binary-vs-data-dir mismatch the update ladder just fixed for live servers, sitting inside backup/restore.
Evidence (verified, unchanged at ce HEAD).
builtin/postgres/backup.yaml:4and:91: both thepgdataconverter andjobs.restorerenderimage: "{{ vars.images.postgres }}".builtin/postgres/vars.yaml:17:majorVersion: "16"; the postgres guard refuses a server on a mismatched data dir, and nothing covers the converter container.- The three legs that ARE guarded: ce
5e79535(pg_upgrade hop + skipped-major refusal), shcf3f4847fb+90d39d87a(pgclient/server major parity). The converter is the unguarded fourth. - Mitigation that exists:
modules/backup/phases.go’sconverterFallbackArchivesre-tars the raw mount and stampsDegradedConverter, so a mismatch degrades rather than losing bytes, but a degraded raw archive restored into a newer package then hits the server guard at restore time, the worst possible moment. - Memory
pgclient_major_must_equal_server: a newer client dumps fine then loads ZERO rows; this failure mode is silent, not loud.
Options. (A) Render the converter with the DEPLOYMENT’s recorded image (now
possible; ce 0a448a0/b23711d record the rendered package version).
(B) Refuse the converter on a major mismatch, keeping the degraded fallback.
(C) Accept the degrade as the design.
Recommendation. A, with B as the refusal when no recorded image exists. The converter’s whole purpose is a logical dump that survives major hops; rendering it with the data dir’s own binary is the only shape that keeps that promise while the package pin moves. C turns every future pin bump into silent fleet-wide backup degradation.
Cost of delay: security (data loss); bounded today only because the postgres pin is deliberately held at 16, and that hold is explicitly temporary.
6. The x9s/HUD direction: land wip/x9s or hold bubbletea?
Section titled “6. The x9s/HUD direction: land wip/x9s or hold bubbletea?”Decision. origin/wip/x9s (88bd572a3, six commits, unmoved since the
first draft) makes shc hud the k9s-UX x9s TUI and retires the bespoke
bubbletea dashboard. It is parked on the release shape: after landing, a
published binary’s shc hud refuses unless .goreleaser.yaml is wired for the
tag.
Evidence (verified).
git show origin/wip/x9s:STATUS.md: every gate green; the two checkpoint defects both fixed on the branch (amake docs/generatepanic:panic: registry: duplicate Commands registration for "hud"atcli/hud.go:57, caused by the docsgen pin being the union of both mains; and a DOA tagged binary). Measured binary cost 126 MB untagged → 175 MB tagged. The tagged build needs credentialed access to the private x9s repo plus an unpublished path-only module, resolved through a gitignoredgo.work.x9shud; untaggedmake buildstays hermetic and network-free.- Two copies exist on the remote:
wip/x9s(rebased, green) andhud-becomes-x9s622fc2f3f(the raw checkpoint, preserved). The main checkout still sits dirty onhud-becomes-x9s, and the owner’s sweeps wipe that checkout (memory:box_auto_reset_job); the branch content is safe on the remote, the checkout copy is the at-risk duplicate. - New since filing, and it changes the arithmetic: a startup panic in the x9s driver is being fixed in a concurrent lane tonight. It is not pushed, so there is nothing to cite yet, which is exactly the point: the branch’s “everything green” claim is now one unmerged fix out of date.
Options. (A) Ship the TUI: wire .goreleaser.yaml (stage x9s in a before
hook, -tags x9shud + GOWORK on the shc build ids), accept +49 MB × 2 OS ×
arches and a private-repo credential in release CI, land, retire modules/hud
after. (B) Land the branch but document shc hud as build-it-yourself, keeping
releases hermetic. (C) Hold, keep bubbletea.
Recommendation. A, after the driver-panic fix lands on the branch. B ships a regression in a working shipped feature; C rots a finished refactor while the bespoke dashboard (21,773 lines, a named gut-class candidate) keeps accruing maintenance. The 49 MB and the release credential are the entire remaining cost, both one-time. Rule A now and let the lane rebase onto it.
Cost of delay: blocks-work; a finished lane is parked and the dirty main-checkout twin invites a sweep accident or a divergent re-land.
7. ★ CORRECTED: the vault password is readable by any local user
Section titled “7. ★ CORRECTED: the vault password is readable by any local user”Decision. Every bootstrap node carries the vault password in a systemd drop-in so the daemon auto-unseals on restart. The previous draft framed this as “plaintext on disk, so harden to 0600 and prove the support bundle excludes it”. Both halves of that framing were wrong. Measured tonight on both leaders.
Evidence (measured tonight, both estates).
- The drop-in is already
600 root:root:/etc/systemd/system/shc.service.d/vault-autounlock.conf, oneEnvironment=line. As the unprivilegedadminuser (uid 1000, no sudo),catof it → Permission denied. So the disk half is already hardened. - But the 0600 is defeated by systemd itself. As the same uid-1000 user,
systemctl show shc.service -p Environmentreturns a 295-byte value whose variable list includesSHC_VAULT_PASSWORD=. Reproduced identically on promanager204.168.129.34and bitspur37.27.215.210. Any local unprivileged account on a bootstrap node can read the vault master password over systemd’s D-Bus property; no sudo, no file access, no exploit. - The support-bundle worry was a red herring:
modules/supportbundle/*.gocontains nosystemctlorjournalctlinvocation at all (grepped), so the bundle never captured a unit dump. The leak was never in the bundle. - Context that keeps the drop-in justified: memory
vault_locks_on_daemon_restart: every bareshcdrestart or reboot leaves the vault LOCKED, a degraded control plane until a human intervenes; and v0.0.16 already fixedshc vault rotateleaving the retired password in the drop-in.
Options. (A) Accept; a host that is compromised loses everything anyway.
(B) Remove the drop-in, rely on peer-unseal + manual cold-start unlock (ce
d93f3137c, proven on both clusters 2026-07-16). (C) Move the value off
Environment= (an EnvironmentFile= pointing at a 0600 root-only file, or
LoadCredential=) so the unit property carries no secret.
Recommendation. C, and treat it as a defect rather than a posture
question. A is now indefensible: “compromised host” is a different threat from
“any local account, including a service account in a container with a host
uid”. B re-introduces the whole-cluster-power-cycle trap the drop-in exists to
solve. C keeps the availability property and removes the exposure, and it is a
one-line unit change plus a renderer change, but verify at implementation
that the chosen mechanism really is absent from systemctl show output, because
that assumption is what failed here.
Cost of delay: security, and now an active exposure on both production estates, not a stable known posture. This is the item whose severity changed most.
8. Second registry for the image-mirror SPOF
Section titled “8. Second registry for the image-mirror SPOF”Decision. The proprietary keycloak/traefik/forward-auth/license-broker images exist only in the GitLab registry. The probe half is live; the mirror half waits on owner action.
Evidence (verified, and one correction).
docs/handoff/2026-07-30-open-core-enforcement-architecture.md:565-604: the probe is landed:apps/verify-pinsruns on release tags and a daily schedule alone (.gitlab-ci.yml:47,:180), refuses (exit 2) on any root yielding zero pins, and its first full run measured 129 unique pins, 219 occurrences, 92 files; all pullable, including all four SHC-built system images.- Correction to the first draft, which said “There is no engineering left;
the CI job exists.” There is no
MIRROR_REGISTRY/MIRROR_REGISTRY_AUTHanywhere in.gitlab-ci.yml(grepped): the mirror-copy job does not exist. §9.9 item 2 states what it needs: the two CI variables plus one addition to theimageslane (a secondproducttagsentry in docker-bake.hcl, or a post-bakecrane copy). Small, but not zero. - The licence posture is settled and makes mirroring free on every tier: a
same-digest ref classifies as a mirror, no licence consulted
(
core/imageref.Mirrorsrule 1).
Options. (A) Create a non-GitLab registry account, set two CI vars, add the copy step. (B) Accept the SPOF.
Recommendation. A, still the cheapest item on the docket, just not free. Twenty minutes of owner time plus an afternoon of CI work closes a single-point-of-failure on the images every paying estate boots from. Prefer quay or ghcr over Docker Hub (rate-limit posture).
Cost of delay: money/availability; a GitLab registry incident during a customer install window is a support fire with no fallback.
9. Offline custody for the two vendor signing seeds
Section titled “9. Offline custody for the two vendor signing seeds”Decision. Both vendor Ed25519 signing seeds (the licence trust root, primary and reserve) still live only on this dev box. Owner custody has been owed since 2026-07-31.
Evidence (measured tonight).
/home/admin/.shc-vendor-signing-key.seed:600 admin:admin, 45 bytes, mtime 2026-07-29 20:53 UTC./home/admin/.shc-vendor-signing-reserve-key.seed:600 admin:admin, 45 bytes, mtime 2026-07-30 00:58 UTC. Both present, same disk, same user.- The reserve-key path is proven end-to-end (memory: licensing dogfood, Phase 0), so seed loss is survivable only while the reserve survives, and it does not survive independently.
- This box also runs the owner’s sweeps and holds live cloud/GitLab credentials
(memory:
token_credential_locations), i.e. it is the most-handled machine in the estate.
Options. (A) Owner stores both in the password manager plus one offline copy in a different building, then either removes them from this box or knowingly leaves a working copy. (B) Status quo.
Recommendation. A, this week. A disk failure or compromise here is a full trust-root event: every customer licence file, the broker and the embedded pubkeys would need coordinated rotation. Fifteen minutes removes the existential half. No code involved.
Cost of delay: security: highest single-event severity on this docket.
10. LicenseRef-SHC-EE text + ce’s missing contributor grant
Section titled “10. LicenseRef-SHC-EE text + ce’s missing contributor grant”Decision. The EE licence identifier is a placeholder with no reviewed text behind it, and ce has no contributor licence grant at all, so the first outside contribution lands AGPL-only and cannot lawfully link into the proprietary binary.
Evidence (verified tonight).
.goreleaser.yaml:390-394: “The product is proprietary. LicenseRef-SHC-EE is the placeholder …license: LicenseRef-SHC-EE (proprietary)”.- ce
origin/mainroot holds exactly one licence-ish file:LICENSE. There is noCONTRIBUTING.md(git ls-tree --name-only HEAD). - The CE/EE posture itself is settled (memory:
ce_ee_open_core_split: AGPL open-core, private superset → public mirror), so this is execution.
Options. (A) Before ce publication: commission the EE text and add a ce CONTRIBUTING.md with a grant (CLA, or the lighter DCO + grant). (B) Publish first, backfill later.
Recommendation. A: treat the grant as a publication gate. B is how projects end up unable to relicense a file a drive-by contributor touched, and the after-the-fact fix means chasing signatures. Minimum viable is one page: DCO plus an explicit “contributions may be incorporated into proprietary editions”. The EE text can start from a standard EULA template and be polished before first sale rather than before publication, but the grant must precede the first external MR.
Cost of delay: money (legal): zero until ce publishes or a stranger contributes; step-function expensive at that instant.
11. Hook-family rename wording
Section titled “11. Hook-family rename wording”Decision. The ce hook family (SetEditionCommands, SetEditionAddNote,
EditionValidate*) needs a name for what the community binary calls the paid
product.
Evidence (re-measured tonight).
git grep -o -E 'SetEditionCommands|SetEditionAddNote|EditionValidate[A-Za-z]*'counts 34 in shc, 42 in ce (76 total).- Correction: the first draft claimed 48 ce / 34 shc and “compounding, 72 →
82 in four days”. Under this measurement shc is unchanged and ce fell from
48 to 42; the legacy sweep (
26b96fa3b) removed residue. The urgency argument from growth does not hold; the correctness argument does. - The reason it still matters: the de-editionize ruling already renamed the
wire key edition→product (
cbeb4f6d1), so the code and the wire disagree in spelling today.
Options for the word. “product” (matches the wire key and the single-product ruling), “licensed”/“paid” (describes the boundary), keep “edition” (contradicts the ruling in the code that implements it).
Recommendation. SetProductCommands family: align with the wire key
already ruled and shipped. One mechanical rename per repo, done inside a
pin-bump window so the seam moves atomically. Decide the help-copy noun in the
same breath (“SHC Pro/Enterprise features” reads fine and invents no brand).
Cost of delay: cosmetic, and no longer compounding.
12. The scrub-spec value-boundary fork
Section titled “12. The scrub-spec value-boundary fork”Decision. Option A (minimal per-shape structural peek) is implemented and
shipped as pkg/scrub; option B (one fixed literal boundary) honours the
no-parser ethos at a measured precision cost. §19.9 also still wants the spec
table ratified.
Evidence (verified tonight).
pkg/scrub/on shcorigin/main:scrub.go,spec.yaml,ansi.go,README.md, tests.spec.yaml’s header states the ruling and the hard invariants it encodes: no bare value-shape rule (a bare 32+ hex match ran ~76% false-positive on the corpus: build hashes, x509 SAN labels, git SHAs), never redact aref+vault://value even under a secret-named key, and scrub runs after ANSI-strip because hulykvs separates a DSN key from its value with SGR escapes.docs/internal/secret-surface-audit.md:302-305: §19.9 is explicitly “a ratify-this-table” item covering the redaction literal, theref+exclusion and the no-bare-hex invariant.- The architecture half is already settled (§22.4: scrub-scanning is not parsing).
Options. (A) Ratify what shipped: structural peek + the spec table. (B) Swap to the fixed literal boundary.
Recommendation. A. The corpus said B’s imprecision matters and that over-redaction “trains people to distrust redactions”; A’s sliver of parsing is four bounded scanners with tests, and A is what runs. Ratify the table in the same ruling; it is a read-and-nod, not a design session.
Cost of delay: cosmetic; the shipped default is the recommended one.
13. ★ CORRECTED: Hostinger: claim the refund, revoke the pasted token
Section titled “13. ★ CORRECTED: Hostinger: claim the refund, revoke the pasted token”Decision. Two owner actions, neither a code task.
Evidence (verified, and the urgency is different from the first draft).
- The renewal bleed is already stopped.
docs/handoff/2026-07-31-execution-marathon.md:1971-1973: Hostinger “sits exactly at its expected contract: two VMs stopped, both subscriptionsnon_renewingthrough 2026-08-31”. The first draft said a forgotten subscription “renews monthly forever”; it will not. That claim came from the capability gap, not the current state. - What is genuinely still owed, verbatim from memory
hostinger_cloud_support: “(1) cancel both subs in hPanel Billing→Subscriptions for the 30-day money-back refund (~$21.30); (2) revoke the ORIGINAL Hostinger token (pasted plaintext in chat, exposed); a replacement full-scope token is already in the .tokens store.” - The API cannot do either:
DELETE /billing/v1/subscriptions/{id}404s and there is no VPS-delete route (405). hPanel is the only door.
Options. Claim the refund in hPanel now, or let the window lapse and keep $21.30 spent on two blank stopped VMs that expire 2026-08-31 anyway.
Recommendation. Do both in one hPanel session. The refund window from the ~2026-07-31 purchase closes ~2026-08-30, the only hard calendar deadline on this docket. The token revocation rides along in the same login, and it is the half that actually matters: an exposed full-scope token on a provider where a leaked VPS is the worst leak class we support.
Cost of delay: money (a closing refund window, no longer a recurring charge) + security (exposed live token until revoked).
14. ★ CORRECTED: the four billable Hetzner primary IPs are not where the sweep said
Section titled “14. ★ CORRECTED: the four billable Hetzner primary IPs are not where the sweep said”★ CLOSED 2026-08-05: PHANTOM. The owner confirms: “I do not have a third hetzner project.” Combined with the measurement (20 primary IPs across both known projects, ZERO unassigned, all on the
2a01:4f9prefix) there is nothing billable idling anywhere the owner owns. Only an idle primary IP bills separately; an assigned one is included with its server.
2a01:4ff:f0:3df7::/64and the “exdeb project” were never findings. A lane reported an empty search; that was restated as “they must be in a third project no credential reaches” (a cause invented to preserve the premise), and then over-corrected to “they do not exist”, equally unmeasured. The lesson, since this pattern recurred three times in one session: when a search comes back empty, say the search came back empty. Do not promote absence into a hidden cause or into a proof of absence. No further work is owed on this item.
Decision. 178.156.157.203, 178.156.191.56, 5.161.67.161,
2a01:4ff:f0:3df7::/64, unassigned, billable-while-reserved primary IPs the
2026-07-31 sweep refused to classify (221–223 days old, zero labels, no SHC
marker). Keep or release.
Evidence (measured tonight, read-only, against both reachable projects).
docs/handoff/2026-07-31-execution-marathon.md:1961filed them as “4 unassigned Hetzner primary IPs on bitspur”.GET /v1/primary_ipswith bitspur/ops’ CIHCLOUD_TOKEN: 8 primary IPs, every one ASSIGNED, oldest created 2026-05-19. None of the four.- The same call with promanager/ops’
HCLOUD_TOKEN(identical toHCLOUD_TOKEN_US_EAST): 12 primary IPs, every one ASSIGNED. None of the four. - Their address ranges support a third project: the two managed projects’ IPv6
are all
2a01:4f9:…, while the orphan is2a01:4ff:f0:…. The likely home is the exdeb project, whose token was embedded in the now-deleted promanager-infra repo and which memory already flags for rotate-or-delete-in-console (promanager-ops/.tokens:38-41).
Options. (A) Assume they were released since the sweep and close the item. (B) Owner logs into the Hetzner console, finds the project that owns them, and releases anything unreferenced. (C) Keep paying.
Recommendation. B, and fold in the exdeb token. A is tempting and might be right, but “not in the two projects I can reach” is not “released”, and the sweep recorded them as unassigned, so nothing in a managed estate references them by construction. One console login answers both this and the owed exdeb credential rotation. Whatever the owner deliberately reserved should get a label, which is what would have exempted it from the sweep in the first place.
Cost of delay: money: small, perpetual, and now unmeasurable from CI, which is its own reason to close it.
15. bitspur-ops’ git remote carries a live glpat- token. It dies 2026-08-11
Section titled “15. bitspur-ops’ git remote carries a live glpat- token. It dies 2026-08-11”Decision. The bitspur-ops checkout’s only credential is a glpat- token in
plaintext in its git remote URL. Rotate, use, or accept.
Evidence (re-probed tonight, read-only).
/home/admin/selfhosted-cloud/bitspur-ops/.git/configremote URL carries the token; the repo has no.tokensfile (ls→ No such file).GET /api/v4/personal_access_tokens/self→ 200: name “Tmp8”,active: true,revoked: false,expires_at: 2026-08-11, scopes include fullapipluswrite_repository,write_registry,manage_runner,k8s_proxy.- It remains the box’s working GitLab PAT and the practical API path to the
parked estate pipeline jobs. Tonight it was used for read-only GETs only:
the token self-check, two projects’ CI-variable name lists, and, for
item 14, two
HCLOUD_TOKENreads feeding read-only Hetzner list calls.
Options. (A) Use it for the remaining parked pipeline work before it self-expires, then mint a properly scoped replacement and scrub every plaintext copy. (B) Revoke now and mint fresh first. (C) Accept.
Recommendation. A, with the replacement scheduled for 08-11, not
“later”. The exposure clock has 7 days left by construction and the
capability it uniquely holds is what the estate rolls were blocked on. After
08-11: one deliberately-scoped PAT via git-credential-store (never the remote
URL), and scrub the four dead plaintext copies, expired or not; an expired
secret in a config file still teaches the habit. C is strictly dominated.
Cost of delay: security (live full-api token in a config file)
+ blocks-work (the parked pipeline jobs).
16. Retire promanager’s old license broker
Section titled “16. Retire promanager’s old license broker”Decision. v0.0.12 baked https://license.selfhostedcloud.org as the default
broker; the old license.corp.promanager.vip was kept up on the condition
“retire it only after both estates are on v0.0.12 and verified against bitspur”.
That precondition is long met; the estates are on v0.0.19.
Evidence (measured tonight).
- DNS:
license.corp.promanager.vip→ 204.168.129.34 (the promanager leader);license.selfhostedcloud.org→ 37.27.215.210 (the bitspur leader). Two hostnames, two different estates. - Both are live:
GET /healthzon each returns HTTP 200{"status":"ok"}(LE staging,-k). - A
license-brokerdeployment is running incorpon both estates, both atapp_version 0.1+7(shc ls -t corp -o json). - No customer binaries predate the cutover: trials end 2026-08-26 and there are no sales (memory: licensing dogfood), so nothing external resolves the old name.
- One caveat to verify at execution, not before ruling: “both keygens hold both estates’ records” should be confirmed by a read-only keygen inventory against the LIVE brokers, not from memory (2026-07-31 state: both production records on the promanager keygen; bitspur’s initially empty, then seeded).
Options. (A) Retire: point anything residual at the public broker, stop the corp stack, keep the DNS name as a redirect for one release cycle. (B) Keep both indefinitely.
Recommendation. A, gated on that one live inventory read. Two brokers is drift surface, and the pagination P0 already taught that the live broker can silently diverge from the declared one. One broker, one truth. Keep the old hostname resolving as a redirect until the next release, then drop it.
Cost of delay: money/cosmetic: a running production stack nobody should be using, plus standing drift risk.
17. The sso two-door entitlement question, now with a second protocol
Section titled “17. The sso two-door entitlement question, now with a second protocol”Decision. The sso entitlement is enforced at two doors encoding different
theories. Door 1 (daemon, ACTIVE) refuses a NEW identity edge to a
non-bundled provider. Door 2 (Keycloak SPI FederationGuard, INERT by
default) watches federation into SHC. §4c framed the paid boundary as door 2
only; R22 and license.go frame sso as free↔paid generally.
Evidence (verified tonight; the mechanism moved in door 1’s favour).
shc modules/app/sso_gate.go: door 1 registerseeIdentityEdgeGateviaSetIdentityEdgeGate: bundled keycloak and any existing edge (re-converge or recovery replay) always pass; only a NEW external edge consults the data-driven resolver, and it re-reads the live posture before refusing so a stale snapshot can never disconnect working auth.- New: door 1 already generalized to two protocols. The SAML socket landed
(ce
9ed5af5), and both enforcement sites now gate onIsIdentitySocket(...):ce modules/app/sso_gate.go:66-73, which isSocketSSO || SocketSAML.integration_service.go:157-164says why in as many words: “The membership test isIsIdentitySocketprecisely so this door and the render door (gateIdentityEdges) cannot drift apart when a third protocol lands.” The unlicensed-SAML-bypass hazard thatdocs/decisions/2026-08-05-oidc-and-saml-sockets.md:153warned about is already closed, and latent anyway: only the bundled keycloak servessamltoday (builtin/keycloak/saml.socket.yaml; no app underapps/serves a saml socket). - Product context: the free tier’s core promise is bundled-Keycloak SSO, which never gates under either answer; huly/matrix native SSO and the auto-redirect campaign are unaffected.
Options. (A) Paid = external IdP anywhere: keep door 1, leave door 2 inert
or advisory. (B) Paid = federation into SHC’s own login only: delete door 1’s
refusal, external-oidc edges become free, arm door 2. (C) Keep both armed.
Recommendation. A. It is the door that works, sits in Go where it is
testable, and has now proved it generalizes; door 2 is a Java SPI with no
pre-commit admin-API interceptor, and Keycloak theme-hiding is cosmetic.
“Bring your own IdP” is a coherent paid line MSP-shaped customers understand. B
gives away the only enforceable half. Whatever is ruled, fix the three
disagreeing surfaces in one commit: license.go’s comment, §4c, and the now
false doc line at ce modules/app/sso_gate.go:76-77 (“Socket is … always
adapter.SocketSSO when the gate fires”), which IsIdentitySocket made untrue.
Note license.FeatureSSO = "sso" must not be renamed: it is baked into every
minted licence.
Cost of delay: blocks-work; the FederationGuard flag, the entitlement copy and the pending oidc-socket rename sweep all inherit this ambiguity.
18. ★ NEW: what does it MEAN to pin a system app? (the postgres deadlock)
Section titled “18. ★ NEW: what does it MEAN to pin a system app? (the postgres deadlock)”Decision. The system-pin row patch is finally runnable, but the spelling it should use is undecided, and the two available spellings mean opposite things. Ruling wanted on which is the pin.
Evidence (measured tonight on both estates).
- The blockers are gone. Both estates are on v0.0.19, so ce
0a448a0’s precedence fix is live; and ce3884c01(in the pin at425bc61b1) removed the reason to fear the re-render:builtin/postgres’spre_updateranscale{postgres: 0}unconditionally, so every same-major re-render stopped keycloak’s store for no upgrade.pg-upgrade-stagenow emits::shc-hook:halt::on both no-hop paths and steps 2–3 never run. That commit also measured the running imagepercona/percona-distribution-postgresql:16and the shipped16.14@sha256:d329970c…as the same manifest digest. - Nothing has been pinned.
shc ls -t master -o json: promanagerpostgresandopenobserveare NULL; bitspurpostgresreads16(written by the refused v0.0.16 update; the class ce8ab2081has since made impossible),openobserveNULL;keycloakandtraefikread the literallateston both. Four stacks, three row shapes, zero real pins. - The ambiguity is real and measurable.
shc search postgreson both leaders returns seven rows matching a bare16:16+0 … 16+5in reposhc, plus16inbuiltin. - The two spellings mean opposite things
(
ce modules/app/source_qualifier.go:93-126): a pin naming exactly what the shipped root declares short-circuits to/usr/lib/shc/apps/<app>ahead of the catalog, so pinning16records a version while still re-resolving to whatever the deb last dropped, which is the very defect the pin exists to close. Only a<version>+<rev>pin reaches the catalog and holds a package rendering, and the master-tenant guard refuses a repo qualifier outright (X403979SYSSRC), which is why precedence was the only place this could be fixed at all.
Options. (A) Pin the bare version (16): cheap, safe today, but a pin in
name only. (B) Pin the catalogue spelling (16+5): a real hold, but a spelling
a shipped app.yaml never carries, and no candidate is known to reproduce the
running rendering. (C) Give system apps a repo qualifier and pin builtin/16:
closes it properly, and is the same gap the catalog needs anyway
(docs/proposals/catalog.md §9.3 calls X409039APPAMB “the live
system-postgres deadlock”).
Recommendation. C as the mechanism, A as the interim. With 3884c01 in
the pin, applying 16 today is a restart-free no-op re-render, so take it now
to get the rows off NULL/latest, but only as a placeholder, and postgres
only: pinning keycloak 26.6 takes the 26.6.1 → 26.7.0 image hop, because
ce moved the builtin images without moving their app.yaml version (same for
traefik 3.6.24 → 3.6.25). Then close it properly with C, which is also what
unblocks postgres 17/18 and gitlab 19.2.1. Note the pin is bypassed by four
unattended paths that pass current = nil (traefik ACME reconcile, keycloak
basePath, openobserve basePath/SSO, CA-rotation cold consumers); a pin is not a
freeze until those are addressed.
Cost of delay: blocks-work (postgres 17/18, gitlab 19.2.1) + security; the control-plane auth database currently re-resolves to whatever the deb ships.
19. ★ NEW: should repo priority break the tie instead of X409039APPAMB?
Section titled “19. ★ NEW: should repo priority break the tie instead of X409039APPAMB?”Decision. This is docs/proposals/catalog.md Q1, and it is the
load-bearing one. Every estate has at least two repositories: the implicit
builtin crawled from /usr/lib/shc/apps (no URL, cannot be unregistered) plus
registered HTTP repos, so the same app legitimately exists twice, and today
that is an error rather than an ordering.
Evidence (verified).
docs/proposals/catalog.md§12 Q1: theprioritycolumn “exists, is indexed, and the schema comment says repos are ordered by it. YetX409039APPAMBrefuses when two repos match.” Either priority is used for something else, or it is an unused disambiguation lever. Marked “Load-bearing, because it is blocking postgres 17/18.”ce modules/app/source_qualifier.go:128-130: the refusal is a barelen(res.Repos) > 1test; nothing consults priority.ce daemon/services_seams_repo_crawler.go:306: the refusal was a deliberate replacement for “the old silentORDER BY”, i.e. priority-ordering is exactly what was taken away, on purpose.- Measured consequence, live: seven rows match a bare
postgres 16across two repos on both estates (item 18).
Options. (A) Priority breaks the tie silently (restores the old behaviour the refusal replaced). (B) Priority breaks the tie and says so in the operation’s output. (C) Keep refusing, and add a repo qualifier for system apps (item 18’s option C).
Recommendation. B plus C, not either alone. A is what 306 deliberately
undid; a silent pick is how an operator installs a different app than they
meant. But a hard refusal with no lever is what deadlocked postgres, and a
catalog UI over many repos will hit it on every duplicated tile. Priority is the
right default resolution as long as the choice is stated; the explicit
qualifier is the right override. Ruling this unblocks item 18, postgres 17/18,
gitlab 19.2.1 and the catalog’s install action at once.
Cost of delay: blocks-work; it is named as the blocker in three separate places.
20. ★ NEW: six corp stacks have no safe pin, and one has none at all
Section titled “20. ★ NEW: six corp stacks have no safe pin, and one has none at all”Decision. After the roll, five stacks per estate are pinned at exact
version+revision. Six are not, deliberately, because no available pin equals
their running image. Ruling wanted: leave them unpinned and never update them,
or publish a package that matches what they run.
Evidence (measured tonight, shc ls -t corp -o json on both leaders).
- Unpinned (
app_version: "latest"), three per estate, six total:affine,affine-db,huly-dbon promanager and the same three on bitspur. affineis the one with no safe pin at all. It runsghcr.io/toeverything/affine:0.26.5(measured viadocker pson bitspur).shc search affineoffers0.26+0,0.26+1,0.27+0,0.27+1,0.27+2, and the version token and the image drifted apart: shc9b13d6874movedapps/affine/vars.yamlfromaffine:0.26.5to0.27.3@sha256:d9d145f9…without changingapp.yaml’sversion: "0.26"(verified:9b13d6874~1:apps/affine/app.yamlreads0.26, and the commit’s diff touches noversionline). So some0.26+Npackage carries the 0.27.3 image, and the index cannot tell you which. Guessing wrong is a one-way AFFiNE 0.26→0.27 data migration.- The two
postgresstacks are milder:affine-dbandhuly-dbboth run the bare tagpercona/percona-distribution-postgresql:16, while every candidate pins16.14@sha256:d329970c…; measured same-digest today (ce3884c01), so a pin is image-identical now and drifts the moment upstream retags:16. - The standing hazard: an unpinned
shc updatere-resolves to the catalogue head (isNewestSelectortreats""/latest/stablealike), which is exactly howminio 2025.9.7+0would cost an operator the admin console.
Options. (A) Leave all six unpinned and hold the operational rule “never
re-render an unpinned stack”. (B) Publish a revision that reproduces each
running image (an affine 0.26+N pinning 0.26.5 by digest), then pin.
(C) Pin the two postgres stacks now (same digest, no risk) and treat affine
separately.
Recommendation. C now, B for affine. The two postgres stacks are a
free win: pinning them costs nothing today and removes them from the
never-touch list. For affine, A is a rule enforced only by memory, and the
whole reason this docket exists is that memory-enforced rules fail. Publishing
one digest-exact 0.26 revision converts “never update this” into “pinned”,
which is a property the system can hold on its own. Also worth ruling in the
same breath: the version-token-vs-image drift at 9b13d6874 is the mechanism
behind both this and item 18’s keycloak/traefik surprise; a guard that refuses
an image bump without a version bump would close the class.
Cost of delay: security (data loss); an unpinned affine is one shc update
away from a one-way migration on a stack holding production content.
21. ★ NEW: the catalog design’s open questions Q2–Q8
Section titled “21. ★ NEW: the catalog design’s open questions Q2–Q8”Decision. docs/proposals/catalog.md §12 numbers eight open questions.
Q1 is item 19 above (load-bearing, blocking postgres 17/18). Q2–Q8 want
rulings but block nothing today; they are read-and-decide, not
re-derivations, and they are stated in full at the source rather than restated
here.
Evidence (verified). docs/proposals/catalog.md:557-599 holds all eight,
each with the section that argues it: Q2 visibility defaults (§5.1 argues for
deriving the default from installability, which inverts the owner’s
proposal for the vapp/vpreset pair); Q3 where vapp/preset data comes from
(§6: archive fetch ≈1.4 MB and keeps repositories dumb, vs. index projection);
Q4 whether app-level presets are worth having (§4.4; zero exist);
Q5 whether the crawler should emit vapp rows (needs no migration);
Q6 whether a preset may declare vapp instances (§4.3; changes what a catalog
item is, so it must be decided before the projection is built);
Q7 whether outbound LDAP is a paid boundary (distinct from the inbound
federation boundary, and distinct from the ldap socket already ruled no,
docs/decisions/2026-08-05-ldap-socket.md); Q8 whether shc preset search
already fetches archives.
Options. (A) One sitting over §12, in order. (B) Rule only Q1 (item 19) and Q6 now, defer the rest until the projection is built.
Recommendation. B. Q1 unblocks live work and Q6 changes the data model, so both must precede any implementation. Q2–Q5, Q7 and Q8 are genuinely cheaper to answer once something exists to look at; deciding them now risks ruling on a shape nobody has held. One correction to fold in whenever §12 is next touched: Q4’s headline example is flagged as unbacked in the doc itself and should be fixed either way.
Cost of delay: cosmetic for Q2–Q5/Q7/Q8; blocks-work for Q6 if the projection starts before it is ruled.
22. Jellyfin’s checkpoint shape: bless the pattern, or converge it?
Section titled “22. Jellyfin’s checkpoint shape: bless the pattern, or converge it?”Decision. The 10.11 one-way EF-Core migration landed behind a refusing gate
(4d946ae94). What may remain is the checkpoint SHAPE: archive retention, and
whether this becomes the house pattern for one-way migrations.
Evidence (verified at HEAD).
apps/jellyfin/vars.yaml:30-41:preUpgradeBackup: "auto" | "skip".autoarchives/configto thejellyfin-preupgradevolume before a migrating version may start and refuses to start it if the archive cannot be written;skipprints the warning and takes nothing, and is deliberately not the default because it “is an assertion about a backup nothing here can verify”. The gate no-ops once the library is on the EF-Core schema.- Same file: the healthcheck was moved off
/System/Info/Public, which 10.11’s loading placeholder answers 200 within a second of container start. - The alternative now exists: ce
bd08eb9makes apre_updaterefusal abandon the update, so future one-way migrations can refuse from the platform instead of from a compose service.
Options. (A) Bless the shipped shape as-is (first-written archive wins,
never overwritten, explicit skip waiver). (B) Converge onto pre_update
refusals now that they gate. (C) Both: keep jellyfin, standardize on
pre_update for new ones.
Recommendation. C. The compose-service gate was built because it was the
only refusal mechanism app packaging had; rewriting a proven, live-rehearsed
gate to use a newer seam buys nothing and risks a working upgrade path. But new
one-way migrations should reach for pre_update, which refuses before anything
starts rather than inside the boot. Retention as shipped (kept forever, never
overwritten) is the right default for a one-way migration nobody can redo.
Cost of delay: cosmetic; the mechanism is shipped and rehearsed; only the house-standard wording is open.
Closed since the first draft (verified in git, with shas)
Section titled “Closed since the first draft (verified in git, with shas)”All of these were on the docket or its still-to-do list hours ago. Each was re-checked against today’s trees, not taken on report.
Mechanisms, ce (all inside shc’s pin c99b51c2865f / 425bc61b1):
- Quota check-and-charge atomicity: ce
fa02ee9(“8 concurrent 500m charges under a 2000m cap admit exactly 4”) with the shc twin4f696c843; pin-bumped50dbfa725. pre_updaterefusals now abort: cebd08eb9(“apre_updaterefusal now ABANDONS the update, instead of degrading the stack”); docs76589ce25.- A refused update is a no-op on the row: ce
8ab2081(“instead of leaving the version pin it never applied”); pin9c1930f15. This is what makes bitspur’s straypostgres = 16row (item 18) a one-off rather than a recurring class. - Bare
ref+vault://at the SYSTEM tier: ced89ca3b; shc guards aligned6a7c548ae; and ce224a4beremoved the shippedX200972VLTENAtext that still described a fallback that no longer exists (pin4f15c894a). ungated_paths:: ce68e0c4b; first consumer shc6eb8d56bc(defectdojo/api/v2/).- The DNS teardown window: ce
7116370(“closes from both ends; no resurrect, no unreachable zone”); pinc64fadf6a. - PersistedEdgeState + recover’s WireIntegrations self-heal: ce
988d4c2(“every-iedge persists its bind-time plug context, secrets as raw templates, never resolved values”) and ce5a52e90(“self-heals the unwired registry and stops abandoning sibling edges”). jobs.unintegrationRUNS at teardown: ce63e44b6(“22 authored teardown jobs, 19 of them security state, executed against the frozen bind-time edge”); extended to the disconnect leg by ce45aae15(-i <ref>!finally REVOKES); shc docs corrected35db6969e(the teardown phase has a runner on BOTH legs, andcleanupwas never a plug/socket phase).- Uninstall reaps the stack access role: ce
7c2ddcb(“so a reinstall inherits no stale grant”). - The node-removal gate on a
deployment_nodeset: ce397823f(“stops permitting the node that runs the work”); pin56addc8fc; decision recordede12be16cf+docs/decisions/2026-08-05-deployment-node-set.md. - The dead task worker (lifecycle Start ctx): ce
9c805ce(“a subsystem’s Start ctx no longer dies because a sibling shares its priority”); pin8b881755f. - A SCHEDULED backup seals into
backup.repoencrypted: cede396a7(“finally reaches the operator’s bucket, encrypted”), with ce729a649making the encrypting engine freeze the stack to take it; retention stops orphaning restic archives: ce66e3abb(“a capture that read nothing stops rotating the last good archive away”) and15c3c81; shipped as v0.0.19 (9ba3fddc6), with the falsified zero-outage claim retracted7d1c2dd5a. - The quiesce bracket’s durable journal (SIGKILL self-heal): ce
27bc51f(“a SIGKILLed capture no longer leaves an app read-only forever”), on top of ce8111eb3(a FAILED backup no longer leaves the database permanently read-only).
Apps and integrations, shc:
- mailcow
server_urlfrom the socket + the mail-activation attribute: shcef478ea2a(“the identity provider learns its address from the socket, and stops reporting a config it never stored”) and84ba14954(“the one mapping pair that makes an admin’s mail switch mean something”), with cec75f8f3on the keycloak side. - The SAML socket, and mautic converted onto it: ce
9ed5af5(“keycloak serves a saml socket, and the plug’s socket key stops being decoration”); shc954d14205(“mautic’s SAML SSO works, on a real saml socket, with no parity secret”); design and traps in367003bf1/docs/decisions/2026-08-05-oidc-and-saml-sockets.md. - LDAP ruled NO:
b2b9b2c10,docs/decisions/2026-08-05-ldap-socket.md(“demand is zero and the pruned face served every realm anonymously”). Note the outbound half survives as catalog Q7 (item 21). - The frappe baked-vs-vapp modes:
d5e936e5f(“apps/is either a volume or the image, chosen at install and refused on update”). - frappe assets and
/files:07ace7b62(“gunicorn serves the assets its own pages ask for”) andbca49442c(“/filesserves through a symlinked host; the loader stops comparing a resolved path to an unresolved prefix”), with the suite fixed to stop passing over a 404ing login page517560ea6. - netbird 0.76.1:
8d0743e86(“the blocker was an encoding bug, not an autogen limit”). This closes the netbird half of the first draft’s item 18 outright: the 32-byte-key blocker was not a real autogen limit, so neither the new autogen mode nor the data-loss waiver is needed.apps/netbird/app.yamlnow readsversion: "0.76". - The sso plugs that reported a login they do not own:
08854d51d(nextcloud + frappe), shipping in v0.0.20 (894d2a186).
Estates:
- Both estates rolled to v0.0.19: measured tonight:
shc versionreturns0.0.19on the promanager leader (204.168.129.34, deb1:0.0.19) and the bitspur leader (37.27.215.210). This closes the “roll before patching a system-stack version row” blocker carried since v0.0.18, and is what makes item 18 actionable at all.
Already closed before the first draft, restated so they stay closed:
telemetry routing (fan-out ruled, most-specific-wins rejected); Hobby
core-metering and the broker hostname (ruled 2026-08-01); the shc update
replay P1 (withdrawn on control evidence 478236568; both rehearsals had
measured a stale Jun-27 binary); the minio-classic / defectdojo-classic
“decide before landing” gate (bf3ba1d64, 316e6e486, 22b5ea619; the
residual product questions are items 1 and 2).
What changed in this refresh, in one place
Section titled “What changed in this refresh, in one place”One item closed, nineteen mechanisms closed under it. Six items survive with their evidence intact (1, 2, 3, 4, 5, 12), five survive with corrected evidence (7, 8, 11, 13, 14), five survive re-argued on stronger ground (6, 9, 10, 16, 17), the old item 18 split (netbird closed, jellyfin became item 22), and four items are new (18–21).
The three corrections matter more than the closures:
- Item 7 got worse, and for a different reason than filed. The drop-in is
already
0600 root:root, butsystemctl show shc.service -p EnvironmenthandsSHC_VAULT_PASSWORDto any unprivileged local user on both production leaders. The support-bundle exclusion the draft wanted proved irrelevant: the bundle never captured systemd at all. - Item 13 got less urgent and stayed urgent for the other half. Both
Hostinger subscriptions are already
non_renewingthrough 2026-08-31, so there is no perpetual charge. The closing refund window (~2026-08-30) and the exposed token are what is left. - Item 14 could not be reproduced. Neither reachable Hetzner project contains the four IPs, and all 20 primary IPs across both are assigned. The item became “find the project”, probably exdeb, whose token is separately owed a rotation.
And two smaller ones: the image-mirror job does not exist (item 8 needs a little engineering, not zero), and the hook-family hit count fell 82 → 76 rather than growing (item 11’s urgency argument does not hold; its correctness argument does).