Per the 2026-07-21 refresh brief: R-39 interim blocks (B4/E1) and the R-36
manual-Save block (C4) deleted — both shipped and proven live; freemail.hu
gate proven (R-4 COMPLETE); golden/floor-lift note now cites two shapes
(rehearsal + virgin HP t740 day-0 lift 0.153.0->0.156.0); A3 loader table
per operations/nodes.md (N100=mkimage/SB-off per record, HP t740=shim/SB
ENABLED); B2 multi-NIC cabled-port gotcha (R-59/R-60 pending); new A5 gate
(agent >=0.93.0 deployed box-side before the first escrow ceremony); D
offboarding pointer to §G (R-25b). DRAFT status and the C7 graduation gate
unchanged. ROADMAP R-25b pointer follows the rename.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UuFPHmHNrCJj1VhY6QdDMU
Found validating v0.69.0 against the live hub. demo-vm-felhom was deleted
on 07-18 and was still on the Customers list AND still raising offsite_stale
(10 events, latest 07-21 17:34, operator email at 19:34) — because
GetCustomers() is report-derived and no lifecycle tier ever deleted a report.
New leg 3 (residue), before the record purge: reports, app_telemetry,
app_log_tails, log_tail_requests, customer_notifications, plus the
credential-bearing appliance_registrations and selfbind_tokens. Audit
(events, notification_log) and F-14 provenance still survive.
Ghost customers are now deletable: 404 means "nothing here", not "no config
row". With no config row the offsite descriptor is unknowable, so the Hetzner
and descriptor legs record skipped_no_config rather than a bare "skipped".
Two more red-proofs.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J55BQE1gE2V4ffud5jweGS
POST /configs/{id}/delete now runs hosts -> RESET -> purge behind three
acknowledgements, a typed customer-id, a stale-preview check and the
ONLINE-host refusal (every gate before any write, so a refusal has zero
side effects). The shallow handleConfigDelete is gone.
Two invariants are asserted, not just commented: ruling 3 is preserved by
construction (leg 2 never sees a host row) and retained escrow custody is
purged exactly once, in leg 3 (leg 2 runs with purgeEscrow=false).
handleCustomerReset's committed half was extracted as commitCustomerReset;
the standalone RESET path is byte-identical to v0.68.1 and its suite is
untouched. Five red-proofs run.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J55BQE1gE2V4ffud5jweGS
New documentation/operations/nodes.md: HP t740 hardware, disks (SanDisk system
SSD 182195804614; Toshiba 1TB NVMe 58BS11AFT8MQ PRESENT AND UNENROLLED, still
NTFS, do not touch), the five-NIC map and the trap that cost the first install,
and the access path - no SSH key is baked, auth is the hub-vaulted G1
break-glass password.
tailscale.md gains demo-hp, the operator-lab-exception warning so a future
product-shape audit does not conclude the product ships tailscale, and the
--accept-dns evidence: the join omitted the flag, MagicDNS rewrote
/etc/resolv.conf, and it was reverted. Harmless at the vacation site, would
have bitten silently at home where split-horizon matters.
OPEN: key expiry still enabled on demo-hp (2027-01-17) - needs an admin-console
toggle or an API token; a pre-auth key cannot do it.
R-59 no-DHCP install must hard-abort (it baked 192.168.100.2 static and
completed - a box that can never call home). R-60 first-boot NIC sweep
self-heal. R-61 the baked root password must be knowable; a fixed well-known
password is explicitly rejected.
Positive evidence same-session: R-21 slice C PROVEN on a SECOND, virgin board
(HP t740) - and the shim loader booted with Secure Boot ENABLED, retiring the
assumption that Felhom installs need SB off. Fresh-box floor lift
0.153.0 -> 0.156.0 during day-0 cited on the publish-train row.
ISO README gains the t740 five-NIC trap: the 4-port igb card gets no lease,
the onboard r8169 port does.
R-58 records the operator ruling (2026-07-21) with the argument verbatim: the
installer should list available storage devices, excluding the install media,
and let one be selected. Third ISO mode alongside unattended-serial and
match-nothing-safety; unattended stays the appliance/factory mode. Slice 1 is
the abort-screen candidate table, same enumeration code, and it collapses the
two-boot dance on its own. Matters most for BYO/reinstall, where the serial is
unknown and a wrong guess is destructive.
ISO README gains the HP section: shim proven on this board by the safety boot,
the uncommitted-armed-profile pattern, verify-from-inside-the-ISO, and a
pointer to the prior-LVM abort that is the one likely failure on a
second-hand disk.
R-55's reboot leg ran operator-present on 9201: immich UI-stopped -> stayed
stopped across pct reboot, calibre-web recreated, zero alerts, ~15s.
R-57 records the lifecycle mechanism with the operator's abandoned-app
requirements verbatim and plant-it as the motivating case, including why the
retired/ directory move was wrong and the v0.158.1 pointer-receiver defect.
PROMPT-TEMPLATE: standard 'For the operator' plain-language section, mandatory
for M+ tasks and anything with a STOP.
ROADMAP rulings (operator, 2026-07-21): R-25b full-teardown cascade with three
acks + typed name (M-sized, spec to follow, no longer blocks R-3); R-11 channel
= direct Messenger, doc is the architect's; R-42 option (a), sidecars follow the
app; R-17 delete the archive - spike-lite found NO tooling verb targets it, so
it is an operator console action; R-4 complete (freemail.hu verified).
R-55 + R-41 slice 1 marked shipped; new R-56 (app difficulty classification -
the constructive half of the glance ruling).
scripts/build-hub.sh v1.23.0: the hub build script was outside any repo. Adopted
verbatim + versioned; the build-dir path is now a symlink to it.
felhom-testing skill: the ~1/5 recovery-code 'known flake' is retired - it was a
real defect the test was correctly detecting.
Mirrors profiles/n100.profile with one deliberate difference: the loader
default stays at the stock signed shim chain. mkimage exists to work around
the N100's AMI firmware GRUB relocation defect, not as a Felhom requirement,
and it costs Secure Boot. HP business firmware should take shim - so try it
first and carry mkimage as the fallback.
Ships the match-nothing disk filter, so built as-is it IS the hp-safety
pre-flight ISO. Arming instructions point at an uncommitted working profile,
per the n100-demo precedent.
Dead primary: degraded in 13 s, exactly one app_start_failed, banner rendered and
self-cleared. Boot orphan: recovered in one attempt with zero alerts. Dead dhclient:
detected in 57 s on process liveness while the lease was still live, healed 120 s after the
kill — the tunnel never dropped, so the outage was prevented rather than observed.
P1 answered as a by-product: bookstack StartedAt == the moment bootrecon StartStack
returned, so unless-stopped did NOT resurrect it. F5 hypothesis confirmed.
New R-55, surfaced by the leg designed to prove the opposite: the boot bind gate recreates
and STARTS every deployed drive-backed app unconditionally, so a customer Stop does not
survive a reboot for those apps. Predates R-52 and does not implicate it, but it narrows
R-52's practical scope and needs a ruling.
R-51's roadmap diagnosis is corrected at the source: aggregation returned StateRunning
("partial") for a running/stopped mix, so the stack read RUNNING and IsDownState was never
consulted about at all — the constraint that row protects was never in tension
with the fix.
New R-54 row closes the INCIDENT-guest-dhclient-killed-2026-07-20 §5 OPEN RISK, and records
the design fact that makes it work: liveness of the DHCP client is itself a probe, because
the damage is timed and the address outlives its cause by 1-2 hours. The static-guest leg is
deliberately deferred to R-50.
New capability-map row is IMPLEMENTED, not PROVEN-LIVE: one leg is live (the watchdog's
healthy cycle on felhom-pve), the three that matter are destructive and operator-present and
have not run.
PROMPT-TEMPLATE §10 gains the seam-discipline row, including that a strings.Contains source
assertion is NOT sufficient — a commented-out call still contains the string.
The operator pressed Re-issue PBS credentials and the chain closed in 13 seconds. The
identical click on 2026-07-18 did nothing at all.
hub 08:39:31Z fresh mint, generation 0 -> 1; descriptor gains secret_generation: 1
(token_id + fingerprint BYTE-IDENTICAL — the invisible re-key shape)
agent 10:39:34 felhom-pbs-apply read felhom-pbs (leg b: the impossible read)
agent 10:39:38 ERROR REJECTED ... applied and DEAD, previous_state=applied
(leg c: the R-39 state, loud)
hub 08:39:45Z consumed_at stamped
agent 10:39:45 one-time token secret consumed (leg a: NO short-circuit)
agent 10:39:45 reconcile (set-only, no --server)
agent 10:39:47 pbsdr: converged state=applied
Corroboration: marker hash moved to afbb3b41… (it was byte-identical to the pre-reissue
marker in the failure); secret mtime 2026-07-18 -> 2026-07-21 10:39:45; new credential
probes 200; three consecutive reports trace applied -> auth_failed -> applied; ZERO
self-heal escalations, one mint, one consume, no consumed-failed.json — the box healed
through the descriptor path before the damper was ever needed.
Recorded for future runbooks: the operator first pressed the OFFSITE re-issue (two
distinct Re-issue actions exist). Harmless to PBS-DR, but it rotated the restic password
and correctly marked the escrow STALE, so the ceremony had to be re-run. Name the surface
explicitly next time.
The v0.68.0 row wrapped itself in a <div>, but the artifacts <form> IS the CSS grid
(display:grid, no inner container). The stray </div> closed the surrounding card from
inside the form and the new <div> was never closed — it swallowed the submit button and
ran to </form>, so the row rendered outside the card and Save landed inline. Reported by
the operator on first use.
The field still submitted (it stayed inside the form), so this was layout damage rather
than data loss, but the unbalanced markup put every section below it in the wrong
container.
Fixed as plain grid cells (grid-column: 2/4), no nested elements.
There was no render assertion on this form at all, which is why a hand-edit broke it
silently. The new test asserts the field is inside the form, the button has not escaped,
the form contains ZERO divs, whole-page div balance holds, and the sections after it
survive. Red-proofed against the broken shape.
R-39's three legs are closed and deployed: the hub stamps a monotonic secret_generation
so a re-key finally moves the descriptor hash; the wrapper gains a narrow read verb so
the non-root agent can read the credential it writes; and ProbeAuth turns a 401 into a
loud auth_failed the existing damper escalates to a fresh mint. Plus a consumed_at
honesty gauge for the applied-but-never-consumed disagreement.
Recorded in the R-39 row, because both are the kind of thing a future reader needs:
- A load-bearing fact the spec did not flag, checked rather than trusted: Apply bails out
if the storage status probe ERRORS and adopt converges without consuming when the
storage reads active, so the fix depended on PVE's 401 behaviour. PVE's storage_info
wraps activation in eval{} and leaves active=0, so a 401 returns HTTP 200 with
active:0 — never an API error. The chain is sound by proof, not inference.
- A defect I shipped and caught: v0.91.0 built the probe seam and main.go never wired it,
so the leg was inert while every test passed. Same class as controller v0.154.0 the day
before. Fixed in v0.91.1 (artifact superseded, not overwritten); v0.91.2 made a healthy
probe observable so "no auth_failed" can never again be confused with "never probed".
The DR-tier capability row is deliberately NOT upgraded to PROVEN-LIVE: the decisive
evidence is STOP-2, the operator pressing Re-issue and the box converging where the
identical click did nothing on 2026-07-18.
R-50b(a) shipped — wrapper sha256 in the manifest + agent reporting + host drift surface,
with unknown-on-either-side reading as quiet rather than drift. (b)/(c) remain open: the
wrapper is still fetched unversioned from raw/branch/main.
Completes the hub half of R-39's fleet fix on top of the generation core (c484aa2).
pbsdrheal gains an auth_failed TRIGGER — a new trigger in the existing machine, not a
new machine. A box whose credential PBS rejects escalates to a fresh mint, never a
re-stage (which would re-feed the secret PBS just rejected), through the EXISTING damper:
a 401 flap must not become a secret-minting chain. With the generation stamp this closes
the loop end to end — agent proves the 401, hub re-keys, generation advances, descriptor
hash moves, agent re-consumes.
consumed_at honesty gauge: a staged secret still unconsumed past a 15-minute grace while
the box reports `applied` is surfaced with its own event. That is the exact 2026-07-18
fingerprint and a disagreement no single tier can see alone. Deliberately a SURFACE, not
a heal — auto-re-issuing on it would mint a second secret on top of an unconsumed one,
which is the mint/consume race R-39(a) already recorded. One event per distinct report,
and an honestly-stuck box does not double-report (its unconsumed secret is the symptom
being healed, not a contradiction).
R-50b(a): ArtifactManifest.WrapperSHA256 + operator field + host-page drift surface. The
PBS wrapper is root-owned 0755 and the pinned sudoers vector, yet installed unversioned
from raw/branch/main and absent from every manifest. Agents >=0.91.0 report the installed
hash; a mismatch is surfaced. An unknown on EITHER side reads as quiet, never as drift —
lighting every host amber on rollout day is how a warning becomes background noise. The
delivery channel itself stays R-50b(b)/(c).
Compatibility unchanged: safe for 0.90.0 agents (unknown JSON key dropped); the re-arm
and auth-honesty guarantees need agent >=0.91.0, so MinAgent moves only after the fleet
has self-updated.
Tests: auth_failed escalate/debounce/recovery-forgets-streak; honesty gauge incl. grace
window, the restage edge (consumed_at deliberately NULLed), consumed-never-alarms, and
honest-stuck-no-double-report; wrapper drift incl. both unknown directions. Red-proof run
at the assertion level: removing the auth_failed arm fails the escalation tests with
reissues=0.
The fleet half of R-39. An ep0 credential re-issue re-keys the SECRET of an existing
token, so token_id, fingerprint, datastore and namespace all come back byte-identical.
The agent re-applies on the descriptor's CONTENT HASH, so a re-issue was invisible to a
converged box: it short-circuited, never consumed the fresh secret, and served a revoked
credential while reporting `applied` — the N100 failure of 2026-07-18.
host_pbs_secrets gains a monotonic per-host `generation`, advanced by every fresh MINT and
by nothing else, stamped into the descriptor as `secret_generation`. That is now the only
field a re-key moves, and it is what re-arms the agent.
DEVIATION FROM SPEC, deliberate: the brief said to return "the new row's id (int64) …
no schema change". There is no row id — host_pbs_secrets is keyed by host_id and UPSERTed
last-write-wins, so a new row never exists, and created_at collides for two mints in the
same second. An additive counter column is the only monotonic source; it uses the repo's
existing idempotent ALTER-TABLE idiom.
RestageHostPBSSecret deliberately does NOT advance it: a re-stage re-arms the SAME secret,
the descriptor content genuinely has not changed, and a bump would cause a pointless agent
refetch loop (that method's own contract says so).
Also corrects a comment that asserted the re-issue refreshes the descriptor "with the NEW
token_id/fingerprint". That is false for a re-key, and believing it is why the descriptor
was never expected to be identical in the first place.
omitempty is load-bearing: a zero generation must not start emitting a new key into every
pre-existing descriptor, which would itself be a fleet-wide spurious re-apply.
Compatibility: agents below 0.91.0 drop the unknown JSON key and behave exactly as today —
inert, not breaking (Scenario C).
Tests: store-level monotonicity + per-host isolation + restage-leaves-it-alone; descriptor
byte-change, omitempty, and sibling-key round-trip; and a FLOW-level test driving
ReissuePBSDR against a fake that models a real re-key. Red-proof run at the assertion
level (not the compiler): commenting out the stamp makes the flow test fail with both
byte-identical blocks printed.
The operator moved the global floor to a version the box did NOT run (0.153.0 ->
v0.154.0) and the managed self-update fired exactly once:
06:57:13Z UpdateState pending, initiated_by=auto-floor
06:57:17Z agent: controller swap requested 0.153.0 -> 0.154.0
06:57:21Z container restarted
06:57:29Z agent: new controller healthy (16 s save -> healthy)
Over a 39-minute window: swap requests 1, agent-driven bootstrap restarts 1,
rollbacks 0, container RestartCount 0. VerifyStartup confirmed on the next boot;
the following periodic check logged "Current version 0.154.0 is up to date" —
the at/above-floor branch correctly doing nothing.
The 2026-07-20 attempt proved nothing because it targeted an already-running
version; that was the whole reason this leg stayed open.
Disclosed in both rows: a hand-deploy of v0.155.0 at 07:17:10 falls inside the
observation window and is what StartedAt shows afterwards. It never goes through
SwapController, so the swap-count assertions hold across the full window — and it
incidentally re-confirmed the at/above-floor branch (0.155.0 running against a
0.154.0 floor -> updater did nothing).
R-23(b) (cosmetic Waiter "recovered" log timing) remains open.
R-48 — the offsite restore controls collapse to one „Visszaállítás…" entry per app plus
a per-app wizard with three described intent cards. Shipped in controller v0.154.0
(3a9d744). Live click-through still pending the operator's floor save.
R-39 — the planned v0.90.1 artifact publish was CANCELLED as a false signal (operator
ruling 2026-07-21). 9596d5a changes zero non-test Go files; its own message says "the Go
binary is unchanged". The fix is the felhom-pbs-apply wrapper, which felhom-pve has
carried since 2026-07-18 and which every new install fetches from raw/branch/main
regardless of binary version. Publishing would have delivered no behaviour change and
advertised a versioned fix the artifact channel never carried.
R-50b (new) — that stop surfaced the real defect: a root-owned privileged host artifact
is delivered unversioned from main, absent from the Day-0 manifest, so the fleet has no
way to answer which wrapper a given host is running.
Row 61's status cell still read "PARTIAL — scope corrected 2026-07-19" while the note
body already ended with the destructive drill of 2026-07-20 and the words "the row now
earns PROVEN-LIVE": 40 files placed after the operator deleted the photos in immich's own
UI and emptied the trash, 1 DB dump replayed rc-0, 11 assets active, no schema drift,
timeline confirmed — taken through the customer-facing buttons, not endpoint shortcuts.
Cell-only correction; the full note history is retained verbatim.
Capability map — the destroy-then-recover drill ran through the customer UI:
photos deleted, TRASH EMPTIED, full files+database restore. 40 files placed
against 6 in the earlier non-destructive run, 1 DB dump replayed rc-0, 11
assets active, no drift, timeline confirmed. That is the proof the 6D
downgrade asked for, so the offsite-restore row earns PROVEN-LIVE. The
customer-restore row records the honest residual: an operator ran it, so the
row's literal 'a customer, not the operator' wording still owes one pass.
ROADMAP R-23(a) — the STOP-2 floor save released the held wait in the SAME
SECOND (hub 18:56:27 CEST = controller 16:56:27Z), out-of-cycle report 2s
later, generation advanced 0 -> 1. Still open: the self-restart single-fire
leg, since the floor was set to a version the box already ran.
Trap recorded: the wake is logx.Debugf, so it is invisible in docker logs at
INFO and lives only in the debug ring.
The operator confirmed the immich timeline renders correctly after the
2026-07-20 reconstitute, which was the last named residual. But that run
restored over an already-good state (11 assets before, 11 after), so it
proves the H4 window is closed and the operation is clean — it is NOT a
destroy-then-recover drill.
Flipping the row on it would repeat the 6D overclaim verbatim: 'immich
end-to-end from offsite alone' must not be claimed by a run in which
nothing was ever lost. Row stays PARTIAL pending a real destructive drill.
ROADMAP R-47 and the capability map's offsite-restore row now carry the live
evidence: a supervised reconstitute of immich from snapshot 49e7cb46 — the
same snapshot that aborted in round 2 — ran stop -> DB-service-only start ->
replay rc-0 -> full start, with no 'already exists', a SUCCESS outcome, and
immich's own 'No schema drift detected' twice where round 2 left it
reporting drift.
The offsite-restore row's residual for a PROVEN-LIVE flip is now only the
immich timeline screenshot (no browser on DooPlex).
Golden 0.153.0 baked and published the same day, sha256 15fdd191f3c660a6...,
first golden carrying all four infra images.
ROADMAP: R-47 collapsed to a shipped one-liner naming both paths (offsite
ReconstituteFromOffsite and local RestoreFromRecoveryUnit), the DB-only
window mechanism, the fail-closed rule, and the no-agent-coupling
declaration (MinAgent stays 0.90.0).
Capability map: the offsite-restore PARTIAL row and the customer-restore
MISSING row both note that H4 is closed in controller v0.153.0 and that they
now await only one clean acceptance run.
Also corrects the stale "closes in v0.149" wording in BOTH rows — v0.149.0
was the F3 dashboard BackupStatus fix; R-47 shipped in v0.153.0.
Capture disproved addendum 1's guess: macOS DOES send a correct NBNS query and
IS answered correctly in 140us; it simply never acts on it. Records the proven
per-client matrix and keeps sidebar discovery explicitly unproven.
The v0.150.0 'orphaned dhclient' cleanup killed guest 9201's DHCP client, not a
host one: LXC guests share the host PID namespace, so the missing eth0, pidfile
and lease file were all true ON THE HOST and all irrelevant. Lease expired ~80
min later, guest lost its address, tunnel/hub/catalog/agent all went dark.
Restored; records the one-command ownership check (/proc/<pid>/cgroup) and
leaves the unsupervised-dhclient risk open for a ruling with R-50.
Also records against R-45 the lesson the loop taught: a phase a client answers
with a one-shot action must be an EDGE the registry serves once, never a level
synthesised into the phase channel.
felhom-pve joins the tailnet as a host package (100.70.170.35, expiry
disabled); DooPlex already advertised 192.168.0.0/24 via its GitOps k3s
tailscale pod (100.107.87.53). ssh felhom-pve now targets the tailnet IP
(direct-over-LAN at home, tunnel when away); felhom-pve-lan = LAN fallback.
--accept-dns=false on the host. Host converted static->DHCP (reservation
keeps .162 at home). Measured: --accept-routes hairpins the local subnet
at home -> travel-only opt-in. PBS is offsite (own WireGuard tunnel), needs
none of this. Recorded not fixed: controller<->agent plane hard-pinned to
192.168.0.162 (agent listen_addr + guest bootstrap.json) -> not yet portable
off-LAN. No code changed. Full doc: documentation/operations/tailscale.md
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XMForrE4c1wZxd9LukxYVt
VERDICT. H1 confirmed: the reconstitution never ran. /backup/offbox/reconstitute
was never hit (reconstituted/safety-dump/replay line counts all 0); the operator
clicked the old missing-only button, which merged 34 files and left the DB
untouched. H2 rejected (the full path never executed; staging used the correct
snapshot 49e7cb46). H3 rejected (no replay happened; DB was genuinely asset:0).
R-44 is exonerated and working: 49e7cb46 carries offsite_run_id 20260719T102851Z
and its dump probes to asset:11 / user:1 — a properly coherent pair, against
round 1's asset:0 / user:0.
H4 (NEW, the more important result). Phase-3 recovery ran the real sequence and
the v0.148.0 path executed correctly — safety dump, stop, start, replay — then
the replay ABORTED: relation "clip_index" already exists. Proven to the second:
10:58:25 controller begins replay
10:58:33 immich-server "Reindexing clip_index" -> "Reindexed clip_index"
10:58:35 dump's CREATE INDEX clip_index fails, exit 3
The replay races the app's own schema repair, because reconstitution starts the
stack before replaying (ImportDump needs a live container). The photos survived
only because pg_dump emits COPY data before CREATE INDEX — luck, not design. The
same start-then-replay shape is on the LOCAL path, so it is a class defect.
Recovery: PARTIAL and STOPPED as instructed. All 11 assets are back and visible
by every DB criterion (status=active, deletedAt null, all files resolve), but the
operation reported failure and immich reports schema drift. Not clean acceptance
evidence; no second attempt made; safety dump read only.
Findings -> R-47 (the replay race), R-48 (restore controls separable only by
layout — the proven cause of this incident), R-49 (1.1 GB of a 1.2 GB immich
backup is ML model cache + a postgres data-dir tar duplicating the .sql dump +
immich's own nightly dump; recorded, not changed).
RULINGS LANDED.
- 00-capability-map.md:61 -> PARTIAL, scope-corrected: 6D's destruction hit the
FILE TREE only; the DB survived in its named volume, so "end-to-end from
offsite alone" overclaimed scope.
- The 704.6 MiB "discrepancy" -> CLOSED, not a defect: it was immich's own
Tarhely widget, never a controller page. Same for round 2's 650MiB->1.4GiB.
- customer-restore row: partial evidence recorded, NOT flipped.
Diagnosis only — no code, label or layout changes; no forget/prune.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P9Nn14TWGzKoqAJAiVwC2s
admin