Files
rust-discovery/deploy/shared-core-upgrade.md
T
2026-11-14 03:06:06 +01:00

109 lines
6.0 KiB
Markdown

# Preserving Linux shared-core runtime qualification
The notebook only builds and orchestrates. The fixed Cloud Linux VM remains the
client; `discovery.ease-technologies.de` remains the server. These bootstrap
tools never create or approve an RMM job, enroll, reset a journal, install a
package or reboot the guest.
The original `cloud_shared_core_upgrade.py` remains frozen at SHA-256
`09b60e99e1bf92dbaa60f0b0a64f9b94e3cefebb110d558b14c381293cbde873`.
Its case `2f8de8b9-2abe-4125-aa12-ae38332a32c6` stopped before any binary exchange:
the generic `nlink == 1` check rejected an existing valid APT publication pair.
The original failed report, staged binaries and APT records remain retained.
`cloud_shared_core_failed_restore.py` restarts only the verified original
runtime, once; its saved actual readback is required before a new V2 case.
V2 is a source copy with a narrow alias admission and earlier object checks:
`cloud_shared_core_upgrade_v2.py`, SHA-256
`d4f857b8ff6e7d9802f863a120a1576ffac5b36facb6a259149409ef25bfbc91`.
The only permitted alias paths are under `/var/lib/rmm-apt-transaction/`:
- `acceptance-rmm-apt-fixture-update-1.1.json`
- `pending-acceptance-rmm-apt-fixture-update-1.1.json`
Both must retain device `2049`, inode `1028362`, UID `0`, GID `984`, regular-file
mode `0600`, link count `2`, and SHA-256
`a88d6b10bbf6c36eaa8dee87eb1a381f2becc36297db8df2b4189791b9165506`.
That content hash comes from the original verified, UI-approved native APT
update result. The mode is the actual stored `lstat` value `33152` (`0100600`).
Neither name is rewritten, unlinked or chmodded. Other hardlinks, symlinks and
unexpected nonregular state objects remain rejected.
The qualified `durable-outcome-v1` release artifacts remain:
| Artifact | SHA-256 |
|---|---|
| Linux agent | `c9c09de0ecb52b556d707cf45accb69186c73b8ec9baeeabfaeff220fafb1d4e` |
| Linux helper | `51c3142ab332216d6a94ff64ff15cbc37ee1d5a99896c14c87bb6f62e63bb946` |
| Unchanged APT provider | `1dc0237dd8df9c9d22568bb1f6b047ede73bc2a5fca5c7fcedb7f24ab2a58b1e` |
Planned, not executed by this preparation:
```text
python3 deploy/cloud_shared_core_upgrade_v2.py prepare --restore-readback <exact-private-original-case-readback.json>
python3 deploy/cloud_shared_core_upgrade_v2.py install <new-case-id>
python3 deploy/cloud_shared_core_upgrade_v2.py finish <same-new-case-id>
```
Each new intent binds both runner hashes, the restore readback, actual alias
metadata, original APT checkpoint, release qualifications and the Linux-MVP
closeout. Local cases use `devstate/shared-core/runtime-upgrade-v2/`; remote
cases use `/var/lib/rmm-bootstrap-history/shared-core-v2-<id>`. A durable
installation receipt forbids repeating installation, including after a lost
reply; only readback/finish of that same case is permitted.
The pre-restart preservation checks run under both existing locks. Old
consumptions, outcomes, signatures, database identities, trust and backup
witnesses remain bound. Private key contents are never read or exported.
General policy/pin remain unchanged, `allow_reboot=false`; the APT endpoint
stays closed and its historical runtime pins remain stale. Native mTLS/Sync,
MCP and central-count qualification are separate final acceptance steps.
V2 preparation checks passed locally: Python syntax, fixed RPC globals, source
format, all 65 qualified release source bindings and the existing native Linux76
core-unit receipt. This document does not claim native V2 installation success.
## Retained native bootstrap assertion failures
The original V1 failure above was a bootstrap verifier error: the existing
legitimate APT hardlink publication was rejected. Neither alias was repaired or
rewritten. Its original-runtime restart and actual readback were separately
recorded before the new V2 case.
V2 case `a2423612-721f-4a6c-a832-b9eef549c41e` subsequently reached the persisted
`installed_preserved` stage: both new binaries were installed, all pre-restart
data witnesses matched, and both daemons started. Its last verifier assertion
incorrectly compared process GID with the account's primary passwd GID. Actual
readback showed the expected service identity, UID993/GID984, matching the
independently verified Linux-MVP closeout. This was a second bootstrap verifier
error; the report remains `completed=false` and is never rewritten as success.
`cloud_shared_core_upgrade_attestation.py` is a separate read-only supplement
for that exact case. It rechecks preserved rows/schema/inodes/trust/backups and
the actual process identity, and binds the stored native VM/server Sync114
evidence to the retained exact signed decision and request. It does not start
services, retry installation, change SQL, alter the original report, or create
jobs. Its prepared SHA-256 is
`59b20ecd2d32ea3b37f2840868eb67ce677ffdb967e079a667e53c7907fce86b`.
## Actual read-only closeout, 2026-11-14
The supplemental attestation was actually executed and passed; its protected
receipt is `readonly-attestation-63730c38-327e-4635-802f-8289c28081de.json`
inside the V2 case. A prior strict row-hash readback rejected28 cached outcome
references that normal signed Sync had renewed. Only those reference changes
were then independently proven: reconstruct the original complete row hash with
identical job/attempt/outcome bytes and the preserved original decision ID;
verify both original Ed25519 decisions and the exact applied request/response,
grant/plan/outcome binding and monotonic journal history. All other historical
rows remain byte-identical. No general outcome exemption or state rewrite.
Native MCP subsequently read the existing Linux device online, rejected the
foreign tenant, and confirmed unchanged30/30/26/0 central counts. Agent/helper
and server share the ready signed Sync114; original18 Root consumptions remain
closed. `devstate/cloud-shared-core-current-runtime.json` binds both actual
attestation and MCP receipts. It supplements the preserved historical MVP
closeout and does not requalify APT for the new binary hashes. All original
failure reports remain unchanged, and no installation or service start was
repeated to obtain this read-only closeout.