Files
rust-discovery/docs/journal-capacity-acceptance.md
T
2026-11-13 22:23:09 +01:00

5.1 KiB

Finite journal capacity during the final Cloud reboot acceptance

On 2026-11-13 the original approved Cloud-VM reboot had already happened, but normal reconciliation stopped before producing a new Agent SyncRequest. The Root database occupied 268,414,976 bytes: 65,531 pages of 4,096 bytes, against the helper's 65,536-page limit. Helper checkpoint 1378 and both journals stayed fault-free; generic transport/dependency errors did not identify the cause. The VM still had approximately 21.6 GB of filesystem space available, so the configured SQLite limit was distinct from disk exhaustion.

The helper persists a capture before the Agent constructs its SyncRequest. That transaction inserts the complete bound snapshot and advances the helper checkpoint together. The configured limit prevented the insert, explaining the unchanged helper checkpoint and absence of a new request. Historical captures, signed responses and consumption tombstones intentionally remain durable; this change deletes or rewrites none of them.

Actual copied-database proof

The bootstrap adapter made a consistent private SQLite backup of the actual Root database. On that copy, the old 65,536-page limit rejected a 524,288-byte test insertion with SQLITE_FULL, code 13. After rollback, the exact same insertion succeeded at the new finite limit of 262,144 pages / 1 GiB. It was rolled back again. The original database's bytes and inode remained unchanged, and both copies of the schema/row fingerprint matched. No package effect, RMM job or reboot was involved in this proof.

Preserving runtime maintenance and normal completion

The shared bounded_journal_capacity now enforces 1 GiB on both Agent and helper, still refusing unexpected page sizes. The Agent uses this same routine when opening its journal, so both binaries were rebuilt. Qualified Bookworm builds, 126 affected Rust tests, formatting and warning-denying Clippy passed. The existing capacity tests verify the new limit while retaining the original consumed grant, start evidence, sequence and replay protection.

The native maintenance versioned only the two binaries. It preserved all 18 Root consumptions, the original unresolved reboot, old signed histories, device identity, database inodes, General policy/pin, provider and all APT attempt/acceptance files. The helper was started first while the Agent remained stopped for independent receipt verification. The qualified binaries are:

  • Agent: 9e9f323075f44e02e3900267b846f124303179171476a23d023708fc8339f9b4
  • Helper: d6fbfc25263243475c50e703c014772c359533548bccc3f09ff91072b0e8b929
  • Unchanged provider: 1dc0237dd8df9c9d22568bb1f6b047ede73bc2a5fca5c7fcedb7f24ab2a58b1e

After the normal Agent resumed, the same original reboot grant 25a4bf3f-772c-487d-825d-73bcbbc418fd completed through signed RootSync. The actual Root ledger then contained 18 consumed, 0 unresolved, 0 faults; the old 17 Root rows and database inodes remained unchanged. No second reboot, new grant, direct outcome update or journal reset was used. The verified final reboot result and restored non-reboot General policy are separate native receipts in the original session.

Protected witnesses, excluded from Git, are under devstate/cloud-reboot/d015c3ef-f3aa-4522-a3c2-fcb6099df17b/:

  • capacity-copy.json: actual SQLite error 13 and unchanged original database.
  • capacity-runtime.json: completed preserving binary maintenance.
  • capacity-upgrade.json: bound index of actual runtime and build receipts.
  • result.json: verified completion of the original single reboot.
  • closed.json: verified restoration of the original non-reboot policy.

The APT endpoint remained disabled throughout. Because runtime measurement includes the helper binary, its earlier closed APT pin becomes historical after this replacement. A separate post-reboot tool remeasures the unchanged provider and performs only the existing copied-database readiness probe after all 18 Root attempts are closed. It does not create a new package acceptance or change any historical acceptance record.

The initial post-reboot readiness run failed before launching the Python probe. The bootstrap had incorrectly added three measurement fields to installation metadata, whose Rust parser permits exactly four fields. Independent readback confirmed the seven current keys, the exact original four-field backup and unchanged values for those original fields. The failed report and all original backups remain retained. A narrowly bound source correction restores a copy of the original four-field bytes while separately versioning the invalid file; it leaves the already correct runtime config/proof unchanged. This recovery requires the original reboot to be closed and does not repeat any package job. Its native completion must be recorded separately from the earlier failure.

Follow-up after the MVP

The 1 GiB limit is finite reserve, not permanent retention management. Durable capture/decision history continues to grow and another exhausted limit must fail closed. Investigate measured growth, storage visibility and a separately specified safe archival policy after the MVP; do not add deletion, compaction or automatic limit increases to this acceptance fix.