Skip to content

validation-v0.2.md

Latest commit

 

History

History
50 lines (36 loc) · 2.82 KB

validation-v0.2.md

File metadata and controls

50 lines (36 loc) · 2.82 KB
Raw

v0.2 Validation

Automated Coverage

  • Exact escaped raw-field spans and nested JSON Pointer paths.
  • Ordered field/residual reconstruction and SHA-256 equality.
  • Empty, invalid-JSON, oversized-line, and canceled-context folds.
  • Source append and same-size mutation rejection before manifest commit.
  • Existing-object corruption detection.
  • Object reuse, batched durability synchronization, doctor, and GC.
  • Incremental unchanged skip, append tail, rewrite/truncate rejection, configuration mismatch, and CDC equivalence with full scan.
  • Exact record-sequence containment, non-contiguous rejection, escape-spelling distinction, and multi-megabyte records.
  • Dry-run and apply contained-session removal, recovery proof, associated-state cleanup, and rollback after database failure.
  • macOS, Linux, and Windows CI builds and tests.

Real Rollout Validation

Three archived rollouts were folded into one disposable object store and restored to separate files without changing their sources:

Source size Parts Reused references New stored bytes Peak RSS
1.16 MiB 136 1 429 KiB 26 MiB
64.54 MiB 1,875 248 32.92 MiB 90 MiB
723.44 MiB 76,842 5,017 233.42 MiB 126 MiB

For all three rollouts:

  • Fold segmentation verification passed.
  • Stored-object reconstruction verification passed.
  • doctor reported zero issues.
  • unfold reported the original byte count and SHA-256.
  • An independent SHA-256 command matched each source and restored file.

The largest fold initially took 324 seconds with one synchronous flush per object. Batched durability synchronization reduced the same workload to 82 seconds. Peak RSS increased to 197 MiB while remaining independent of total source size.

Codex Compatibility

A restored rollout and copied state database were placed in an isolated Codex home. The official Codex CLI found the original session ID, resumed it, and completed a turn. The live Codex home and source rollout were not modified.

Real fork chains were also checked with exact containment. Several ancestry-linked sessions correctly returned contained=false because Codex had regenerated or inserted records. This confirms that fork ancestry must not be used as deletion evidence.

Contained Removal

The complete CLI flow was run against an isolated copy of the current Codex SQLite schema:

  1. Fold an archived contained rollout.
  2. Run proof-only removal and confirm the thread, source, associations, and global-state references remain.
  3. Run apply removal and confirm the archived thread, dynamic tools, spawn edges, exact global-state references, and source rollout are gone while the container remains.
  4. Run doctor against the retained fold store with zero issues.
  5. Restore the removed rollout from its retained manifest and verify the original byte count and SHA-256.