PF Simulation r34 — continuation and attachment corrections

Validated revision: r34-preserved-python-layers-byok-byos-continuation-attachments
Execution record: validation_r34_byos/executions/20260922T044145786600Z

Changes delivered

• Continuation: new checkpoints use lossless string/image deduplication and compression. The reader restores exact provider-native history, reasoning/signatures, tool pairing and completed-tool cache. Intact legacy checkpoints remain readable. Capacity errors now expose safe numeric diagnostics instead of replacing the reason with a generic unavailable message.
• Attachments: each active attachment has a Remove control. Removal retires its private continuation state, preserves historical messages/citation metadata and invalidates affected continuation. Identical filename/content uploads reuse the same active attachment without duplicate quota reservation.
• Failure recovery: identical re-uploads repair missing or corrupt retained bytes. Removal commits its audit record before physical deletion. Failed cleanup remains visible and retryable, preserves the appropriate storage charge, and cannot feed removed bytes to a new investigation. BYOS retries complete storage work without replaying model requests or scientific execution.
• Compatibility: standalone and SaaS retain Python source layers. SaaS retains per-user BYOK, regular managed storage and optional BYOS. No scientific solver correction or capability restriction was introduced.

Release validation

The repackager checks file integrity only; its generic notice about changed material is separate from the functional acceptance results recorded above.

The 25 numerical lifecycles comprise ten original standalone/SaaS workflows, five hosted personal-key workflows using managed storage, and ten hosted personal-storage workflows across S3-compatible and NAS transports. They cover OpenAI, OpenAI-compatible endpoints, DeepSeek, Gemini and Anthropic, and include private layer creation, numerical validation and execution with the layer active.

Regression runs use resumable batches bound to the tested source and exact collected case identities. Completed batch logs, JUnit and receipts remain available, and aggregate results are checked against them. Earlier incomplete runs are retained separately in PF_Simulation_r34_Incomplete_Validation_History.zip and are not counted as release acceptance. One was interrupted by an environment restart; another exposed a validation false positive that confused ordinary source filenames with private checkpoint fields. The privacy check now distinguishes fields and actual native payload markers from filenames.

New regressions cover image-heavy logical checkpoints exceeding the old 2 MiB storage ceiling, save/reload between continuations across all five protocols, exact native-history retention, cached proposals executed once, owner isolation, legacy loading, corrupt checkpoint rejection before network/tool work, context-budget blocking, stored/expanded capacity bounds and decompression defenses. Attachment coverage includes standalone/hosted authorization, CSRF, concurrent-operation exclusion, database rollback, unlink/acknowledgement failure, duplicate integrity repair, pending deletion, restart/retry, shared remote references and quota handling.

Limits of the result

The provider responses are scripted HTTP/protocol fixtures. Numerical execution is real local subprocess execution, and storage tests use S3-compatible and NAS loopback HTTP services. These results do not establish live paid-model accuracy, live customer cloud/NAS behavior, production PostgreSQL, Docker isolation or deployed load capacity.

Checkpoint compression preserves state; it does not make continuation unlimited. The complete stored record remains bounded to 2 MiB, the expanded JSON to 64 MiB, and the existing history/tool/segment limits still apply. Provider context and token limits are separate. A checkpoint already discarded by an older release cannot be reconstructed from chat text; start a new question when no valid saved state remains.

Removal cannot retract data already sent to a model or remove quotations retained in historical messages. It removes the active attachment payload and future file access. Completed runs, proposals and private layers remain retained.

Upgrade and reproduce

Use the corrected complete r34 ZIP. Preserve the existing assistant database, evidence, personal-storage configuration and credential-vault key; restart the application and reload the browser. See R34_CONTINUATION_ATTACHMENTS_GUIDE.md inside the archive for removal/retry and continuation details. This delivery is a code package, not a deployment to an existing server.

From the extracted project directory:

python release_tools/audit_r34_core.py verify
python release_tools/validate_r34_byos_release.py verify
python release_tools/validate_r34_byos_release.py execute --baseline /path/to/original/r34.zip

The final command requires a POSIX environment, the project research/development dependencies and Node.js. This acceptance-runner requirement does not change the application launchers. The archive retains its execution record, individual command logs, JUnit XML and numerical evidence under validation_r34_byos/.

Archive: PF_Simulation_Standalone_and_SaaS_r34.zip
Archive bytes: 517,386,147
Archive SHA-256: eee029b2f3c342313cccb561c8c69495e8d29bdfde34fb543f2ea4fb17f8e355
Original r34 comparison SHA-256: 34f9e7d54c7909e56dfe16c43e14ac9b07eb23178014e5961a51e22c8fe57c73