fix(sessions): preserve ciphertext on wrong-key pop

[harshitethic]

Summary

• distinguish an expired Fernet token from ciphertext that cannot authenticate with the configured key when EncryptedSession.pop_item() unwraps the backend's atomic pop result
• keep the existing behavior of discarding expired items
• restore the exact popped encrypted envelope and stop when authentication fails, preventing a wrong key from draining the rest of the recoverable session history
• add a SQLite-backed regression test proving a wrong-key pop leaves ciphertext intact and the correct key can still read all original messages

Verification

• git diff --check
• Python compile check for the modified session module and regression test
• branch reduced to a single focused commit with no workflow/helper files in the final diff

Fixes #5005

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7366a7f3a8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Do not classify corrupt tokens as wrong-key ciphertext

When the newest envelope has a damaged, truncated, or tampered payload, the second decrypt raises the same InvalidToken as a wrong key, so this classifies the corrupt record as recoverable. pop_item() then restores it and returns None, permanently hiding any valid older items behind that corrupt tail instead of isolating it; the implementation needs an independent key identifier or another mechanism that can distinguish a key mismatch from record corruption.

AGENTS.md reference: AGENTS.md:L115-L115

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Restore popped ciphertext atomically

If another writer appends after the backend pop completes but before this add_items call, the old ciphertext is appended after the new item, so a correct-key reader observes reordered history and the next pop targets the wrong tail; if restoration fails, the ciphertext is lost outright. The pop and conditional restoration therefore need a backend-level atomic/transactional operation rather than two independently awaited mutations.

AGENTS.md reference: AGENTS.md:L117-L117

Useful? React with 👍 / 👎.