[Improve] Show automatic memory saves in session and task transcripts

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Show successful automatic memory saves in both Fast session and standard task transcripts using the shared persisted MemorySaved event contract and expandable renderer.
• Include only successfully persisted, redacted distilled content in each event.
• Make task events deterministic and idempotent across in-process and drainer retries, while preserving private-task/session eligibility rules.
• Keep the existing deployment-settings test-lock timeout fix for parallel DB-backed server suites.
• Remove repository-skill caching changes from this PR; repository-skills caching remains owned by PR [Fix] Fast replies wait tens of seconds on skill discovery before every turn #3066.

Why this change was made

Automatic task-turn memory saves were persisted to Brain but were not visible in task transcripts. The shared event and renderer give session and task users the same clear feedback without introducing a second notification path. Retry-safe insertion prevents fallback/drainer retries from duplicating transcript rows.

Impact

Fast sessions and standard tasks show a subtle expandable “Saved to memory” indication through their existing transcript history delivery paths. Privacy checks remain before task distillation, and task event payloads are redacted before persistence. Repository-skills behavior is unchanged in this PR.

Validation

Focused producer, retry/deduplication, task transcript delivery, shared UI, Fast transcript, cloud-agents, web, lint, typecheck, and pre-push checks were run. The full cloud-agents suite has one unrelated existing OpenCode plugin-seed failure; the focused memory suites pass.

[roomote-community]

No code issues found. See task

• Expired repository-skill catalog entries are never evicted, so stale per-user scopes accumulate indefinitely.
• Memory transcript entries can list distilled facts that failed to persist.
• Task follow-up turns cannot publish their own memory-save transcript entries because the event key is constant for the run.
• A failed task memory-save event is not retried after the Brain summary persists.

_{Reviewed e2bffad}

[roomote-roomote]

Resolved the final task-memory review issue and pushed in e2bffad2.

• Summary CAS and task_messages MemorySaved event insertion are now atomic in one transaction. If event publication fails, the summary is rolled back and the durable drainer can retry both; a committed summary cannot strand the indication.
• Added regression coverage for event-publication failure followed by successful retry.
• Repository-skills caching remains removed from this PR and owned by PR [Fix] Fast replies wait tens of seconds on skill discovery before every turn #3066.
• Focused task-memory producer, retry/deduplication, transcript delivery, typecheck, lint, and pre-push validation pass.

The outstanding review thread is addressed; current CI for the latest head is running.