Add runtime safeguards for readonly sessions, LLM failures, and summary reuse

[junyangyan4-svg]

Summary

• make subagent and helper sessions recall-only so ephemeral work does not pollute long-term memory
• add a cooldown guard for permanent 4xx LLM API failures to stop runaway retry/token burn
• dedupe community summary generation by reusing cached summaries and embeddings when membership is unchanged

Why

• Subagents and helper sessions are useful consumers of memory, but they are poor producers of long-term memory. Marking them readonly keeps recall available while reducing noise, database churn, and unnecessary maintenance work.
• When the configured memory LLM starts returning permanent 4xx failures such as 403/404/422, the current behavior can keep hitting the provider every turn. A short cooldown turns a runaway failure mode into a visible, bounded one.
• Community summaries are expensive. If a community's members have not changed, or the same member set reappears under a new detected community id, the existing summary and embedding are still valid and can be reused.

Implementation notes

• add ReadonlySessionRegistry and skip ingest/afterTurn/compact/session_end maintenance for readonly session patterns like agent:...:subagent:..., temp:*, and slug-generator-*
• add LlmFailureGuard and pause only on permanent 4xx provider failures while resetting immediately after a successful call
• store a deterministic member_signature for community summaries, backfill existing rows, and reuse prior summaries/embeddings when the signature matches
• document the new safeguards in both README.md and README_CN.md

Verification

• npm test
• npx tsc --noEmit

[adoresever]

Thanks for the PR. There are conflicts with main now (llm.ts was updated in #38). Could you rebase onto the latest main and resolve? Also since this is a large change (+485 lines), could you split it into separate PRs for each feature (readonly sessions, LLM failure guard, summary reuse)? That would make review easier.

[junyangyan4-svg]

I've rebased onto the latest main. You can review the change now.

[junyangyan4-svg]

Superseded by split PRs per review request: #41 (readonly sessions), #42 (summary reuse), and #43 (LLM failure guard).