feat: add tiered durable state and memory bounds

README.md

@@ -375,11 +375,32 @@ probe and `/ready` reports worker availability. Poll and cleanup summaries are
 bounded in memory, and the last 200 worker crash records are synchronized with
 the durable state branch for post-restart diagnosis.
 
+Version `2026.08.22.1` adds tiered durable outcome storage and hard runtime memory
+bounds. The live `agent-state` branch contains only pending outcomes, an
+aggregate checkpoint, and the most recent terminal window; completed raw
+outcomes are sanitized, gzip-compressed, and uploaded to monthly
+`agent-archive-YYYY-MM` branches. A restart restores only the live branch, so it
+retains lifetime success statistics and architecture evidence without cloning or
+parsing the full history. The first upgrade automatically compacts a legacy
+outcome file after 2,000 records; later compactions are incremental.
+
+Candidate preparation now retains at most 100 distinct models (up to three
+routes per model) per cycle by default. ModelScope repository-tree, config,
+model-card, modification-time, and page caches use bounded LRU eviction, which
+prevents memory use from growing with every poll cycle. These limits can be
+tuned with `--candidate-pool-limit`, `MODELHUB_AGENT_OUTCOME_COMPACT_THRESHOLD`,
+`MODELHUB_AGENT_RECENT_OUTCOME_LIMIT`,
+`MODELSCOPE_DETAIL_CACHE_MAX_MODELS`, and `MODELSCOPE_PAGE_CACHE_MAX_PAGES`.
+Cold archive upload is best-effort and retried on later state synchronizations;
+the aggregate checkpoint remains the authoritative recovery input even while an
+archive shard is awaiting upload. ModelHub Git traffic is excluded from proxy
+routing in the container to avoid deployment-specific Git failures.
+
 ## Deploy
 
 Create a tag and submit the repository URL plus tag in "我的适配智能体".
 
 ```bash
-git tag -a agent-v28 -m "ModelHub agent 2026.08.21.1"
-git push origin main agent-v28
+git tag -a agent-v29 -m "ModelHub agent 2026.08.22.1"
+git push origin main agent-v29
 ```