Which issue or RFC does this PR close?

Closes #1322.

Rationale for this change

Caller-controlled scope_id values could permanently populate several process-level Runtime and relational-provider caches. A long-running HTTP or MCP Server therefore exhibited linear memory growth when receiving valid requests with many one-off scope IDs, even when no Source or Memory data was persisted.

The Runtime needs a bounded lifecycle for inactive scope-local objects while preserving per-scope serialization for operations that currently hold or are waiting for a lock.

What changes are included in this PR?

• Add a lease-aware LRU cache for scope-local Runtime resources.
  • Retain up to 128 inactive scopes by default.
  • Never evict a scope while an operation holds or waits for its serialization lock.
  • Allow concurrent active scopes to temporarily exceed the configured capacity, then trim the cache when operations complete.
• Evict the corresponding relational context, source lock, activation lock, and experience lock together.
• Clear cached scope resources when the Runtime closes.
• Keep nested scoped operations within the same Runtime lifecycle so graceful shutdown waits for all related work.
• Add RuntimeConfig.scope_cache_size and the POWERCONTEXT_SERVER_RUNTIME_SCOPE_CACHE_SIZE environment variable.
• Add the low-cardinality Prometheus metric:
  • powercontext_server_runtime_scopes{state="active"}
  • powercontext_server_runtime_scopes{state="cached"}
• Ensure metrics never use scope_id as a label.
• Document the new configuration in the English and Chinese configuration references.
• Add regression coverage for:
  • many one-off scope IDs through the public HTTP boundary;
  • cache convergence to the configured capacity;
  • lock holders and waiters remaining protected during cache pressure;
  • environment-based configuration.

Are there any user-facing changes?

Yes, all changes are additive:

• RuntimeConfig.scope_cache_size defaults to 128.
• Server deployments can configure it with POWERCONTEXT_SERVER_RUNTIME_SCOPE_CACHE_SIZE.
• /metrics now exposes active and cached scope counts using a bounded state label.

There are no breaking HTTP, MCP, OpenAPI, or persisted-data changes. If an inactive scope is evicted, its ephemeral services are recomposed on the next request; persisted scope data is unaffected.

How was this change tested?

Targeted Runtime, provider, Server, and metrics tests:

.venv/bin/pytest -q \
  tests/builtin/runtime \
  tests/builtin/persistence/test_provider.py \
  tests/e2e/test_builtin_runtime.py \
  tests/test_server.py \
  tests/test_server_metrics.py

Result: 67 passed.

Repository validation:

UV_CACHE_DIR=/tmp/powercontext-uv-cache uv lock --locked
.venv/bin/prek run -a
.venv/bin/ty check
.venv/bin/zensical build -s
.venv/bin/python -m pytest --doctest-modules \
  --deselect tests/builtin/inference/test_pydantic_ai.py::test_instrumented_generation_spans_exclude_schema_retry_content

Result: 464 passed, 7 skipped, 1 deselected.

The deselected inference tracing test was also run independently and fails because the current Pydantic AI instrumentation includes schema-retry content in span attributes. No inference or tracing implementation is changed by this PR.

Manual validation sent 250 successful POST /v1/context/prepare requests using distinct scope IDs:

• Before: Runtime locks, contexts, source locks, and activation locks each grew to 250.
• After: cached Runtime scopes and the corresponding provider resources converged to the default limit of 128, with 0 active scopes after the requests completed.

AI usage statement