fix: align shipped short-term memory capacity with component defaults - #1847

Purpose of the change

Normal installation examples and generated wizard configuration currently override the 64,000-character short-term-memory default with 500. That causes ordinary messages to reach the summarization threshold much earlier than the component and full configuration defaults imply.

Description

• Use 64000 in the shipped CPU/GPU/Nebula/event samples, install documentation, Helm configuration, and retrieval-agent examples.
• Remove the wizard's redundant 500 override so its partial configuration resolves through the existing default merge.
• Describe the budget accurately: Python string lengths of recent episode content plus the saved summary. It is separate from the summarizer model's context window.
• Preserve explicit user values, enabled/disabled flags, session overrides, validation bounds, and the existing summarization/eviction algorithm.
• Bump the Helm chart from 0.2.0 to 0.2.1; keep appVersion: v0.2.6.

Fixes/Closes

Fixes #1772. Related prior work: closed, unmerged #1821 proposed disabling short-term memory; this change keeps enablement intact. #1518 overlaps backend/example wiring, and #1720 concerns lifecycle behavior; neither scope is folded into this PR.

Type of change

• Bug fix
• Documentation update

How Has This Been Tested?

Linux, Python 3.12.14, frozen dependencies:

• 73 passed across wizard, configuration, episodic configuration, and existing short-term-memory tests.
• Generated Qdrant/Milvus/SQLite configurations resolve through MemMachine._with_default_episodic_memory_conf to 64000.
• Real ShortTermMemory with the existing mocked LLM does not summarize at 501 characters or exactly 64000 total characters. One additional Unicode code point triggers exactly one call, both with and without a saved summary.
• Samples load normally; explicit 500, disablement, and per-session settings remain authoritative.
• Repository-wide Ruff and formatting pass; changed Python files pass ty; git diff --check passes.

Additional macOS validation:

• helm lint deployments/helm passed.
• Default/event and declarative helm template output both loaded through the actual Configuration model and default merge, with capacity 64000 and enablement preserved. Only the container log path was relocated to a temporary host directory for validation.
• Chart version/appVersion assertions passed.
• Node 22 mint broken-links passed with no broken links.

Full base-environment ty check packages reports 23 unresolved imports for absent optional SDKs (sentence_transformers, nebulagraph_python, pymilvus, litellm, spacy), with no other diagnostic categories. The full optional-dependency and OS/Python matrix was not run locally.

• Unit Test
• Test Script (Helm render and embedded-config validation)

Checklist

• I have signed the commit(s) within this pull request
• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code
• I have made corresponding changes to the documentation
• I have added unit tests that prove my fix is effective
• New and existing unit tests pass locally with my changes (focused suites above)
• I have checked my code and corrected any misspellings

Maintainer Checklist

• Confirmed all checks passed
• Contributor has signed the commit(s)
• Reviewed the code
• Run, Tested, and Verified the change(s) work as expected

Further comments

Implementation, code review, and validation were performed with OpenAI Codex under the contributor's authorization. Human self-review remains unchecked. The commit is signed off and has GitHub's verified signature. No latency improvement or full CI result is claimed.