fix(context): preserve standalone engine compression arguments

[newjordan]

Summary

• forward only the optional compression keywords a context-engine plugin declares
• preserve /compress <focus> for engines that implement the documented focus_topic argument but not the built-in-only force extension
• propagate implementation TypeErrors instead of masking them with a second, reduced-argument call
• document the supported standalone context-engine install, activation, update, restart, and removal flow

Why

Hermes currently calls every engine with focus_topic and force, catches any TypeError, and retries with only current_tokens. That keeps older strict-signature plugins alive, but it also drops manual focus for current third-party engines and can execute a broken compressor twice while hiding its original error.

Signature-aware forwarding preserves backward compatibility without conflating an engine bug with an unsupported keyword. The documentation update makes the standalone route explicit for third-party engines, which aligns with the repository policy that external integrations should not be added under the core plugins/ tree.

This was exercised while building a standalone local embedding context engine that uses the existing ContextEngine and register_context_engine() surfaces; no product-specific code is added to Hermes core.

Validation

• uv run --python 3.11 --with pytest python -m pytest -q tests/run_agent/test_compress_focus_plugin_fallback.py tests/run_agent/test_plugin_context_engine_init.py tests/agent/test_context_engine.py tests/agent/test_compress_focus.py tests/agent/test_compression_concurrent_fork.py — 64 passed
• uvx ruff check agent/conversation_compression.py tests/run_agent/test_compress_focus_plugin_fallback.py tests/agent/test_compression_concurrent_fork.py — passed
• npm run build from website/ — Docusaurus production build passed for English and Simplified Chinese
• git diff --check — passed

[alt-glitch]

This was generated by AI during triage.

Related: this narrows the same signature-aware context-engine compatibility space as #64342 and #16450. #64342 additionally carries memory-context and broader compression work, so this remains a related subset rather than a duplicate.

[newjordan]

Superseded by #65933. After triage, the runtime portion was removed to avoid duplicating #16450 and #64342; the replacement retains only the standalone context-engine documentation. The original head branch was deleted while closing this PR, and GitHub will not reopen a PR after that branch is recreated, so #65933 carries the recovered docs-only commit.