fix: bound memory capture and recall output

[xdeepsnyx]

Summary

• Preserve explicit Supermemory capture mode and prevent auto/session-end ingestion when automatic capture is disabled.
• Harden Supermemory explicit writes with safe metadata and bound profile output to visible result counts.
• Add a shared memory/search output cap and apply it to session search and provider-backed memory search/profile tools.

Motivation

Provider-backed memory can become counterproductive if automatic capture ignores conservative settings or if recall/search payloads grow large enough to flood active context. This PR keeps curated memory useful while bounding automatic ingestion and recall output.

Changes

• memory.auto_capture: false now gates Supermemory on_session_end ingestion.
• capture_mode: explicit is preserved by config normalization and skips automatic turn capture.
• Explicit Supermemory writes add non-sensitive metadata such as saved_at, platform, and origin where available.
• Supermemory profile output is limited to the configured recall count while still returning total counts and a truncated flag.
• Adds memory.search_result_char_limit with a default of 10,000 characters.
• Adds tools.memory_search_caps to cap memory/search-style tool results with a visible truncation notice.
• Applies the cap to session_search and memory-provider search/profile/recall-like tools.

Test plan

env -u SUPERMEMORY_CONTAINER_TAG \
    -u SUPERMEMORY_CUSTOM_CONTAINERS \
    -u SUPERMEMORY_ENABLE_CUSTOM_CONTAINER_TAGS \
    python -m pytest \
      tests/plugins/memory/test_supermemory_provider.py \
      tests/tools/test_session_search.py \
      -o 'addopts=' -q

Result: 72 passed in 1.85s

Closes #21690

[teknium1]

Thanks for addressing both Supermemory capture controls and recall growth.

Problems

• capture_mode: explicit is preserved and skips sync_turn, but the changed on_session_end() guard only adds _auto_capture; with auto_capture=true, explicit mode still calls ingest_conversation() automatically (plugins/memory/supermemory/__init__.py, changed session-end guard; current equivalent at :720-744).
• The new session_search cap is applied only to browse/discovery returns. Current routing returns _scroll() and _read_session() directly (tools/session_search_tool.py:678-739), leaving those message-bearing paths unbounded.
• The session-search test is based on _summarize_session, which current main no longer contains; current main uses the four-shape dispatcher at tools/session_search_tool.py:619-740.

Suggested changes

• Apply one automatic-capture predicate, including explicit mode, to every automatic ingest path and test session-end behavior.
• Bound all four session-search shapes using valid structured output and truncation metadata, then test oversized read and scroll results.

Automated hermes-sweeper review.

[teknium1]

This guard still permits automatic session-end ingest when capture_mode == "explicit" and auto_capture is true. If explicit means explicit writes only, include that mode here as well and add a session-end regression test.

[teknium1]

This helper is only used for browse/discovery returns in this patch; the scroll and read shapes still return unbounded payloads. Apply the boundary policy to every session_search shape and preserve valid structured output after truncation.