fix: rename deduplication tool to memory_deduplicate to avoid collision

[Displayer226]

Summary

This PR resolves the naming conflict for the memory_compact tool by renaming the metadata-based deduplication version to memory_deduplicate.

Motivation

Fixes #431.
This is offered as an alternative to PR #433.

While #433 proposes removing one of the implementations, this PR argues that both are essential for a complete memory lifecycle:

1. Deduplication (memory_deduplicate): A lightweight, metadata-based tool in src/tools.ts used for structural hygiene (archiving identical or redundant facts).
2. Summarization (memory_compact): A heavyweight, LLM-based tool in index.ts used for increasing knowledge density by merging related memories into refined entries.

By renaming instead of removing, we preserve both functionalities for the user.

Key Changes

• Renamed registerMemoryCompactTool to registerMemoryDeduplicateTool in src/tools.ts.
• Changed the exposed tool name from memory_compact to memory_deduplicate.
• Updated registerAllMemoryTools to reflect the new name.

Testing

• Verified that both tools can now be registered simultaneously without gateway startup errors.
• Manual test of memory_deduplicate confirmed it correctly identifies and archives duplicate metadata entries.

[rwmjhb]

Review: REQUEST-CHANGES

The rename approach is the right call — both tools serve different purposes (dedup vs progressive compaction), so they should coexist under distinct names. Closing #433 in favor of this direction.

Must fix:

1. Governance test is broken — test/memory-governance-tools.test.mjs:157 still asserts memory_compact is registered. After the rename, this test fails with AssertionError: tool memory_compact should be registered. This file is also not in the default npm test script, so CI won't catch it.

2. Parameter style mismatch — the renamed memory_deduplicate uses camelCase dryRun (default: preview), while the surviving memory_compact uses snake_case dry_run (default: write). A caller using { dryRun: true } against memory_compact silently gets write behavior instead of preview.

Worth considering:

• The rename is a breaking change for any existing user or LLM agent calling memory_compact expecting dedup behavior — consider documenting the migration in the PR description or a changelog note.
• Branch is behind main — please rebase.

[Displayer226]

Hello !

Thanks for the review! I've updated the PR to address all the 'Must fix' items and rebased on the latest master:

• Parameter Standardization: Standardized memory_compact parameters to camelCase (dryRun, minAgeDays, similarityThreshold) and
  set dryRun: true as the default for safety, matching the memory_deduplicate style.
• Fixed Governance Test: Updated test/memory-governance-tools.test.mjs to target the correct tool name.
• Changelog: Added a note in CHANGELOG.md regarding the breaking rename and parameter changes for version 1.1.0-beta.11.
• Rebase & CI Cleanup: Rebased on master and fixed two existing test regressions in plugin-manifest-regression and
  session-summary-before-reset caused by recent changes to selfImprovement hook defaults.
• Extra Verification: Added a new check in test/plugin-manifest-regression.mjs to ensure both tools remain correctly registered with their new camelCase interface and safe defaults going forward.

The full test suite is now passing locally (npm test).

[AliceLJY]

LGTM — clean rename + parameter standardization, addresses all prior review feedback.

Verified:

• memory_compact (progressive summarization) and memory_deduplicate (metadata dedup) now coexist with distinct names
• Parameters consistently use camelCase across both tools
• dryRun defaults to true (safe preview mode) for both tools
• Governance test updated to target memory_deduplicate
• Plugin manifest regression test extended with assertions for both tools' parameter schemas
• CHANGELOG clearly documents the breaking changes
• selfImprovement: { enabled: false } additions fix unrelated test regressions from upstream

Minor note (non-blocking): In plugin-manifest-regression.mjs, the updated registerTool mock silently drops calls where toolOrFactory is a function but meta is undefined (neither if branch matches). Original code would have thrown on meta.name. Since this is test-only, low risk, but worth noting in case a future tool registration path hits that gap.

Assigning to @rwmjhb for final merge decision.