docs(design): scope/collection filter evaluation (#76)

[jphein]

Summary

Evaluates the four design alternatives raised in #76 for whether mempalace_search should expose a cross-collection filter now that the storage layer supports multi-collection-per-palace as a first-class capability.

Recommendation: Option 1 today (status quo, document), with Option 2 (collections= parameter, default [\"mempalace_drawers\"]) as the deferred trigger once a second MCP-visible collection earns its read surface.

Reasoning

PR #8 (0b945e1) retired mempalace_session_recovery — both write and read paths. The fork's MCP surface now reaches exactly one collection (mempalace_drawers). The premise of #76 ("the palace has multiple physically separate stores") is true at the storage layer but no longer true at the MCP/CLI surface this fork exposes.

The recovery-collection split (2026-04-25 → 2026-05-05) is itself the case study for why Option 3 (per-collection sibling tools) goes wrong: writes shipped before the read tool did, checkpoints became invisible to mempalace_search, and the structural fix was to retire the whole side collection. That lesson, ratified in the README's "what this fork has learned" section, should set the design bar: every new sibling collection has to earn its own read tool before any write lands.

Options evaluated

1. Status quo, document. Cheapest; ratifies the post-PR-refactor(mcp): retire mempalace_session_recovery collection + read tool #8 reality.
2. collections= parameter. Generalises the retired kind= filter; ranks across collections via RRF. Forward-compatible addition.
3. Per-collection sibling tools. Most additive; echoes the recovery-collection failure mode. Deferred indefinitely.
4. Federated default. Single tool reads all collections, fused. Highest implementation cost; calibration regressions are silent. Deferred until ≥3 collections.

Includes

• Trade-offs matrix across complexity, performance, UX, backwards compat, and the "echoes a known failure mode" axis
• Impact analysis on palace-daemon /search, SME's MemPalaceDaemonAdapter, and the MCP tool surface
• Ranking concerns for cross-collection fusion (embedder identity, distance distribution comparability, metadata-filter parity)
• Explicit trigger conditions for revisiting Option 1 → Option 2
• fork-changes.yaml entry with rendered FORK_CHANGELOG.md and README.md

SME impact

Under the recommendation, the SME-side cleanup is straightforward: remove DEFAULT_KIND, remove the --kind CLI flag, delete the xfail'd test_kind_content_excludes_stop_hook_checkpoints. Unblocks PR #7's cleanup with no further coordination.

Test plan

• Confirm the design doc renders correctly in FORK_CHANGELOG.md and the README fork-change-queue table
• Verify no source code is changed (design only; runtime unchanged)
• Review recommendation against the recovery-collection retirement lesson

Closes #76.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.