feat(dsl): SYS EXPLAIN REMEMBER dry-runs the pipeline

[KailasMahavarkar]

Step 2 of retrieval-observability. `SYS EXPLAIN REMEMBER "q"` runs gather + fuse + temporal filter without materializing nodes, running the reranker, expanding nucleus, or mutating recall counts. Returns candidate slot ids with per-signal scores + the Step 1 telemetry block.

Example

```python
r = gs.execute('SYS EXPLAIN REMEMBER "European capitals" LIMIT 3')

r.kind # 'plan'
r.data

{

"verb": "REMEMBER",

"query": "European capitals",

"limit": 3,

"candidates": [

{"slot": 9, "id": "mem1", "fused_score": 0.42,

"vector_sim": 0.52, "bm25_score": 0.0, "recency_score": 1.0,

"graph_score": 0.0, "co_bonus": 0.0, "recall_boost": 0.0},

...

]

}

r.meta["signals"] # same shape as real REMEMBER: fusion / recency /
# stages / reranker / nucleus / sentence-query-expansion
```

Why

Tuning fusion weights or diagnosing "why did this rank where it did" required running real REMEMBER, which bumps `recall_count` on every candidate. EXPLAIN is side-effect free - safe to call repeatedly during interactive debugging.

Also the foundation for Step 3 (`ANSWER` verb), which will consume the same candidate plan before handing context to a reader LLM.

Implementation

• `RememberQuery` handler gains internal `_plan_only=False` kwarg. When True: returns after fusion + temporal filter as `Result(kind="plan", data={candidates}, meta={signals})`. Skips materialization, rerank, nucleus, recall-count bumps, similarity buffer.
• `SYS EXPLAIN` handler dispatches `RememberQuery` inners via `self._executor._remember(inner, _plan_only=True)`.
• New `_executor` back-reference on `SystemExecutor`, wired from `store.py` at construction.
• Empty-store and empty-gather branches also respect `_plan_only`.

Tests

• `test_sys_explain_remember_returns_plan_without_side_effects` - asserts plan shape + no recall-count mutation
• `test_sys_explain_remember_empty_store_returns_empty_plan` - degenerate case

Full suite

1758 passed, 101 skipped (+2 new), zero regressions.

Follow-up

Step 3: `ANSWER` verb. Retrieves + formats context + calls reader LLM via `litellm`. Returns `{answer, cited_slots, signals}`. Reuses the plan-only path internally to grab candidates before generation.

🤖 Generated with Claude Code