feat: vector search should respect --path filtering

[tenfourty]

Summary

Currently --path filtering works for FTS but is ignored by vector/semantic search. The vector pipeline scans the full embedding table and scores all documents, with path filtering only applied after ANN scoring. This means kbx search "topic" --path memory/meetings/ returns FTS results scoped to that path but vector results from the entire KB.

Proposed behaviour: Pre-filter the vector candidate set by document path before ANN scoring, so both FTS and vector search respect the same --path argument consistently.

Use Cases

• Scope to a date range of meetings: kbx search "deployment" --path memory/meetings/2026/03/ — only search March 2026 meetings
• Search only entity files: kbx search "infrastructure lead" --path memory/people/ — find people matching a role
• Agent-driven scoped search: Callers that know exactly which KB subtree is relevant can reduce noise from unrelated documents
• Project-scoped search: kbx search "blockers" --path memory/projects/ — only search project entity files

Current Behaviour

# FTS results are correctly scoped to meetings
kbx search "migration timeline" --fast --path memory/meetings/ --json
# → only results from memory/meetings/

# Hybrid search: FTS results are scoped, but vector results come from everywhere
kbx search "migration timeline" --path memory/meetings/ --json
# → FTS hits from memory/meetings/ + vector hits from memory/people/, memory/notes/, etc.

The inconsistency means --path is unreliable in hybrid mode — users can't trust that all results come from the specified path.

Implementation Notes

LanceDB Pre-Filtering

LanceDB supports where clauses on vector search that filter before ANN scoring:

# Current: no path filter on vector search
results = table.search(query_vector).limit(limit).to_list()

# Proposed: pre-filter by path prefix
results = (
    table.search(query_vector)
    .where(f"source_path LIKE '{path_prefix}%'")
    .limit(limit)
    .to_list()
)

Pre-filtering reduces the ANN candidate set, which is both more correct and potentially faster (smaller search space).

Glob Pattern Support

--path currently accepts glob patterns (e.g. memory/meetings/2026/0[1-3]/). The vector pre-filter needs to handle both forms:

Pattern	LanceDB where clause
memory/meetings/	source_path LIKE 'memory/meetings/%'
memory/meetings/2026/03/	source_path LIKE 'memory/meetings/2026/03/%'
memory/people/*.md	source_path LIKE 'memory/people/%.md' (approximate)
Complex globs	Fall back to post-filter (fetch more candidates, filter in Python)

For simple prefix patterns (the common case), translate directly to a LIKE clause. For complex globs with character classes or alternation, fetch a larger candidate set and post-filter — still better than no filtering at all.

FTS Parity

FTS already filters by path correctly. The fix is isolated to the vector search path — ensure it receives and applies the same --path argument that FTS does. The hybrid merge step should then combine two equally-scoped result sets.

Acceptance Criteria

• kbx search "query" --path memory/meetings/ returns only results from memory/meetings/ in both FTS and vector results
• kbx search "query" --path memory/people/ --json — all results have path starting with memory/people/
• Simple prefix patterns use LanceDB where pre-filter (no post-filter overhead)
• Complex glob patterns fall back to post-filter with expanded candidate set
• No regression in search performance for queries without --path
• --explain (feat: search explain/trace mode for retrieval diagnostics #68) shows the path filter applied to both FTS and vector pipelines