Summary

Fixes #2295.

Allow recall to select only global/untagged observations by using an empty tag
set with tags_match="exact".

This enables users to switch between shared observations and tagged observation
scopes without adding a synthetic "global" tag or maintaining negative filters.

Root cause

The centralized tag filters treated None and [] as "no filter" before
checking the matching mode.

As a result:

• SQL retrieval applied no filter and returned every scope.
• Python post-processing returned every result.
• An exact empty compound tag group excluded all untagged results.
• Link-expansion skipped Python filtering when the tag list was empty.

This contradicted the existing scope semantics where the empty tag set
represents the global observation scope.

Fix

• Interpret an empty tag set as the global/untagged scope in exact mode.
• Match both historical NULL tags and current empty-array tags.
• Keep parameter indexes unchanged because the empty-scope SQL clause requires
  no bind parameter.
• Apply the same semantics to flat SQL filters, Python post-processing, and
  compound tag groups.
• Run link-expansion post-filtering for empty exact scopes.
• Preserve existing behavior for all other modes: empty tags still mean no
  filtering for any, all, any_strict, and all_strict.

Tests

Added regression coverage for:

• flat SQL filtering with None and [];
• bind parameter numbering after an empty exact scope;
• Python-side filtering of NULL, empty, and tagged results;
• compound exact-empty tag groups in SQL and Python;
• recall API filtering that returns an untagged memory while excluding a tagged
  memory.

Verification

Run on Linux x86_64 with Python 3.11, Node.js 22, and PostgreSQL 17 + pgvector:

• 91 passed - tests/test_tags_visibility.py
• 10 passed - tests/test_graph_filtering.py
• ty check hindsight_api passed
• Full ./scripts/hooks/lint.sh passed