fix(recall): allow exact filtering of untagged/global observations (#2295)

[nicoloboschi]

Summary

Fixes #2295. Supersedes #2322.

Allow recall to select only untagged/global observations by passing an empty tag
set ([] or omitted) together with tags_match="exact". This lets users switch
between shared and tagged observation scopes without a synthetic "global" tag or
a hand-maintained negative filter.

Root cause

The centralized tag filters treated None/[] as "no filter" in every mode,
returning every scope. There was no way to recall only the global/untagged scope —
the one that observation_scopes="shared" consolidation writes to.

Fix

In exact mode, an empty/absent tag set now means the global/untagged scope and
emits an untagged-only clause (tags IS NULL OR tags = '{}', matching both
historical NULLs and current empty-array storage). Applied consistently across:

• the flat SQL builders (build_tags_where_clause, build_tags_where_clause_simple)
• Python post-processing (filter_results_by_tags)
• compound tag-group leaves (_build_group_clause, _match_group)
• link-expansion post-filtering (run unconditionally so the empty exact scope applies)

No bind parameter is consumed for the empty-scope clause, so following parameter
indexes stay aligned. All other modes (any, all, any_strict, all_strict)
still treat empty/absent tags as "no filtering" — unchanged.

http.py field descriptions and the regenerated clients/OpenAPI/docs document the
new exact-empty scope.

Relationship to #2322

This implements the same fix as #2322 but additionally updates the RecallRequest
API descriptions and regenerates the client SDKs (Go/Python/TypeScript), OpenAPI
spec, and docs — so the user-facing surface is complete. It also adds the
parameter-offset regression test from #2322. Opening this to supersede #2322.

Tests

• SQL builders: flat + compound, None and [], param-offset preserved
• Python post-filter: NULL / empty / tagged results under exact-empty scope
• Recall API: tags=[] + tags_match="exact" returns the untagged memory and
  excludes the tagged one

Verification

• tests/test_tags_visibility.py tag-filtering tests pass (34 passed for the
  exact/untagged/empty/global subset incl. the recall API test)
• ty check passed; ruff check + ruff format clean; generated files in sync

[wangzupeng12061]

Thanks for mentioning #2322 and keeping the regression test. If appropriate, would it be possible to add me as a co-author for the root-cause investigation and the parameter-offset regression test contribution?