feat(search): normalize semantic query (NFKC + casefold)

[lg320531124]

Problem

Semantic find/search queries are passed verbatim to the embedder and retriever, so colloquial or CJK full/half-width variants of the same intent do not match indexed content. Real examples from daily usage:

Query typed	Indexed as	Match?
ＯｐｅｎＶｉｋｉｎｇ (full-width)	OpenViking	❌
OpenVAKING (case variant of a typo)	openvaking	❌
Harms agent (double space)	hermes agent	❌

search_service.py only does query.strip(); the hierarchical retriever forwards the query verbatim. The embedder tokenizes "ＯｐｅｎＶｉｋｉｎｇ" and "OpenViking" to different vectors, so recall silently drops on CJK-keyboard and casual input.

Change

Add _normalize_search_query(query) (NFKC + casefold + whitespace collapse + strip) and apply it at the two semantic entry points in viking_fs.py:

• VikingFS.find (after _ensure_non_empty_search_query)
• VikingFS.search (after _ensure_non_empty_search_query)

def _normalize_search_query(query: str) -> str:
    if not query:
        return query
    return re.sub(r"\s+", " ", unicodedata.normalize("NFKC", query).casefold()).strip()

• NFKC folds full/half-width Latin/digits to canonical ASCII
• casefold handles Unicode case folding (stronger than str.lower(), e.g. ß → ss)
• whitespace collapse + strip removes accidental double spaces / leading-trailing gaps
• Idempotent and only widens recall — already-normalized ASCII is byte-for-byte unchanged

Why this layer

Applied at the viking_fs layer (not SearchService) because session/memory/tools.py calls viking_fs.search directly, bypassing SearchService. Fixing it at the lowest shared entry point covers all callers.

Why grep is left alone

grep takes a regex pattern. NFKC + casefold would corrupt explicit character classes (e.g. [A-Z], [Ａ-Ｚ] collapse differently), and grep already exposes a case_insensitive flag for case folding. Normalizing there would be a behavior change, not a recall fix.

Tests

tests/unit/test_search_query_normalization.py — 7 cases, all pass:

• CJK full-width → ASCII (ＯｐｅｎＶｉｋｉｎｇ → openviking)
• mixed-case casefold (OpenViking, OpenVAKING)
• whitespace collapse (Harms agent → harms agent)
• leading/trailing strip (\tquery\n)
• empty + None passthrough
• idempotency

======================== 7 passed, 4 warnings in 2.28s =========================

Pure-function unit tests (no vectordb/embedder fixture) — the helper is side-effect-free, so integration tests would pay full storage init cost for a string transform.

Scope

+22 lines in viking_fs.py, 1 new test file. No public API change, no config knob, no behavior change for already-normalized queries.

Related

Continues the search-recall robustness line of #2874 (tool-input compaction) and #2900 (grep empty-result fallback). Does not overlap either.

[lg320531124]

CI note: the API & CLI Integration Tests job is red, but this is a pre-existing CI-infra flake, not introduced by this PR.

The failure is identical across all currently-open PRs (#2874, #2936, #2937, #2938) and reproduces on #2874 which predates my other work:

1. Build ragfs-python native extension → ❌ ERROR: maturin build produced no wheel (the Rust→Python binding fails to build a wheel in CI)
2. test_cli_search.py::TestSearchGrep::test_grep_basic / test_grep_case_insensitive → ERROR at the add-resource setup step (AssertionError: add-resource failed after retries)
3. find/search tests → SKIPPED (Upstream ... marker — the upstream VikingDB service is unreachable from the runner)

This PR is pure-Python (touches only 'openviking/session/memory/tools.py'-style modules), does not touch crates/, maturin, the ov CLI binary, or any CI workflow — so it cannot be the cause of a Rust-binding build failure or an upstream-service reachability issue.

Evidence that maintainers are not blocking on this flake: #2934 was merged today (2026-07-02 01:07 UTC) with the same API & CLI Integration Tests → FAILURE status, and check-deps → SUCCESS + main-branch 02. Main Branch Checks → success both stay green.

Local unit tests for this PR pass (see PR body). Happy to help root-cause the maturin/upstream flake separately if useful — just let me know.