fix(mcp): harden semantic search inputs and store-error handling

[DeusData]

What does this PR do?

Distilled from #938 with co-author credit to @dpersek — the four residual hardening points that were still open after the semantic-search work landed on main, adapted to the current design (semantic_limit/semantic_offset, sg_semantic_page_t, the certified top-K fetch loop). Refs #915.

1. extract_semantic_keywords (src/mcp/mcp.c) — a semantic_query array containing a non-string element (["publish", 42], [null], [["publish"]]) was silently narrowed to its string members, so a malformed call ran as a different query and the caller never learned. It now returns -1 and search_graph answers with the same isError response as a bare-string semantic_query (message extended to say every element must be a string).

2. run_semantic_query_core fails closed on a store error — a vector scan that FAILED was rendered as "0 semantic matches" plus the "use a moderate/full index or broader keywords" hint, sending the caller off to broaden keywords against a broken index. cbm_store_vector_search now distinguishes the two states main conflated:

   • no node_vectors table at all (lean index) → CBM_STORE_NOT_FOUND → unchanged behaviour: empty page + the moderate/full-index hint. On main this case only worked because the no such table prepare failure was swallowed as zero matches; a sqlite_master probe makes it explicit.
   • the table exists but the scan fails → CBM_STORE_ERR → isError "semantic search failed …", for semantic-only and combined (structural + semantic) calls alike — no structural page pretends the semantic half found nothing.

3. vs_min_cosine_score (src/store/store.c) — the keyword L2 norms and the node norm were recomputed inside the per-(node, keyword) loop. Keyword norms are now computed once per search (vs_keyword_norms), the node norm once per node. Math-identical: the integer sums are exact, so sqrt((double)ma) * sqrt((double)mb) yields the same doubles whether or not it is factored out of the pair loop.

4. vs_append_result — strdup results were unchecked. It now returns CBM_STORE_OK/CBM_STORE_ERR, frees the partial row on failure, and only counts a row once every copy succeeded, so the caller's existing free path stays valid.

Left out of #938 (already covered or contradicted by main's current design): its three run_semantic_query_core call sites (main has one), the fixed 250-candidate window (main certifies top-K via the fetch-expansion loop so limit cannot change ranking), and the row memset (a failed row is never counted, so it is never read).

Verification

• Reproduce-first: tool_search_graph_semantic_query_rejects_non_string_elements and tool_search_graph_semantic_store_error_fails_closed (mcp suite) were built against main's src/ with only the tests applied — exactly those two fail (382 passed, 2 failed, 4 skipped, both at the isError assertion). With the change: mcp mcp_mutation_guard store_search cli → 714 passed, 0 failed, 4 skipped (the four are pre-existing Windows-platform skips).
• make -f Makefile.cbm lint-ci green (cppcheck, clang-format, NOLINT check); scripts/check-no-test-skips.sh OK; no raw fopen( in the diff.

Checklist

• Every commit is signed off (git commit -s) — required, CI rejects
  unsigned commits (DCO, see CONTRIBUTING.md)
• Tests pass locally (make -f Makefile.cbm test) — the four suites exercising semantic search; full 3-OS ladder is left to CI
• Lint passes (make -f Makefile.cbm lint-ci)
• New behavior is covered by a test (reproduce-first for bug fixes)