test(semantic): route the semantic tool tests through the gate

[AlexMost]

Third of four PRs for #112. Converts runSearch (~120 call sites behind it) and
the six files in test/semantic/tools/ to reach tools through the registration,
so the repo-wide rule PR 4 records has no exception.

grep -rn '\.handler(' test/semantic/ now returns nothing; the only surviving
handler-direct calls in the repo are the two commented edit_note WRITE_FAILED
tests whose subject is the CallToolResult envelope.

Triage — every previously-green test the migration broke

Seven tests failed once their calls crossed the gate. None was relaxed; each was
either a false pin on an unreachable handler check, or a fixture passing a
parameter the advertised schema does not declare.

Unreachable handler checks (5). The schema already states the constraint, so
registerTool's .strict() wrapper rejects with INVALID_PARAMS and
details.issues before the handler's own check can run. The pins on
INVALID_ARGUMENT were pinning a branch no client can reach. Rewritten to assert
INVALID_PARAMS plus the offending field — the same triage read_notes got in
PR 1 (task 1.4), and exactly the divergence ADR-0015 is being written to record.

Test	Schema constraint	Was	Now
rejects thresholds below 0 and above 1	threshold: z.number().min(0).max(1)	INVALID_ARGUMENT	INVALID_PARAMS, path: 'threshold'
rejects an empty query array	query: …array(z.string()).min(1).max(8)	INVALID_ARGUMENT	INVALID_PARAMS, path: 'query'
rejects a query array longer than 8	same	INVALID_ARGUMENT	INVALID_PARAMS, path: 'query'
handler treats empty path_prefix array as empty filter	filter.path_prefix .min(1)	INVALID_ARGUMENT	INVALID_PARAMS, path: 'filter.path_prefix'; renamed to rejects an empty path_prefix array at the gate
handler treats empty exclude_path_prefix array as empty filter	filter.exclude_path_prefix .min(1)	INVALID_ARGUMENT	INVALID_PARAMS, path: 'filter.exclude_path_prefix'; renamed likewise

The dead code these expose is left in place deliberately — readThreshold's
value < 0 || value > 1 branch in src/modules/semantic/tool-helpers.ts and the
empty-path_prefix arm of the filter fallback are now provably unreachable
through the gate. Removing them is a source change, not a test migration; it
belongs with the same sweep as the set_property finding below.

Fixtures naming the only vault (3). In single-vault mode vaultParamShape
contributes no vault parameter at all, so { vault: 'v', … } is an unknown key.
The subject of each test was the lexical-only fallback / semantic_status /
missing backend — never vault routing — so the redundant argument was dropped and
the assertion kept intact. Same shape as PR 2's VAULT_NOT_FOUND triage.

• search-notes.test.ts — returns lexical-only matches (no throw) when vault has no semantic backend
• search-notes.test.ts — reports unavailable when the semantic module is globally off (no backend)
• get-similar-notes.test.ts — throws SEMANTIC_INDEX_NOT_FOUND when vault has no semantic backend

New coverage

Per task 3.4, a rejects a vault argument in single-vault mode case was added to
get-similar-notes.test.ts and find-duplicates.test.ts, asserting
INVALID_PARAMS with path: '<root>' and a message naming vault. Net test
count: 1331 → 1333.

Notes

• find_duplicates and get_similar_notes resolve with an array, and
  toToolResponse only populates structuredContent for a plain record — so
  callTool reads them off the text channel, as the helper's doc comment
  describes. Expected, not a gate bug. Their result types are module-internal, so
  the tests derive them from the tool's own handler rather than exporting from src/.
• Assertions on reg.spec.inputSchema (hybrid) and registerTool(tool).spec.description
  (search-notes) were left untouched — they already cross the gate. The raw
  tool.inputSchema.safeParse assertions likewise stay on the unwrapped schema,
  which is their actual subject.
• No schema, response shape, or error code was changed. Every commit is green on
  npm test && npm run lint && npm run typecheck.

Carried forward, not fixed here

set_property is the only caller of inferTypeAndValidate, and all three of its
UNSUPPORTED_VALUE_TYPE throws are unreachable under the declared value union
(null/undefined, a non-string/number list element, a non-scalar value — the gate
cuts all three first). The code is still declared in OperationsErrorCode. Found
in PR 2; deserves its own change rather than a drive-by.

Refs #112

🤖 Generated with Claude Code