fix(snowflake): unblock multi-schema ingest and relationship discovery (#204)

* feat(setup): drop redundant Snowflake schema prompt; fall back to free-text on listSchemas failure

Snowflake setup previously asked for a single schema as free text, then
ran a multiselect against the discovered schemas — two schema questions
back-to-back, with the first being only a session bootstrap. The SDK's
`schema` is optional, so the bootstrap step is unnecessary.

- Remove the free-text Snowflake schema prompt; only pass `schema` to
  snowflake-sdk when one is configured.
- When `listSchemas()` fails (e.g. role lacks SHOW SCHEMAS), prompt the
  user for a comma-separated list, persist it as `schema_names`, and use
  it as both the table-list filter and the multiselect default. Applies
  to every driver with a scope-discovery spec, not just Snowflake.
- Update docs to lead with `schema_names`; keep `schema_name` as a
  documented single-schema shorthand.

* fix(snowflake): keep introspecting when primary-key discovery is denied

The PK query joins INFORMATION_SCHEMA.TABLE_CONSTRAINTS and
INFORMATION_SCHEMA.KEY_COLUMN_USAGE, which require grants the
connection role may not have. Previously a 'SQL compilation error:
Object ANALYTICS.INFORMATION_SCHEMA.KEY_COLUMN_USAGE does not exist
or not authorized' aborted the entire introspect — schemas, columns,
and row counts were all discarded over a missing nice-to-have.

Wrap the constraint query in try/catch, log a one-line warning per
schema, and return an empty PK map. Columns end up with
primaryKey=false; relationship inference still has FK and profiling
to fall back on.

* fix(scan): unblock relationship discovery on Snowflake

Two adjacent bugs prevented the scan's relationship pipeline from producing
any joins on a Snowflake warehouse:

- relationship-profiling.ts fell through to a default `GROUP_CONCAT` branch
  for unknown drivers. Snowflake has no GROUP_CONCAT, so every per-table
  profile query failed with "Unknown function GROUP_CONCAT". Add an explicit
  Snowflake branch that uses LISTAGG with a literal '\x1f' delimiter
  (Snowflake requires the delimiter to be a constant, so CHR(31) is rejected).
- description-generation.ts destructured `connector.sampleTable` and
  `connector.sampleColumn` into bare locals, losing the `this` binding when
  the class-method connectors (Snowflake, Postgres, MySQL) were invoked.
  Every sample call threw "Cannot read properties of undefined (reading
  'assertConnection')" and degraded LLM descriptions to metadata-only
  prompts. Call the methods through the connector instead.

Without these, even after the primary-key probe is allowed to fail softly,
the scan ends up with 0 validated relationships and an empty `joins:` block
in every shard YAML.

* test(scan): cover table-ref helpers

* feat(scan): plumb tableScope through live-database introspection port

* feat(scan): apply tableScope during metadata fetch

* feat(scan): enforce table scope at fetch boundary

* feat(scan): pool Snowflake sessions and batch enrichment for faster ingest (#206)

* feat(cli): add RSA key-pair auth option to Snowflake setup wizard

Extends the interactive Snowflake setup flow with an authentication-method
prompt (password vs RSA/JWT key-pair). The RSA branch collects a private-key
path (env/file/absolute) and an optional passphrase; the resulting connection
config records `authMethod: 'rsa'` with `privateKey` and `passphrase` instead
of `password`.

* feat(scan): pool Snowflake sessions

* fix(scan): reuse structural snapshots and cleanup connectors

* feat(scan): parallelize relationship profiling

* feat(scan): batch table description generation

* docs: document Snowflake ingest concurrency knobs

* fix(scan): close Snowflake ingest perf verification gaps

* fix(scan): keep batched description failure bounded

* feat(scan): dispatch query-history probes by connection driver

Extract historic-sql dialect resolution into a shared helper so the
status-project readiness check and the local ingest factory agree on
which connections enable query history and which probe to run. The
status command now picks the postgres/snowflake/bigquery probe based on
the connection's driver instead of always reporting against postgres,
which previously caused snowflake connections with queryHistory.enabled
to surface a misleading "driver is snowflake" failure.

Also drops a noisy console.warn from Snowflake primary-key discovery —
INFORMATION_SCHEMA.KEY_COLUMN_USAGE is commonly ungranted for read-only
roles and the FK + profiling paths handle the empty PK map already.

* fix(llm): allow StructuredOutput tool and raise maxTurns for generateObject

The Claude Code agent SDK announces an internal pseudo-tool named
StructuredOutput in the system/init message whenever outputFormat is set
to { type: 'json_schema' }. The runtime's isolation check built its
allowedToolIds set only from MCP tool ids and treated StructuredOutput
as an unexpected host-injected tool, so every generateObject call threw
"Claude Code runtime isolation failed: tools=StructuredOutput ..." and
the table-descriptions and relationship-LLM-proposal enrichment stages
recorded null output across the board.

Whitelist StructuredOutput specifically in generateObject's
allowedToolIds — the check also enforces missing_tools symmetry, so
generateText and runAgentLoop, which do not see StructuredOutput, must
not require it.

generateObject also ran with maxTurns: 1, which the model intermittently
breached when it emitted thinking text before the structured response.
Raised to 5 to give the schema-bound call enough headroom without
allowing unbounded loops. The existing tests now exercise the path with
an init message that announces StructuredOutput so the regression cannot
slip back in.

* chore(scripts): add ktx-reset.sh project-cleanup helper

Convenience script for repeatable ingest testing: takes a project
directory and prunes everything except ktx.yaml and .ktx/secrets/, so
the next ktx setup or ktx ingest run starts from a known-clean state.

packages/cli/src/context/llm/claude-code-runtime.test.ts

@@ -91,9 +91,14 @@ describe('ClaudeCodeKtxLlmRuntime', () => {
     });
   });
 
-  it('validates structured output with the caller schema', async () => {
+  it('validates structured output with the caller schema and whitelists the SDK StructuredOutput tool', async () => {
     const schema = z.object({ answer: z.string() });
-    const query = vi.fn((_input: any) => stream([initMessage(), resultMessage({ structured_output: { answer: 'yes' } })]));
+    const query = vi.fn((_input: any) =>
+      stream([
+        initMessage({ tools: ['StructuredOutput'] }),
+        resultMessage({ structured_output: { answer: 'yes' } }),
+      ]),
+    );
     const runtime = new ClaudeCodeKtxLlmRuntime({
       projectDir: '/tmp/project',
       modelSlots: { default: 'sonnet' },
@@ -341,7 +346,10 @@ describe('ClaudeCodeKtxLlmRuntime', () => {
   it('passes scrubbed env to object generation and agent loops', async () => {
     const schema = z.object({ answer: z.string() });
     const objectQuery = vi.fn((_input: any) =>
-      stream([initMessage(), resultMessage({ structured_output: { answer: 'yes' } })]),
+      stream([
+        initMessage({ tools: ['StructuredOutput'] }),
+        resultMessage({ structured_output: { answer: 'yes' } }),
+      ]),
     );
     const objectRuntime = new ClaudeCodeKtxLlmRuntime({
       projectDir: '/tmp/project',

packages/cli/src/context/llm/claude-code-runtime.ts

@@ -47,6 +47,13 @@ const BUILTIN_TOOLS = [
 
 const KTX_MCP_SERVER_NAME = 'ktx';
 
+// SDK-internal pseudo-tool that the Claude Code CLI announces in its
+// system/init message whenever outputFormat: { type: 'json_schema' } is set.
+// Structured output is returned via result.structured_output (not through
+// canUseTool), so the tool only needs to be whitelisted for generateObject's
+// init isolation check; generateText / runAgentLoop never see it.
+const STRUCTURED_OUTPUT_TOOL_NAME = 'StructuredOutput';
+
 function isResult(message: SDKMessage): message is SDKResultMessage {
   return message.type === 'result';
 }
@@ -238,7 +245,12 @@ export class ClaudeCodeKtxLlmRuntime implements KtxLlmRuntimePort {
         projectDir: this.deps.projectDir,
         model: modelForRole(this.deps.modelSlots, input.role),
         env: this.deps.env,
-        maxTurns: 1,
+        // Structured output occasionally takes more than one assistant turn —
+        // the model may emit thinking/text before the StructuredOutput tool
+        // call, or the SDK may count assistant + tool-result as separate turns.
+        // 5 leaves headroom without enabling unbounded loops; the json_schema
+        // constraint still forces the final answer to be the schema.
+        maxTurns: 5,
         tools: input.tools,
       }),
       outputFormat: { type: 'json_schema' as const, schema: jsonSchema(input.schema as z.ZodType) },
@@ -247,7 +259,7 @@ export class ClaudeCodeKtxLlmRuntime implements KtxLlmRuntimePort {
       query: this.runQuery,
       prompt: [input.system, input.prompt].filter(Boolean).join('\n\n'),
       options,
-      allowedToolIds: new Set(mcpToolIds(input.tools ?? {})),
+      allowedToolIds: new Set([...mcpToolIds(input.tools ?? {}), STRUCTURED_OUTPUT_TOOL_NAME]),
       expectedMcpServerNames: expectedMcpServerNames(input.tools),
     });
     const error = resultError(result);