Artifact mcp capability: authenticated tool calls fail with generic tool_error while the identical calls succeed outside the artifact bridge (Kite/Zerodha connector)

### What happened?

get_profile takes no arguments and returns basic account info — it isn't a market-data or high-load endpoint, yet it fails identically to the others. Rules out argument-shape and rate-limit theories specific to any one tool.
The one tool that succeeds (search_instruments) is the one Kite tool that doesn't require a fresh authenticated call to Zerodha (it's answered from a cached/static instrument master).

### What did you expect to happen?

mcp.callTool("Kite", "get_quotes", {...}) (and the other authenticated tools) should succeed through the artifact bridge, the same as it does when the assistant calls the connector directly, since it's the same connected account/session.

Actual behavior
code: tool_error
message: Failed to get quotes // (or "Failed to execute get_profile" / "Failed to get latest trading prices" /
// "Failed to get OHLC data" / "Failed to get historical data", per tool)
server: Kite
retryable: undefined

search_instruments alone succeeds and returns real data.

### Steps to reproduce

- Connect the Kite (Zerodha) MCP connector to your Claude.ai account (Settings → Connectors).
- Publish an Artifact declaring capabilities: {mcp: {servers: [{server: "Kite", tools: ["get_profile","get_quotes","get_ltp","get_ohlc","get_historical_data","search_instruments"]}]}}.
- In the page, call await claude.use("mcp") then mcp.callTool("Kite", "", ) for each of the six tools above.
- Observe: get_profile (zero arguments), get_quotes, get_ltp, get_ohlc, and get_historical_data all reject with {code: "tool_error", ...}. Only search_instruments resolves successfully.
- In the same Claude.ai session, ask the assistant (in chat, not inside the artifact) to call the same Kite tools directly (e.g. get_profile, get_quotes) — they succeed instantly with real account/market data.
- Fully disconnect and reconnect the Kite connector (Settings → Connectors) and repeat step 3 — behavior is unchanged.

### Area

MCP Connector (adding/managing servers)

### MCP Server (if applicable)

Kite/Zerodha connector

### Error messages or logs

```shell
https://claude.ai/code/artifact/d0e43904-1424-483a-8b45-45eae8d87022
```

### Additional context

_No response_

Start with the artifact URL and reproduce the six Kite/Zerodha calls, comparing them with direct connector calls in the same Claude.ai session. No repository file or test is named; done means determining why authenticated calls fail through the artifact bridge and confirming that those calls succeed there without breaking search_instruments.

Vom Indexierungsmodell aus dem Issue-Text verfasst.