temperature and top_p sent together break every current Claude model via the custom provider

## Summary

When configuring the built-in AI investigation agent with `LLM_PROVIDER=custom` pointed at Anthropic's OpenAI-compatible endpoint (`https://api.anthropic.com/v1`), every agent run fails immediately. The UI shows a generic error:

> The model provider is not configured or could not be reached.

The actual server-side error (visible in container logs) is:

```
LLMStreamError: `temperature` and `top_p` cannot both be specified for this model. Please use only one.
```

(For `claude-sonnet-5` specifically, `temperature` alone is already rejected: `` `temperature` is deprecated for this model. ``)

## Root cause

The agent's default sampling settings (bundled `model-tuning` config, `defaults.sampling` = `{temperature: 0, topP: 1}`) are applied per model via `samplingFor()`, and the zod schema backing both the bundled tuning table **and** the operator-supplied `AGENT_MODEL_TUNING_PATH` override requires both fields as non-optional numbers:

```ts
strictObject({
temperature: z.number().min(0).max(2),
topP: z.number().min(0).max(1),
})
```

There's no way, via `AGENT_MODEL_TUNING_PATH` or any other configuration, to send only one of the two — both are always forwarded as `temperature` + `top_p` in the OpenAI-compatible request body.

Anthropic's API (tested on `claude-sonnet-5`, `claude-sonnet-4-6`, `claude-haiku-4-5` — current-generation models) rejects any request specifying both parameters simultaneously:

```json
{"error":{"code":"invalid_request_error","message":"`temperature` and `top_p` cannot both be specified for this model. Please use only one.","type":"invalid_request_error","param":null}}
```

`claude-sonnet-5` additionally rejects `temperature` outright even alone (it's an adaptive-thinking model; sampling params were removed for it).

## Steps to reproduce

1. Get an Anthropic API key from console.anthropic.com.
2. Set:
```
LLM_PROVIDER=custom
LLM_API_URL=https://api.anthropic.com/v1
LLM_MODEL=claude-haiku-4-5
LLM_API_KEY=
```
3. Start an AI investigation against any connected database.
4. UI shows: "The model provider is not configured or could not be reached."
5. Server log shows the real cause (see Summary above).

## Confirmed independent of app version

Compared the compiled `model-tuning` and agent request-construction chunks between `latest` (0.15.0) and `dev` (0.16.0) — byte-identical output (same webpack content hash for those modules), so this affects both.

## Environment

- Image: `ghcr.io/libredb/libredb-studio:latest` (0.15.0) and `:dev` (0.16.0)
- Confirmed via direct calls to `api.anthropic.com` (bypassing the app) that: the API itself works fine when only one of the two params is sent, and that the plain non-tool-calling `custom` provider path (which only ever sends `temperature`) works correctly. The bug is specific to the tool-calling agent path, which always sends both.

## Suggested fix

Either:
- Make `topP` optional in the model-tuning schema (send `top_p` only when explicitly configured for a model), defaulting to omitting it, or
- Add provider/endpoint-specific request shaping — skip `top_p` (or `temperature`) when the target API returns this specific 400, or when `LLM_API_URL` targets a known Anthropic-compatible endpoint.

Start with the model-tuning configuration and the samplingFor() path described in the issue, then trace how the zod schema and AGENT_MODEL_TUNING_PATH values become the tool-calling request body. Done means the custom Anthropic-compatible agent no longer sends incompatible sampling parameters together, while supported model configurations continue to work.

Written by the indexing model from the issue text.