feat(routing): add grob_autotune MCP tool for offline classifier calibration (T-P5b)

[Destynova2]

Summary

Closes the Phase P routing-intelligence work. Pairs with #268 (T-P5a) by exposing a higher-level MCP surface dedicated to classifier tuning.

Two actions

`action=suggest` — returns the seven classifier weights/thresholds (5 weights + 2 tier thresholds) as `TuneSuggestion` entries:

```json
{
"action": "suggest",
"suggestions": [
{ "key": "weights.tools", "current": 1.0, "proposed": 1.0,
"rationale": "current value; manual tuning recommended (see auto-tune-routing.md)" },
...
]
}
```

The MVP reports `proposed == current`; future revisions will infer patches from observed traffic. Operators can iterate manually using the rationale field as a placeholder.

`action=apply` — accepts a list of `{key, value}` patches and persists them via the existing `config_guard::persist_and_reload` pipeline:

```json
{
"action": "apply",
"patches": [
{ "key": "weights.tools", "value": 5.0 },
{ "key": "thresholds.complex_threshold", "value": 6.0 }
]
}
```

This is sugar over batching multiple `grob_configure` calls into one MCP round-trip; the same deny-list (`api_key`, `providers`, `dlp`) and whitelist (`weights.`, `thresholds.`) apply.

Files changed

• `src/routing/classify/autotune.rs` — new module with `TuneSuggestion`, `AutotunePatch`, `current_snapshot()` + 4 unit tests.
• `src/routing/classify/mod.rs` — registers the autotune module.
• `src/server/mcp_handlers.rs` — `handle_autotune` handler, dispatch registration, tool description in `tools/list`; existing `test_inject_builtin_tools_*` tests updated for the new tool count.

Test plan

• `cargo check --lib` clean
• `cargo clippy --lib --tests -- -D warnings` clean
• `cargo test routing::classify::autotune` — 4 passed
• `cargo test server::mcp_handlers::tests::test_inject` — 3 passed
• All prek pre-commit + pre-push hooks green
• CI green (auto-merge enabled below)

Note on scope

The implementation intentionally ships the surface without speculating on the inference strategy. The `proposed` field is wired and forward-compatible with future inference logic. See `docs/how-to/auto-tune-routing.md` for the manual tuning workflow.