feat(security): collect credentials via elicitation for createAccessibilityAuthConfig and createLCASteps

[SavioBS629]

Both tools previously accepted credentials as tool arguments, so the values transited the model context. They now prefer MCP elicitation: when the connected client advertises elicitation support, missing credentials are requested directly from the user via the client — the value flows user -> client -> server and never enters the LLM's tool-call arguments, context, or logs.

A shared, capability-gated helper (elicitCredentialsIfSupported) drives this. When the client can't elicit, the user declines, or the request errors, it falls back to the existing argument-based flow — so behavior is unchanged for clients without elicitation and it is safe to ship to remote MCP (degrades gracefully).

• createAccessibilityAuthConfig: username/password are now optional; elicited when absent and supported, otherwise the tool errors asking for them.
• createLCASteps: adds a requires_authentication flag so credentials are only elicited when the test case actually needs login; the credentials arg remains a fallback.

Backward-compatible (no argument removed). Tests cover the helper's branches and the elicit / no-capability / decline / error paths for both tools.

[SavioBS629]

Elicitation for credential tools — behavior & rationale

createAccessibilityAuthConfig

┌────────────────────────────────────┬───────────────────────────────────────┐
│             Condition              │                Result                 │
├────────────────────────────────────┼───────────────────────────────────────┤
│ Credentials passed as args         │ Used directly — no prompt             │
│                                    │ (backward-compat)                     │
├────────────────────────────────────┼───────────────────────────────────────┤
│ Credentials omitted + client       │ Server prompts the user via the       │
│ supports elicitation               │ client form; entered values are used  │
├────────────────────────────────────┼───────────────────────────────────────┤
│ Credentials omitted + client does  │ Returns error "Username and password  │
│ not support elicitation            │ are required…" (graceful fallback)    │
├────────────────────────────────────┼───────────────────────────────────────┤
│ Credentials omitted + client       │ Returns the same "required" error     │
│ supports it, but user              │ (creds still missing)                 │
│ declines/cancels                   │                                       │
└────────────────────────────────────┴───────────────────────────────────────┘

createLCASteps

┌───────────────────────────────────────┬────────────────────────────────────┐
│               Condition               │               Result               │
├───────────────────────────────────────┼────────────────────────────────────┤
│ requires_authentication false/unset   │ No prompt — most test cases don't  │
│                                       │ need login; proceeds normally      │
├───────────────────────────────────────┼────────────────────────────────────┤
│ Credentials passed as args            │ Used directly — no prompt          │
│                                       │ (backward-compat)                  │
├───────────────────────────────────────┼────────────────────────────────────┤
│ requires_authentication: true + creds │ Server prompts for the test case's │
│  omitted + client supports            │  login; entered values are used    │
│ elicitation                           │                                    │
├───────────────────────────────────────┼────────────────────────────────────┤
│ requires_authentication: true + creds │ Falls back to the existing arg     │
│  omitted + client does not support    │ path (proceeds without creds,      │
│ elicitation                           │ unchanged behavior — no error)     │
└───────────────────────────────────────┴────────────────────────────────────┘

Why we implemented it this way (arg path kept as fallback)

1. Backward compatible — no argument removed; existing callers that pass credentials work unchanged. Zero breaking change.
2. Uneven client support — elicitation is newer; not all clients (nor the Remote MCP wrapper yet) support it. Capability-gating means we prompt only where supported and degrade gracefully elsewhere.
3. Safe for Remote MCP — gated on client capability, so deploying can't break the hosted server; elicitation just stays dormant there until the wrapper forwards it.
4. Incremental, low-risk — ships the security win now (creds out of model context on capable clients) without forcing a breaking change; we watch for issues before tightening.

Note — new requires_authentication flag (LCA only): Unlike the accessibility tool (creds always required), most LCA test cases don't need a login, so there's no reliable way to know when to prompt. The flag is an explicit "this test case needs login" signal — the agent sets it instead of passing credentials, and the tool then elicits them. Additive and optional (defaults false), so it changes nothing for existing calls.