tests: harden Android signing/permission/approval path coverage

[kwsantiago]

Standalone backlog (split out of #319, closed as interop-complete). This is a refinement / tech-debt item, not a client-interop gap.

Problem

The Rust core is now heavily unit-tested, but the Android (Kotlin) signer layer is thin — ~2 unit + ~7 instrumented tests across ~100 files. For a key-holding app, the request-handling / permission-resolution / approval / rate-limit wiring on the Android side is under-tested (happy-path dominant).

Proposed work

Add focused tests on the security-relevant Android paths:

• Nip55ContentProvider / Nip55Activity request parsing + result mapping (incl. error/reject/timeout paths).
• Permission decision + duration mapping round-trips (Kotlin <-> Rust enum mapping).
• Approval flow state (biometric failure, cancellation, kill-switch active).
• Rate-limit / velocity wiring (the delegation to the Rust limiters).
• Bunker request gating (handleApprovalRequest authz + rate-limit + concurrency cap).

Batch / multi-event signing (added by #320 / PR #325)

The batch path is brand-new and security-sensitive; it currently has no instrumented coverage:

• TOCTOU consent-bypass regression test (highest priority): assert the signed/rejected set equals the set rendered on the approval screen. Specifically: accumulate N requests, capture the displayed snapshot, then deliver a late onNewIntent and confirm (a) once a decision is committed (decisionLocked) further intents are ignored, and (b) handleApproveBatch/handleRejectBatch act on the displayed snapshot, never the live pending list. This guards the HIGH finding fixed in 9a9bf6c.
• Same-caller accumulation invariant: a different-caller or get_public_key intent resets the batch; cross-caller mixing never happens.
• MAX_BATCH_SIZE (20) cap: the 21st request is dropped, not accumulated.
• One-biometric-for-N semantics: a single presence gate covers the batch; per-request preApprove + grant + audit still fire for every event (parity with the single path).
• Result serialization round-trip: serializeBatchResults output parses back to the expected {id, package, signature, result, rejected?} per request, including the batch-reject (rejected: true) case.

Acceptance

• Meaningful coverage of the unhappy paths that gate key use, not just happy-path.
• A regression test that fails if the batch signed set can diverge from the displayed set.