fix(ai-red-teaming): make extraction/membership/inversion work in the TUI

[rdheekonda]

Problem

Running the traditional-ML privacy attacks through the TUI agent did not match the SDK:

• Extraction ran with an empty query pool -> 0 target queries, producing a bogus "0% fidelity, cloned" finding. The SDK path worked because the notebook explicitly fetches /pool and passes it in.
• Membership inference had the same latent gap (needs member/non-member sets).
• Model inversion had no agent tool at all - the dispatch map only had extraction/membership/evasion, so the agent could not run it (SDK-only).
• Even when extraction ran, the agent's summary showed only fidelity/agreement/queries and reported per-class fidelity, the budget curve, and transfer as "missing follow-ups" - though the SDK computes them and they are persisted to the finding's metrics_detail.

Evasion was unaffected (it perturbs a single input, no external dataset).

Fix (capability-only)

• Extraction (generate_extraction_attack): derive the query pool from the target's sibling /pool endpoint when none is given; fail loudly on an empty pool instead of a 0-query clone; surface the full SDK result to the agent (soft fidelity, KL divergence, per-class fidelity, fidelity-vs-budget curve, surrogate accuracy); actually run the transfer test (measure_transfer defaults on, exposed as a tool param).
• Membership (generate_membership_attack): derive member/non-member sets from /members and /nonmembers when not supplied; fail loudly if either is empty.
• Inversion (generate_inversion_attack): new agent tool (confidence / nes). Infers input_dim/input_shape from a /pool sample, reconstructs per class, reports per-class confidence + classes recovered.
• Agent doc lists the four black-box classifier tools.

Bumps capability 1.12.0 -> 1.13.0 (immutable content requires a bump).

Tests

14 new unit tests (tests/test_attack_runner.py): pool/member-set derivation, fail-loud guards, rich-metric surfacing, transfer wiring, and the new inversion tool (dispatch registration, shape inference, per-class output, unknown-attack error).

Validation

• pytest tests/test_attack_runner.py -> 152 passed, 1 pre-existing failure (TestGenerateMultimodalAttack::test_generates_compiling_script_with_image) confirmed failing on origin/main, unrelated to this change.
• Pending: E2E via TUI against the ml-extraction-* targets to confirm extraction now queries the pool (non-zero fidelity), membership derives its sets, and inversion runs.