feat(airt): multimodal prompt_matrix (text×media cross-product) + agent transparency (1.6.10)

[rdheekonda]

Summary

Multimodal attacks could only pair prompts to media 1:1 (prompts/prompts_csv by index/basename). So "I have 4 text prompts and 5 images, run all combinations" produced 5 trials, not the 20 a user expects — a common pattern that wasn't supported. This adds a cross-product mode.

• generate_multimodal_attack gains prompt_matrix=[...]: every prompt runs against every media item → N×M trials (4 prompts × 5 images = 20). Configured transforms apply to each combination. Distinct from prompts/prompts_csv (which remain 1:1).
• The generated workflow builds the media-set list (matrix or 1:1) and logs it: Running 20 multimodal set(s) (matrix: 4 prompts x 5 media) then [1/20] … [20/20].
• Agent transparency (per feedback): the agent doc now (1) insists on calling execute_workflow in the same turn as generate_* — never stop after "workflow generated" (that leaves 0 trials and looks like a silent failure), and (2) instructs the agent to narrate every step (what's running, how many trials, any target error) — no black box.
• Generated-loop variables renamed to descriptive names (no bare i).

Bumps capability to 1.6.10.

Validation

• Offline generate_only: generated workflow compiles, emits PROMPT_MATRIX = [...], uses the cross-product loop (for matrix_prompt in PROMPT_MATRIX), applies transforms=TRANSFORMS per set, and uses no bare i.
• End-to-end on dev: 4 prompts × 5 images → Running 20 multimodal set(s) (matrix: 4 prompts x 5 media), [1/20]…[20/20], Assessment complete (assessment efe16775). 20 trials as expected.
• py_compile clean on attack_runner.py + tools/attacks.py.

Notes

• Builds on feat(airt): Nova Sonic S2S target for TUI + fix judge model on multimodal findings (1.6.9) #81 (multimodal S2S + judge-model fix, 1.6.9).