fix(tau2): zero-fill missing/empty sessions in aggregate score

[elronbandel]

Summary

• Closes #223 (Layer 2 — the tau2-specific hidden denominator).
• Tau2's aggregate_sessions was silently dropping sessions without a simulation file, while total_tasks kept reporting the planned count. Headline benchmark_score was averaged over a smaller denominator than every other field implied (e.g. Kimi+claude_code+tau2/retail returned 1.0 over 3/100 sessions).

What changed

• Iterate over sessions directly in TAU2Evaluator.aggregate_sessions, counting aggregated_sessions as the actual contributor count.
• Rescale avg_reward from compute_metrics to the planned denominator (total_sessions) — missing/empty sessions contribute reward=0, matching the per-session score() contract introduced in #152.
• Expose aggregated_sessions in the metrics dict and logger.warning(...) when it diverges from total_tasks, so downstream readers can detect partial coverage without grepping logs.

Why this shape

Every other benchmark adapter (swebench, bfcl, browsecompplus, gsm8k, hotpotqa, appworld) raises on a missing planned-session file — tau2 was the only fail-quiet aggregator after #152. #152 correctly made the per-session score() return 0 for "agent returned text instead of tool calls", but in the aggregator it switched from a hard error to a silent skip — those two paths disagreed. This PR makes them agree without re-introducing the original #152 crash.

Layer 1 (core_aggregate pre-filtering to COMPLETED) is left for a follow-up — once tau2 is honest, the residual gap is just non-COMPLETED sessions, which successful_sessions / total_sessions already exposes uniformly.

Test plan

• uv run pytest tests/benchmarks/test_tau2_aggregate.py -v — 5/5 pass, including a new test_missing_file_counts_as_zero_reward and the updated test_empty_sessions_count_as_zero_reward (now asserting score=0.4 over 2 sessions instead of 0.8 over 1).
• uv run ruff check and ruff format --check clean.
• Re-aggregate an existing affected run (exgentic evaluate --aggregate-only ...) and confirm benchmark_score drops from the inflated value to the planned-denominator value.