Add automatic Graph selection with shared Cua-S1 worker budgets

[Levius-Fubuki]

Purpose

Add opt-in automatic selection among eager, exact Graph and rule-bucket execution, sharing one request clock, capture ledger, cache and memory budget. The latest follow-up integrates the graph lifetime fixes from the preceding PRs. Eager remains the default; automatic selection does not guarantee the fastest mode.

Stacked on #37. Incremental runtime comparison. The PR diff remains limited to runtime Python source; validation helpers, tests and generated evidence are preserved separately. This optimizes the Python multimodal worker, not the native Rust/CUDA worker.

Test Plan

System1-Omni Version / Commit: 10b1522e483999e1830f715bb9970128cfaf323f.

Fresh validation started 2026-09-30 on RTX 4090 24 GiB, driver 595.71.05, Torch 2.14.0+cu130, Transformers 5.17.0 and PEFT 0.21.0; BF16 base with unmerged adapter, without FLA/causal-conv1d.

• Validation checkout and pinned setup: PYTHONPATH=src:recipe/cua_s1 python -m pytest tests/cua_s1 -q.
• Exact GPU commands and exit statuses, using upstream-verified weights and a fresh output directory.
• Runtime/validation source mapping: git diff <runtime> <validation> -- src is empty. Source hashes and numerical/performance reports are independently verified.

Test Result

• CPU suite: 264 passed in 21.68s.
• Fresh full-checkpoint automatic-worker validation passes 266 full-logit comparisons and 61 complete responses, including changed inputs and boundaries. Shared-manager eviction, capture/time/resident budgets, explicit retirement and real HTTP/model-isolation checks pass.
• A dedicated two-entry cross-mode test evicts a bucket while retaining an exact graph, then verifies three changed-content full-logit survivor replays and released victim resources.
• 3,720 timed predictions cover two cold-cache runs across five synthetic workloads and five modes: eager, default exact, exact admission-window 32, bucket and auto. Complete responses, source hashes, timing summaries and shared resource counters are independently verified.
• Auto total time is 14.8–30.9% below eager in this matrix. Relative to the best fixed mode for each run, auto’s extra time ranges from +0.23% to +9.62%. Positive values mean auto is slower. This does not establish a universal advantage over explicit modes; keep it opt-in and select modes using the intended workload.
• GitHub rust CI passed on the above runtime head.

Full evidence and tradeoffs · Raw primary report · Independent verification summary.

Timing is serial and synthetic on one GPU. Cold-cache totals include capture and parity-check costs; model loading, HTTP parsing and final cleanup are excluded. Two runs are not confidence intervals or proof of production workload behavior.

Self-review

Agent-assisted full-diff review and validation limits found no actionable correctness or architecture defect. Runtime scope, commands and claims were checked against the recorded sources/results. The contributor remains responsible for understanding the changes; this does not replace maintainer review.

[hsliuustc0106]

Review of 164a3457f94b (incremental stack changes).

Reviewed exact/bucket selection, shared cache/admission/request accounting, bounded timing/history state and cross-mode eviction. No independent actionable correctness defect found in the inspected paths.

I would hold adoption until representative workload measurements demonstrate that automatic selection improves on explicit exact/bucket modes after including capture and parity-check costs. Also validate cross-mode eviction and surviving-graph replay on real GPUs. The heuristic is not itself evidence that auto chooses the fastest mode. This optimizes the Python multimodal worker, not #19's native Rust/CUDA path.

Validation: 264 archived CPU tests passed against this head's runtime source with Transformers 5.17.0 and tokenizers 0.23.2. Local Torch remains 2.13 rather than pinned 2.14. CUDA capture/stream/event behavior was mocked; no full-checkpoint GPU parity or new latency measurement was performed.

The validation suites were retrieved from the immutable archived revisions linked in the PR descriptions; they are not retained in the current PR diffs.