refactor(optimize-sorting): restructure benchmark flow with catastrophic gate

[colehurwitz]

Changes

• Added gate_catastrophic_t{1,2,3} GateNodes: New first-line gates after run_benchmark that only catch catastrophic accuracy drops (>10% below baseline mean). On HALT, skip confirmation runs and go straight to archive. On PROCEED, continue to confirm_benchmark.

• Moved confirm_benchmark_t{1,2,3} before accuracy gate: Confirmation runs (2 additional benchmarks for 3-run average) now always execute when not catastrophic, rather than only on accuracy pass.

• Added new gate_accuracy_t{1,2,3} GateNodes after confirm_benchmark: These check the averaged 3-run result. Output pass (PROCEED) or fail (HALT) only — no reloop since 3-run average is definitive.

• Updated _accuracy_gate_command(): Handles averaged format by checking if br['accuracy'] is a dict and using br['accuracy']['mean']. Removed catastrophic check (now handled by dedicated gate). Removed reloop output.

• Updated _per_unit_gate_command(): Handles averaged per_unit_accuracy format — checks br['per_unit_accuracy'][uid]['mean'] when value is a dict.

• Updated _archive_prompt(): Describes new flow semantics — catastrophic gate halt produces single-run format, accuracy gate produces averaged format.

• Edge rewiring: New flow per tier:

  • run_benchmark → gate_catastrophic --PROCEED→ confirm_benchmark → gate_accuracy --PROCEED→ archive
  • gate_catastrophic --HALT→ archive (catastrophic failure, skip extra runs)
  • gate_accuracy --HALT→ archive (revert based on averaged result)
  • Tier 3: gate_accuracy --PROCEED→ gate_per_unit_accuracy → archive

• Graph size: 31 nodes, 39 edges (was 28 nodes, 36 edges)

---

Test & prompt fixes (Closes #1432)

• Test node sets: Added gate_catastrophic_t1/t2/t3 to tier node sets; total 28→31, per-tier counts +1 each
• Test edge count: 36→39
• Test edge topology: Updated all assertions to match new flow (run_benchmark → gate_catastrophic → confirm → gate_accuracy → archive)
• Test reloop count: 5→2 (accuracy gate reloops removed; only config gate + per-unit gate remain)
• New test: test_catastrophic_gates_are_gate_nodes
• Builder T1 prompt: Removed stale gate_accuracy_t1 reloop reference
• Builder T2 prompt: Removed stale gate_accuracy_t2 reloop handling section

[github-actions]

Sentrux Quality Report

Absolute

Scanning ....
[scan] git ls-files: 629 total, 615 kept, 14 dropped (ext:14, meta:0, big:0)
[build_project_map] 615 files, 109 unique dirs, 101 cache misses, 5.7ms
[resolve] 1165 resolved, 1572 unresolved (of 2737 total specs)
[resolve_imports] project_map 5.8ms, suffix_idx 1.3ms, suffix_resolve 19.2ms, total 26.3ms
[build_graphs] 615 files | maps 2.8ms, imports 26.6ms, calls+inherit 7.2ms, total 36.5ms | 1164 import, 8401 call, 1 inherit edges
sentrux check — 3 rules checked

Quality: 4450

✗ [Error] max_cc: 6 function(s) exceed max cyclomatic complexity of 30
    factory/cli/_ceo_helpers.py:_execute_ceo (cc=46)
    factory/cli/_ceo_helpers.py:_validate_ceo_flags (cc=43)
    examples/chess_evolve.py:main (cc=33)
    factory/cli/_task_builder.py:_build_ceo_task (cc=32)
    factory/cli/run.py:cmd_run (cc=32)
    factory/outer_loop/mutations.py:validate_and_repair (cc=31)

✗ 1 violation(s) found

Diff (vs base branch)

Scanning ....
[scan] git ls-files: 629 total, 615 kept, 14 dropped (ext:14, meta:0, big:0)
[build_project_map] 615 files, 109 unique dirs, 101 cache misses, 5.9ms
[resolve] 1165 resolved, 1572 unresolved (of 2737 total specs)
[resolve_imports] project_map 6.0ms, suffix_idx 1.3ms, suffix_resolve 19.5ms, total 26.7ms
[build_graphs] 615 files | maps 2.5ms, imports 27.0ms, calls+inherit 6.4ms, total 35.9ms | 1164 import, 8401 call, 1 inherit edges
sentrux gate — structural regression check

Quality:      4447 -> 4450
Coupling:     0.80 → 0.80
Cycles:       4 → 4
God files:    3 → 3

Distance from Main Sequence: 0.35

✗ DEGRADED
  ✗ Complex functions increased: 68 → 69

[colehurwitz]

Superseded by unified PR — see #1443