feat(spec-009): orchestration × complexity matrix — full implementation

[zenprocess]

Summary

Operationalizes the orchestration × complexity matrix design inside Pawbench. Inspired by Fabian Wesner's One-Shot Shop Challenge (announcement) — the empirical demonstration that orchestration architecture beats model choice (Team Mode 85% vs Sub-Agents 57% on the same model, 143 E2E tests).

This PR closes Pawbench phases P3 through P7 of the matrix design.

What's new

Modules

• complexity.py — ComplexityTier enum + heuristic inference (B2)
• quality.py — ArtifactQuality + Python analyzer (ruff/mypy/radon, all optional) + generic fallback + pluggable registry (B4)
• dqs.py — Composite Dispatch Quality Score v1 + dqs_spread
• orchestration.py — OrchestrationShape vocabulary + run_with_shape executor + merge-turn synthesis (B1)
• ablation.py — Pure ablation matrix over already-collected results, interpretation thresholds, removal candidates (B7)
• context_tier.py — manifest-only context stripping (B6)

Scenarios

• pawstyle-orchestration-matrix.json (new) — 4 independent feature blocks, one per complexity tier, designed to differentiate orchestration shapes
• All 8 existing scenarios retroactively tagged with complexity_tier

CLI surface

--orchestration flat,waves,scatter-gather,team-mode,subagents
--ablate quality,format_compliance,tool_accuracy,useful_ratio,steering_rate
--context-tier {standard,manifest-only}
--verification-runs N
--no-quality-analysis

Report fields

• dim5_artifact_quality (per-scenario + aggregate)
• quality_by_tier — display / crud / transactional / cross_cutting
• orchestration_results + orchestration_dqs_spread — the headline SLI
• ablation — per-component delta + removal candidates
• dqs — composite + breakdown + verification reliability

Tests

• 53 new tests across 6 test_spec009_* files
• Full regression: 151/151 green (98 existing + 53 new)
• Network-free unit tests via lazy aiohttp/engine imports in orchestration.py

Explicitly out of scope (deferred follow-ups)

• waves shape currently degenerates to subagents; team-mode degenerates to scatter-gather. Real DAG-aware waves and shared-scratchpad team-mode need precise operational contracts before they ship distinct executors.
• artifact_quality is reported but does NOT feed DQS — calibration data required first (≥100 dispatches), then formula change in a separate PR.
• verifier_agreement_rate is structurally surfaced but degenerate at 1.0 until an LLM-judge verifier replaces deterministic score_turn. Plumbing is in place; the judge is a follow-up.
• Leaderboard rendering of new dimensions is a separate PR.

Companion work

• Axiom §17 — normative orchestration shape and complexity tier vocabularies (v0.6.0)
• CACP additive fields — complexity_tier, verification_runs[], artifact_quality, fixture_gap status

Test plan

• pytest tests/test_spec009_* — 53/53 green
• pytest tests/ full regression — 151/151 green
• CLI surface smoke (pawbench --help shows all 5 new flags)

🤖 Generated with Claude Code

[codecov-commenter]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

ℹ️ You can also turn on project coverage checks and project coverage reporting on Pull Request comment

Thanks for integrating Codecov - We've got you covered ☂️

[sonarqubecloud]

Quality Gate failed

Failed conditions
C Reliability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE