Establish EdgeFlow engineering and experiment governance

[hanklin9188]

Summary

This PR converts EdgeFlow's existing design and E00–E30 catalog into an operational engineering and scientific-governance system.

It selectively adapts small, composable practices from mattpocock/skills—shared domain language, focused ADRs, tracer-bullet tickets, public-seam TDD, tight diagnosis loops, primary-source research, architecture deepening, progressive disclosure, and two-axis code review—without allowing a generic agent workflow to override EdgeFlow's G0–G8 validation authority.

What changed

Shared language and decisions

• add AGENTS.md with separate engineering and scientific work streams
• add CONTEXT.md with canonical EdgeFlow terminology
• add six ADRs for local-first control, agent/data-plane separation, raw-record truth, executable validation, PASS-only policy evidence, and deferred learned components
• add docs/agents/ configuration and primary-source research conventions

Ordered execution program

• add a dependency-ordered E01–E30 backlog
• define completion states from planned through released
• define mandatory gates for measurement calibration, primary 3B baselines, fairness/quality, workload policy, causal interventions, amortization, kernel integration, learned components, Copilot grounding, usability, and clean reproduction
• define exact execution, artifacts, acceptance criteria, and prohibited claims for each phase

Validation parity

• add specs/validation_requirements.yaml with 22 stable requirements
• record each requirement's current implemented / partial / planned status honestly
• map rules to enforcement paths, tests, artifacts, owners, backlog tickets, and blocked release profiles
• add scripts/audit_validation_parity.py
• add governance contract tests and a focused GitHub Actions workflow

EdgeFlow-specific skills and templates

• add change-design, ticketing, performance-debugging, and two-axis code-review skills
• add a formal experiment issue form and PR template
• add validation implementation and primary-source research templates
• preserve the existing manifest-tracked experiment report template

Why

The repository already has strong platform engineering, but formal performance evidence remains incomplete. The main risks are now evidence debt, specification/implementation drift, and roadmap work that is not represented as an executable dependency graph.

This PR does not claim new performance results and does not change any existing runtime measurement. It establishes the process required to produce future claims safely.

Validation

Latest head: 44a50aceac9e15546bce6c3f58909e2329f7b98b

• CI / Python 3.11–3.13: passed
• Security job: passed
• Governance Contract: passed
• Ruff: passed
• Full tests: 34 passed, 2 expected GPU-dependent skips per Python job
• Package validation: passed
• Result verification: passed through the existing CI sequence

The standalone overlay package was also checked with:

python scripts/audit_validation_parity.py --schema-only
python -m py_compile scripts/audit_validation_parity.py tests/test_governance_contract.py

Operational backlog created

• Formal evidence program: execute E01–E30 by dependency and evidence level #5 — complete E01–E30 evidence program
• E01: calibrate CUDA, wall-clock, and end-to-end timing boundaries #6 — E01 timer calibration
• E02: establish RTX 4080 SUPER thermal and background-load measurement policy #7 — E02 thermal/background policy
• E03: calibrate repetition counts and winner-reversal risk #8 — E03 repetition calibration
• Validation parity: close the first public-evidence certification gaps #9 — first public-evidence validation parity closure

Current ready frontier after merge

1. E01 timer calibration (E01: calibrate CUDA, wall-clock, and end-to-end timing boundaries #6)
2. E02 thermal/background-load policy (E02: establish RTX 4080 SUPER thermal and background-load measurement policy #7)
3. E03 repetition-count calibration (E03: calibrate repetition counts and winner-reversal risk #8)
4. validation parity closure (Validation parity: close the first public-evidence certification gaps #9)

The primary 3B E04 baseline remains blocked until those foundations close.

Attribution

The workflow is an EdgeFlow-specific adaptation informed by the MIT-licensed mattpocock/skills repository. No upstream code is used as EdgeFlow validation authority; executable EdgeFlow validators remain final.