Publish evidence-guided alpha workflow

[Evostructs]

What

• Publish the evidence-guided alpha workflow, public examples, licensing, security policy, and CI hardening.
• Add behavior replay, inspection, verification, optimization, and the independent Codex CLI adapter.
• Add an explicit OpenAI-compatible thinking-mode control for short exact-label replays.
• Publish reproducible Codex and official DeepSeek API reports with hash-bound audits.

Why

Instruction compression needs evidence about preserved behavior, not token reduction alone. This change adds frozen candidates, asset-specific replay suites, negative controls, provenance, and explicit claim limits.

The DeepSeek follow-up also exposed an adapter boundary: provider reasoning can consume a replay's 128-token output allowance before final content is emitted. The new option lets users explicitly disable compatible-provider reasoning for these short tasks; the default remains provider-controlled.

DeepSeek V4 follow-up

• Official API models: deepseek-v4-pro and deepseek-v4-flash
• Three trials, seed 20260817, explicit non-thinking mode
• 534 calls, zero operational errors
• V4 Pro matched source and candidate on all valid blind observations; V4 Flash matched both on 63/66 and showed the same permission-boundary errors on each asset
• Neutral mutants degraded under both models, showing material but incomplete control sensitivity
• Raw reports, hashes, invalid-case handling, privacy scan, and limits are recorded in examples/project_instructions/deepseek-v4-replay-audit.2026-08-17.json

Safety and scope

• No credentials, bearer values, local user paths, Codex paths, or handoff files are included.
• The clean source distribution contains all 10 published DeepSeek reports and excludes local failed diagnostics.
• This PR makes no general model-ranking or behavior-equivalence claim.

Checks

• ruff format --check .
• ruff check .
• mypy denser
• pytest -q — 170 passed, 2 deselected
• Clean sdist and wheel build
• Fresh wheel install and CLI smoke test
• Package-content and sensitive-shape scans