feat: audit Codex context savings

[Evostructs]

What

• add denser audit, a fail-closed behavior-fidelity gate with a required known-bad control
• add a versioned Codex CLI text-only/v1 profile for pre-bundled text decisions
• publish a reproducible, seeded standard-vs-text-only benchmark and complete per-call evidence
• package the work as 0.2.0a1 with an updated citation and case study

Why

A shorter instruction file is not the same as lower end-to-end model input. This change measures provider-reported full input and refuses to call unchanged scores convincing unless the workload also catches a known-bad change.

Recorded result

Codex CLI 0.147.0, gpt-5.6-sol, medium reasoning, three trials per case, randomized submission order:

• release-operation decisions: 27/27 in each profile; 20,294.11 -> 18,154.00 input tokens/call (-10.55%)
• automation permission routing: 15/15 in each profile; 20,619.00 -> 18,434.00 input tokens/call (-10.60%)
• 84/84 calls passed; 0 operational errors; 0 transport fallbacks

The first strict run exposed one missing output-contract behavior. text-only/v1 fixes that boundary, and the complete 84-call audit was rerun.

Checks

• 181 tests passed locally (2 integration tests deselected)
• Ruff formatting and lint passed
• mypy passed
• wheel and source archive built successfully
• wheel installed in a fresh Python 3.12 environment; CLI reported 0.2.0a1
• staged-content sensitive-pattern scan found no credentials or local user paths

Scope

text-only is only for tasks whose complete input is already in the request. Coding, repository, shell, network, plugin, app, skill, or memory work must keep the standard profile.