feat: add behavior-validated context selection

[Evostructs]

What changed

• add denser minimize-context for versioned visible-context bundles
• greedily test optional components and fail closed on regressions, changed behavior, operational errors, or weak negative controls
• repeat a final audit and require a caller-set provider-reported full-input reduction target
• add bounded concurrent Codex CLI replay with thread-isolated per-call evidence
• publish a reproducible two-task local-file Codex pilot and make it the README's primary result

Why

The earlier text-only result mostly proved that unused capabilities cost tokens. This change tests the more useful claim: keep normal local tools available, remove only irrelevant visible context, and retain components whose removal changes behavior.

Evidence

• real Codex CLI 0.147.0, gpt-5.6-sol, medium reasoning, standard capability profile
• two local-file tasks: complete and selected bundles both 6/6 in the final three-trial audit
• known-bad control: 0/6
• provider input: 277,871 -> 243,210 tokens, 12.47% reduction
• release and CI policy removals were rejected as regressions
• 190 tests passed; Ruff lint/format and strict mypy passed
• wheel and sdist package smoke passed
• key-pattern scan found no OpenAI, GitHub, Google, AWS, Slack, or private-key material