fix: make workflow token budgets opt-in

[Whamp]

Summary

• Make tokenBudget an opt-in soft spend gate and tell models never to infer it as a planning target.
• Align the workflow-authoring skill, lifecycle guidance, phased-budget example, and HIGH/ULTRA effort directives with that policy.
• Extend the Pi behavior is weird when workflow execution is triggered via the editor component keyword. #89 choice harness to verify omitted ordinary budgets and exact explicit caps across both script and named workflow invocations.
• Preserve the existing provider-visible definition ceiling: 4,276 bytes against the accepted 4,283-byte limit.

Addresses #89.

Evidence

• npm run build
• npm run docs:check
• npm run context:check
• npm run guidance:check
• npm run release:verify — 0 warnings
• Focused changed-area tests — 160/160 before final review fixes; affected tests rerun after each fix
• npm run delivery-choice -- --model openai-codex/gpt-5.6-sol:high — 3/3 scenarios passed
• npm test — 1,149/1,150 locally. The only failure is the unchanged faux-provider object-schema test at tests/agent.test.ts:335, which also fails when run alone and is outside this diff.

[QuintinShaw]

Thanks — this is exactly the right fix for the #89 feedback, and fast. Making tokenBudget a pure opt-in gate and telling models it's not a planning target hits the root cause rather than papering over symptoms, and sweeping the skill, lifecycle doc, examples, and the HIGH/ULTRA effort directives in one pass keeps everything consistent. I pushed one small commit on top to align the review checklist in references/review.md with the opt-in wording — hope you don't mind. Merging.