fix: restore agreed token budget wording

[Whamp]

Summary

Follow-up correction to #127.

#127 compressed the reviewed tokenBudget schema wording to remain under the existing provider-visible byte ceiling. That was the wrong tradeoff: the agreed language was the requirement, so this PR restores it verbatim and deliberately revises the ceiling instead.

Optional user-requested soft spend gate, not a planning target. Do not set tokenBudget unless the user explicitly supplies a cap or asks you to choose one; never infer or invent one from task size. If omitted, the configured defaultTokenBudget applies; without one, the run is unlimited. Reaching the gate blocks later agent() calls; concurrent in-flight work can overshoot.

• Pins the full sentence with a strict equality regression test.
• Revises the accepted provider-visible definition ceiling from 4,283 to 4,392 bytes.
• Regenerates the context measurement and guidance baseline.

Evidence

• npx tsx --test tests/workflow-tool.test.ts tests/workflow-prompt-budget.test.ts — 35/35
• npm run build
• npm run docs:check
• npm run context:check
• npm run guidance:check
• npm run release:verify — 0 warnings
• npm run delivery-choice -- --model openai-codex/gpt-5.6-sol:high — 3/3 scenarios passed against the exact wording

[QuintinShaw]

Thanks — restoring the exact wording the eval validated is the right call; shipping a paraphrase the harness never saw would quietly invalidate the 3/3 result. I pushed a small commit fixing the comment in the budget test (it credited #127 with the verbatim adoption — that's this PR; #127 was the compression) — hope you don't mind. Merging.