feat: add configurable execution token budgets

[qinxuye]

Summary

• Add an execution-scoped token budget shared by nested agents, with checkpoint-aware pause/resume accounting and guards before model and tool calls.
• Add configurable soft-warning percentages and deterministic hard-stop handoff: preserve already-created, registered files and report partial or blocked completion without spending another model call.
• Add administrator defaults and maximums, personal overrides, bilingual settings controls, and a typed policy-resolver extension. Wire the policy into normal tasks, preview, builder chat, and workforce creation.

Behavior and scope

• No token cap is enabled by default. With no effective cap, there is no budget warning or stop, and no budget-only checkpoint write.
• Blank personal settings inherit administrator/environment settings; administrator maximums still apply. The soft-warning percentage defaults to 80% and is configurable.
• Accounting uses provider-reported input plus output tokens, including cached input once. Existing billing and quota behavior remains separate.
• Enforcement happens at call boundaries. In-flight calls can exceed the configured cap; a final answer already produced by a paid call is retained. This is not an exact billing cap or a guarantee of lower token usage.
• Budget termination preserves available registered output links, identifies incomplete work, and does not claim that existing files have passed content validation.
• No documentation files or experimental harness scripts are included.

Validation

• Backend regression suite: 2,932 passed across tests/core/agent, configuration, token accounting, execution-budget web policy, workforce creation, and service-manager initialization tests.
• Frontend regression suite: 49 passed across settings, execution budgets, user preferences, and translations.
• Pre-commit checks passed, including Python lint/format/type checks and frontend lint/TypeScript checks.
• Added regression coverage ensuring unlimited executions do not create an extra budget-only tail checkpoint.

Real-model task checks during development

Ran real saved tasks with the same synthetic business fixture to generate an XLSX workbook and CSV. The enabled runs deliberately used a 25,000-token cap and 30% soft threshold to exercise the mechanism; these are test values, not proposed defaults.

Model	Budget	Model calls	Reported tokens	Outcome
GPT-5.6 Luna	Disabled	2	20,602	XLSX and CSV delivered
GPT-5.6 Luna	25,000 / 30%	3	31,496	XLSX and CSV delivered; final in-flight call exceeded the cap
DeepSeek Flash	Disabled	6	65,178	XLSX and CSV delivered
DeepSeek Flash	25,000 / 30%	3	35,316	Stopped with XLSX available and CSV incomplete; reported partial completion

Enabled runs emitted one warning each; disabled controls emitted none. Verified accounting, call-boundary stopping, downloadable artifacts, CSV values, and XLSX source rows/formulas. Native Office recalculation and visual rendering were not part of these checks. An initial run with a mismatched Python environment was excluded and rerun with the task environment aligned.

These are single-run functional checks, not statistical performance results. They demonstrate enforcement and partial delivery, not a reliable token-saving effect from the soft warning.