feat(phase-12): add token cost controls

[lyxnbclass]

What changed

• add provider-independent session token budgets and per-request output limits
• estimate message and tool-schema input cost before every model step
• stop with budget_exhausted before contacting the provider when the request cannot fit
• preserve reported token usage across resumed sessions and expose /budget
• document configuration, threat model, validation, and known limits

Why

Long-running tool loops can continue spending model tokens even when the user intended a bounded task. This stage adds an explicit opt-in guard without hard-coding volatile provider prices or claiming exact billing.

User impact

Users can set --token-budget / MINICODE_SESSION_TOKEN_BUDGET and --max-output-tokens / MINICODE_MAX_OUTPUT_TOKENS. Existing behavior is unchanged when neither limit is configured.

Validation

• 107 passed stage-related regression
• 298 passed, 3 skipped full regression
• branch coverage 85.58%
• Ruff and Mypy passed
• compileall, sdist/wheel build, and offline MockModel demo passed