fix: preserve requested output budgets across native compression

[WindLX]

Fixes #69

Proxy-side issue: ranxianglei/billion-context#1678

Summary

Preserve the pre-clamp requested ceiling on the OpenAI transport using x-bili-output-budget: <requested>:<raw-clamped>. Keep max_tokens's existing raw-input clamp for direct requests. A coordinated billion-context change consumes this metadata before rebuilding/compression, recalculates the reservation against rebuilt input, and removes the header before upstream forwarding.

The header is sent even when the two values are equal, allowing the proxy to preserve deliberately small user budgets rather than apply its historical starvation heuristic. CLI transport remains unchanged because /alpha/generate is not a supported compression route. No extra retry, effort change, catalog/window inflation, config option, or history migration.

Compatibility and review

This is a cross-plugin protocol proposal, not a standalone fix. Older proxies/direct routes retain the clamped JSON budget. Both implementations must ship for the affected native DSH integration to benefit. Header naming and handshake design need maintainer agreement; this PR is intentionally a draft. The plugin version and Harness peer ranges are unchanged.

Verification

• npm run typecheck: pass.
• npm test: 714 passed.
• npm run build: pass; generated lib/index.js and source map included according to the repository's existing tracked artifact surface.
• npm run test:engine: passed against DSH 0.2.0-rc.1 on the same adapter change before PR staging.
• Regression fails without the new header, passes with it; both transports retain the raw clamp, and an explicit 1024 remains identifiable as explicit.
• Loopback-only adapter -> native interceptor -> proxy -> fake upstream: raw 1024 becomes upstream 131072; explicit 128 remains 128; handoff is not leaked upstream.
• Locally deployed coordinated pair allowed the affected user to continue their real session. No authenticated live-provider replay of private historical content was performed.

Only synthetic fixtures and source changes are included; no user conversation, credentials, local config or traces are published.