Stop sending contextWindow as the request max_tokens

[TheGreatAxios]

Summary

• Setting contextWindow no longer changes the request output budget; every provider branch keeps the shared source default
• contextWindow still drives compaction and the status-bar meter via contextWindowFor, and the provider-branch behavior is now documented in IMPLEMENTATION.md
• Defers a dedicated maxOutputTokens setting to a follow-up; no new setting is added here
• No in-tree default or operator config sets contextWindow (only tests reference it), so no existing deployment changes behavior

Verification

• bun test src/config/inference-sources.test.ts src/provider/context-window.test.ts src/provider/openai-compatible-adapter.test.ts src/pricing-metadata.test.ts src/cost/cost-summary.test.ts tests/unit/context-window.test.ts passes (73 pass, 0 fail)
• bun test src/config.test.ts passes (94 pass, 0 fail)
• bun run check passes (lint, typecheck, build, 6982 pass, 0 fail)
• New regression test proves contextWindow 400000 yields max_tokens 16384, not 400000, while contextWindowFor still returns 400000

Fixes CL-7784

[linear-code]

CL-7784