fix: retain fitting OpenAI read baselines

[jalehman]

What

Keep the newest successful singleton read exact when the complete OpenAI provider request is proven to fit its token window.

Why

The Unicode-safe byte fallback projected a 61.9 KB canonical page even though the full Terra request used only about 64k message tokens. That prevented staged ingestion proposals from preserving required update baselines.

Changes

• Preserve the byte-safe fallback first
• Restore only the newest singleton read
• Count the provider-transformed OpenAI request
• Reject dense Unicode and mutation outputs

Testing

• bun test test/ai/gateway-tool-loop-openai-budget.test.ts test/ai/tool-loop-context-exact-results.test.ts test/ai/gateway-tool-loop.test.ts
• bun run typecheck
• Autoreview clean

[jalehman]

Merge gate: exact production transcript replay retained the 61,900-byte latest read at 107,045 estimated total tokens; focused tests 33/33, typecheck, repository verify, static/security checks, mechanical tier, LLM skill tier, JSONB parity, and specialized slow tests passed. Shard 7/9/10 failures exactly reproduce on the prior head and are unrelated database/extraction failures. Serial failed on an unrelated temporary Git index.lock race. Proceeding under the explicitly authorized ingestion repair/deploy.