Keep Codex source and STATE probes on the pinned plaintext protocol

[HTExplicit]

With the production compression setting enabled, borrowing cleared stale Content-Encoding headers and then silently re-entered the business zstd encoder at dispatch. Source acquisition and both target STATE probes therefore differed from the pinned ranxi plaintext wire protocol. The existing header regression passed only because its test process had compression disabled.

Mark only source acquisition and the two-shot target protocol as plaintext observations. Keep normal business and Pelican generation on the existing compression setting; do not toggle that setting. Record source and per-shot request encoding in additive diagnostic fields. The existing authenticated probe comparison can also compare plaintext with the configured codec on one fixed candidate, with identical plaintext body/tier/identity/exit and the same eight-request bound. Probe-only results never qualify business traffic.

Validation: enabling the real compression condition reproduced the unexpected zstd headers before the fix. The corrected source/target dispatch contract, preserved native headers and actual models/tiers, normal-generation compression, unchanged global switch, isolated encoding comparison and ordinary native HTTP tool-continuation regressions pass. Live wire-encoding comparison and complete model acceptance remain pending deployment; prior body-only comparisons do not establish plaintext protocol parity.