fix: hoist leading system/developer prefix out of the openai fold space

[ranxianglei]

fix: hoist leading system/developer prefix out of the openai fold space

Symptom (restart regression, omp @ glm-5.3)

• Panel pre-restart: 3 active blocks (383k compressed), sent 146k.
• Panel post-restart: Blocks: none, sent balloons to 546k, all ranges reappear uncompressed.
• /home/dog/.omp/acp-omp.log on restart: prime-fold … blocks=0 + fold-replay rejected reason=fp m00001..m00264 want 1d8b6f39 got 432bb7b2 @0..264 — positions resolve, fingerprint mismatches; 20 s later the first live request replays the same calls successfully (event=replayed).

Root cause

The openai codec was the only one folding the system prompt into the message-id space:

codec	system location
anthropic	top-level system field (out of fold space, issue #64)
responses	instructions + systemParts, never in msgs
openai	messages[0] → fold piece at pos 0 (inside)

Two failure modes follow:

1. Compression eats the system prompt. Dump sweep (~/.omp/acp-omp-dumps/req_76*.json, sid 01a027f3): pre-compress requests carry systemLen=45151 (824–828 msgs); the moment compression covered m00001..m00264 (which includes the system piece) the wire dropped to systemLen=0 on every subsequent request. The model received no system prompt at all after the first compression.
2. Restart replay rejected. primeFold reconstructs system content from host state (runtime injections differ across restarts: 45k live vs ~7.4k reconstructed). Since the system piece sat at pos 0 inside the id space, every span fingerprint covering it diverged, and the guard correctly rejected the replay → "Blocks: none" until the first live request re-derived ids in live space.

Fix

openaiToCore now returns { msgs, systemText }:

• the contiguous leading system/developer prefix is hoisted out of the fold space (joined with \n\n);
• mid-conversation system messages (rare, host-synthetic) stay in the fold space unchanged, keeping the originalRole round-trip;
• conversation ids no longer depend on system content at all.

mirrorOpenaiToCore needs no change: the mirror emits systemText as messages[0], which now folds through the same hoist — mirror and live spaces converge regardless of systemText, which is exactly what restart replay needs.

Re-injection stays with the existing injectOpenaiSystem (callers add [originalSystem, compressPrompt]).

Tests

• new tests/openai-system-hoist.test.ts (6): prefix hoisted + ids identical to system-free body; empty systemText; mid-conversation system kept; system-content changes never shift ids (restart regression); mirror converges with live for arbitrary systemText; re-injection round-trip.
• tests/wire-bili-message-roundtrip.test.ts: the two tests encoding the old contract (leading-only system/developer bodies) rewritten — prefix hoist + mid-conversation originalRole round-trip.

Suite: 428/428, tsc --noEmit clean.

Downstream

• omp: capture systemText at the fold and re-inject the original on rebuild (follow-up PR + e2e restart regression with live system ≠ reconstruction).
• proxy: server.ts openai branch currently only injects the compress prompt via injectOpenaiSystem; after this change it must prepend the original systemText too (follow-up PR).
• pi: does not use the openai codecs — zero impact.