fix(bench): grid compressor — non-thinking model, degenerate-output guard, prompts in artifact

[drewstone]

Grid run #1's integrity flags, fixed: the thinking worker model burned the compressor's 1024-token cap on reasoning and emitted a ~1-word prompt — silently substituting a different treatment (prompt REMOVAL) for the requested 50% compression. The cells still PROMOTED (the domain was saturated, so the score leg was vacuous), which is exactly why this had to fail loud.

• Compressor = a fixed non-thinking model (COMPRESSOR_MODEL, default deepseek-v4-flash), maxTokens 2048.
• Degenerate output fails loud: <20% of the target word count throws with the output in the message — a wrong treatment can never silently enter a cell.
• The artifact now persists the actual per-cell prompts, so the treatment is inspectable post-hoc (run fix: persist final runtime stream failures #1's couldn't be).

Run #1's honest residue (recorded in the ledger): machinery validated end-to-end with the program's first PROMOTED verdicts; the science needs a non-saturated domain — rerun queued with a weaker worker. The steering cost-mechanics observation (refine early-stops on resolve → half sample's cost at equal score) stands.

[tangletools]

✅ Auto-approved PR — 146c7ed1

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-10T22:07:56Z}