fix: raise eval push payload limit to 100 MiB (match prime#818)

[hubert-marek]

Companion to PrimeIntellect-ai/prime#818.

_MAX_SAMPLES_PAYLOAD_BYTES is the per-sample ceiling push_traces (verifiers/v1/utils/platform.py) enforces before POSTing eval samples to /evaluations/{id}/samples; its comment notes it mirrors the Prime Evals client's request ceiling. prime#818 raises that client ceiling 25 → 100 MiB, so this mirror is raised to match.

Why: computer-use environments (cua_world_v1, osworld_v2_v1) embed a full 1920×1080 screenshot per step (~1.5–3.5 MiB each), so a single rollout's sample exceeds 25 MiB and this check raises, skipping the upload — uv run eval --push logs:

• sample 0 is too large to upload (53564890 > 26214400 bytes) — cua_world (~51 MiB)
• sample 0 is too large to upload (78391638 > 26214400 bytes) — osworld_v2 (~75 MiB @ 50 steps)

so those runs get no hosted rollout link. 100 MiB (with prime#818) gives real headroom for screenshot-heavy rollouts without compressing traces.

🤖 Generated with Claude Code

---

Note

Low Risk
Single constant change with no logic edits; slightly larger HTTP bodies on upload, bounded by the same checks as before.

Overview
Raises _MAX_SAMPLES_PAYLOAD_BYTES in platform.py from 25 MiB to 100 MiB so the verifiers --push path stays aligned with the Prime Evals client request ceiling (prime#818).

push_traces still uses this constant for per-sample rejection and batch splitting before POSTing to /evaluations/{id}/samples; only the ceiling changes, so screenshot-heavy computer-use rollouts that were skipped as “too large” can upload instead of failing the check locally.

^{Reviewed by Cursor Bugbot for commit 066420a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Raise eval push payload limit to 100 MiB in _MAX_SAMPLES_PAYLOAD_BYTES

Increases _MAX_SAMPLES_PAYLOAD_BYTES in platform.py from 25 MiB to 100 MiB to match the limit set in prime#818. Any payload size enforcement using this constant now allows up to 100 MiB.

^{Macroscope summarized 066420a.}

[macroscopeapp]

Approvability

Verdict: Approved 066420a

Single-line configuration change increasing a payload size limit constant from 25 MiB to 100 MiB. The change is minimal, self-contained, and has clear intent with no complex runtime implications.

^{You can customize Macroscope's approvability policy. Learn more.}