[asutermo]

Summary

Fix the mainline context / eval-author work-budget failure without reducing skill content, skipping comparisons, or relaxing similarity thresholds.

• Raise both revision-guarded SkillEvaluator v0.7.0 scalar-work limits from 100m to 300m, recorded as full-pairwise-300m-v1.
• Preserve the 512-chunk ceiling, vector validation, exact pairwise algorithm, and five-minute timeout. Exceeding the work/time budget remains incomplete, not passing.
• Retain only check-specific, bounded integer guard metadata in diagnostics. Do not estimate missing dimensions or publish raw provider/source metadata.

Incident and budget rationale

Run https://github.com/NVIDIA-NeMo/eval-author/actions/runs/41132259985 (job https://github.com/NVIDIA-NeMo/eval-author/actions/runs/41132259985/job/124337600119) evaluated 69 files / 417 chunks and hit scalar_comparison_limit (100000000) after about four seconds. PR #166 was already merged and its dual-guard version fix was deployed; this is real exhaustion of the existing allowance, not the former 25m clustering override or a timeout.

The recorded model route is nvidia/nvidia/nemotron-3-embed-1b. Its documented output is 2048 dimensions, but the actual incident response dimension was not retained; this PR does not claim it was observed. At 2048 dimensions, 417 chunks require 177,635,328 scalar work, and the existing maximum 512 chunks requires 267,911,168. A 300m bound accommodates that maximum while retaining a finite guard.

Verification

• Tests-first regressions reproduced the prior work-budget rejection and missing safe diagnostics.
• Installed pinned v0.7.0: 136 focused/native tests passed, zero skips.
• Network-disabled Linux ARM64 Docker: 25 native tests passed. The actual wrapper/CLI and downstream collector parser accepted synthetic 417×2048 and 512×2048 probes; 512×2294 (300,091,904) remained incomplete. Native boundary tests admit 512×2293 (299,961,088).
• Excess chunks and invalid provider vectors remain rejected. Synthetic vectors are deliberately labeled; these results do not establish real-content semantic findings or live-provider completion.
• Ruff, formatting, Python 3.12/3.13 typing, copyright, license and whitespace checks passed. Independent review passed with no security or logic findings.
• Final stable-tree full suite: 3,427 passed, 25 skipped, 289 warnings. The isolated-runtime evaluator tests skipped in this general environment passed separately against the actual pinned evaluator and in Linux Docker.

Scope and limitations

Five files only: wrapper, diagnostics, their tests, and existing CI documentation. No skill content, dependencies, lockfile, credentials, provider configuration, or workflow trigger changes. Separate from the anti-cheat work in #164.

Linux ARM64 Docker is not GitHub's AMD64 runner. No provider requests were made; provider latency, actual incident dimensions, and downstream LLM completion remain unverified. The next mainline live run is the end-to-end confirmation, and incomplete handling remains intact.