chore: sync public mirror from internal

[evalops-mirror]

Summary

• sync the sanitized public tree from evalops/mono
• keep evalops/maestro as a generated public mirror of the private source of truth
• preserve public-owned CI and trusted-publishing workflows from the public checkout
• internal source SHA: 9286ef6efb16e40356a9001fcb4a7563502e07d8
• last generated public sync base: c305093fca8d521d2ef5093f32ab42c8573fcd58
• previewed public-tree drift: 1 file(s) to copy/update and 0 stale file(s) to delete
• public-only commits since last generated sync: 0
• to pause generated syncs while hand-editing this branch, apply the sync-hold label to this PR; the sync workflow skips every push while it is set

Source-of-truth status

Public Mirror Drift Audit

• package: @evalops/deixic-code
• private source: https://github.com/evalops/mono@main (9286ef6efb16)
• public projection: https://github.com/evalops/maestro@main (c305093fca8d)
• files to copy or update: 1
• stale files to delete: 0
• result: drift detected
• invariant: public_projection_has_drift

Sample Changed Paths

• copy/update packages/local-host-rs/src/model_catalog_data.json

Guidance

Let internal main generate and merge the public sync PR before relying on public main.

Drift sample

• copy/update packages/local-host-rs/src/model_catalog_data.json

Public-only commits since last generated sync

• none detected since last generated sync

Validation

• generated by the sync-public-release-mirror workflow in public-tree mode

Test Plan

• generated by the maestro-sync-public-release-mirror workflow in public-tree mode
• public-source-provenance require-internal-pr confirms internal source PR lineage
• generated mirror PRs use the slim public CI path (fmt/clippy/smoke + provenance); full cargo test/evals already ran on internal main for this SHA
• bot review threads on generated mirror PRs are advisory and do not block the next sync force-update

Staged Rollout

• Staging is unnecessary for this generated mirror PR: it does not independently promote user-visible behavior. It mirrors already-reviewed internal source from evalops/mono@9286ef6efb16e40356a9001fcb4a7563502e07d8, including existing hidden/evaluation surfaces, and keeps public package parity behind the established public-source-provenance gate.

---

[evalops-mirror]

Resolved 0 review-bot thread(s) on this generated public mirror PR.
Left 0 human-started thread(s) unresolved.

Review of mirrored code belongs on the mono source PR.
See https://github.com/evalops/mono/issues/9100 for the triage pattern.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 061750e975

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Clamp enlarged output ceilings to the remaining context

When openrouter/nvidia/nemotron-3-ultra-550b-a55b uses the catalog default, this value becomes RequestConfig.max_tokens, but build_config_with_usage only applies clamp_output_to_remaining_context to local models and hosted auto-compaction waits until 85% of the 262,144-token window. The new 182,520 ceiling therefore leaves only 79,624 tokens for the prompt, so longer requests that are still far below the advertised context limit are rejected by OpenRouter for context overflow (or require input truncation); the previous 32,768 value did not create this failure range. The same issue affects the new approximately 943K ceilings for ~z-ai/glm-latest and z-ai/glm-5.3, so catalog-derived hosted budgets should be clamped against the actual request input or kept within the compaction headroom.

Useful? React with 👍 / 👎.