docs: correct v0.6.14 release contracts

[jdchawla29]

Summary

• disclose the trace payload categories and separate backend trace, local span, file-tracking, and CLI analytics behavior
• align runtime resources, rollout deadlines, verifier submission, and Docker socket warnings with the implemented SDK contracts
• document experimental stability, the complete Harbor adaptation result, and serialized live-environment acquisitions
• keep the Fireworks material focused on one serverless importance-sampling workflow

Why

The release documentation stated stronger privacy, isolation, and timing guarantees than the SDK implements and omitted parts of two public surfaces. These changes make the docs describe the shipped code paths and security boundaries directly.

User impact

Users can make informed telemetry and Docker-socket decisions, understand timeout return behavior and runtime resource handling, consume structured Harbor failures, and choose the supported Fireworks cookbook path without release-history or comparison material.

Validation

• jq empty docs/docs.json
• validated all 36 v6 MDX pages for navigation coverage, internal paths, anchors, and component balance
• uv run --extra dev --extra modal pytest -q hud/tests/test_settings.py hud/telemetry/tests/test_exporter.py hud/telemetry/tests/test_instrument.py hud/eval/tests/test_file_tracking_observer.py hud/eval/tests/test_local_runtime.py hud/eval/tests/test_hosted.py hud/eval/tests/test_rollout.py hud/integrations/harbor/tests/test_contract.py hud/eval/tests/test_docker_provider.py (268 passed)
• git diff --check origin/main..HEAD

---

Note

Low Risk
Documentation-only changes that correct overstated guarantees and disclose telemetry and Docker-socket risks; no runtime or API behavior changes in this PR.

Overview
Aligns v6 documentation with shipped SDK behavior for telemetry, runtimes, verifiers, Harbor, experimental APIs, and the Fireworks cookbook—without changing code.

Telemetry now lists what trace spans can contain (prompts, model output, tool I/O, instrumented call args/results, workspace artifacts) and separates backend uploads, local span files, file-tracking spans, and CLI analytics controls.

Runtime and tasks clarify that rollout_timeout caps execution through grading but not how long Task.run / Taskset.run may block (post-expiry cancellation, trace reporting, telemetry flush). HUDRuntime resource handling and LocalRuntime serialized acquisitions on a shared live Environment are documented.

Experimental pages add stability notices; Harbor documents AdaptResult / ValueError when no tasks exist; verifier and Compose docs describe hosted verifier serialization and a Warning that local service_access mounts the host Docker daemon.

Fireworks material is trimmed to the one-turn importance-sampling serverless path; loss-function comparison, BFCL, and dedicated-vs-serverless sections are removed. The training guide points to that narrower cookbook scope.

^{Reviewed by Cursor Bugbot for commit b64a6f6. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b64a6f6121

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".