Add Auto-FL report skill

[holgerroth]

Summary

Adds a productized nvflare-autofl-report companion skill for generating reproducible final artifacts after an Auto-FL campaign has stopped, reached a cap, hit a hard blocker, or been manually interrupted.

This is a follow-up to #4780 and is intentionally stacked on that branch. The new work is contained in commit 08105a368; once #4780 lands, this PR's diff will collapse to the reporting feature only.

User Experience

After stopping a campaign, the user can ask their coding agent:

Use the NVFlare Auto-FL Report skill.
Generate the final report for the stopped campaign in ./job.

The skill verifies stopped state and deterministically produces:

• autofl_final_report.md
• autofl_report_summary.json
• a refreshed progress.png

The report includes baseline/best results, candidate lineage and inherited code changes, manifests and hashes, exact commands, runtime/failures, literature checkpoints with measured follow-on outcomes, and comparability warnings.

Design

• Refuses to finalize state with final_response_allowed=false unless the user explicitly confirms an abrupt interruption after execution is independently checked.
• Does not mutate job source, results.tsv, candidate manifests, or campaign state.
• Works without Git and does not auto-commit.
• Reuses the product Auto-FL progress plotter instead of copying research plotting logic.
• Distinguishes the imported autofl.yaml budget from executed baseline/best commands.
• Warns when training compute changed or a test-like metric guided repeated candidate selection.
• Associates each literature checkpoint with subsequent measured candidates until the next checkpoint and preserves campaign-recorded source identifiers without claiming independent citation verification.

No files under research/auto-fl-research and no H100-specific assets or instructions are changed.

Validation

• 72 passed across report, Auto-FL runner/guard/plotter, importer, skill admission, and release-bundle tests.
• Black, isort, flake8, and git diff --check pass.
• Full docs HTML build completes; remaining warnings are pre-existing elsewhere in the docs tree.
• Forward-tested against a copied 149-row, 8-client CIFAR-10 campaign ledger: the report reproduced the 0.6870 baseline, 0.8218 best score, 34-candidate retained lineage, 10 literature checkpoints, and the executed local-epoch comparability warning without touching the live campaign.

Dependency

• Depends on Add Auto-FL agent skill #4780 (nvflare-autofl, candidate manifests, campaign state, ledger, and progress plotter).

[codecov-commenter]

Codecov Report

❌ Patch coverage is 87.20682% with 60 lines in your changes missing coverage. Please review.
✅ Project coverage is 56.68%. Comparing base (7df6a4c) to head (23f6a6e).

Files with missing lines	Patch %	Lines
nvflare/app_common/autofl/job_importer.py	87.04%	60 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4845      +/-   ##
==========================================
+ Coverage   56.49%   56.68%   +0.18%     
==========================================
  Files         969      971       +2     
  Lines       92210    92679     +469     
==========================================
+ Hits        52096    52535     +439     
- Misses      40114    40144      +30     

Flag	Coverage Δ
unit-tests	56.68% <87.20%> (+0.18%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[holgerroth]

Addressed the external hardening findings in 424f898. Baseline and retained-best selection are now strictly status-based, while name/command baseline heuristics remain limited to guard-compatible attempt accounting. Unreadable candidate manifests fail closed; state-only and ledger-referenced pending evidence are covered. Zero-runtime prose renders n/a, relative --plotter semantics are documented, and pending literature rendering dead code was removed. Added regressions for baseline_plus_prox, corrupt/list/empty manifests, state-only pending evidence, ledger-referenced manifests, divergent retained-vs-observed warnings, plot preservation before refusal, and zero runtime. Validation: 79 Auto-FL tests passed; Black, isort, flake8, git diff, and docs build passed.

[holgerroth]

/build