feat: add manifest input for paperbanana sweep

[eureka0928]

Closes #182

Summary

paperbanana sweep today takes eight comma-separated axis flags plus --input and --caption — a 200+ character line of bash that's miserable to diff, share, or commit alongside the paper. paperbanana batch already solved this with YAML/JSON manifests; sweep is the symmetric gap.

paperbanana sweep --manifest examples/sweep_manifest.yaml

Example manifest:

input: sample_inputs/transformer_method.txt
caption: "Overview of our encoder-decoder architecture with sparse routing"
max_variants: 20

axes:
  vlm_providers: [gemini, openai]
  image_providers: [google_imagen, openai_imagen]
  refinement_iterations: [2, 3]
  optimize_inputs: [false, true]
  auto_refine: [false]

Implementation

• load_sweep_manifest() in core/sweep.py (~50 lines): parses YAML/JSON, validates required keys (input, caption), enforces types on optional keys (pdf_pages: str, max_variants: int >= 1), rejects unknown axis keys, resolves input paths relative to the manifest's parent directory (same convention as load_batch_manifest).
• --manifest / -m flag on sweep: mutually exclusive with the eight axis flags. When set, --input and --caption become optional and default from the manifest.
• Invocation-level flags stay CLI-only (--output-dir, --config, --format, --dry-run, --verbose, --auto-download-data) because they're invocation concerns, not sweep-plan concerns.
• examples/sweep_manifest.yaml with inline commentary on what belongs in the manifest vs. on the CLI.
• 15 new tests covering YAML + JSON, relative-path resolution, each validation branch, and unknown-axis rejection.

Design notes

• Axis keys in the manifest use plural names (vlm_providers) to match CLI flags (--vlm-providers), not the singular internal axis names in build_sweep_variants (vlm_provider).
• Unknown axis keys are rejected up front so typos like typo_axis don't silently no-op.
• pdf_pages from manifest only overrides when set — the CLI --pdf-pages falls through if the manifest omits it.

Test plan

• pytest tests/test_core/test_sweep.py — 48 tests pass (15 new)
• Full suite: 544 pass, 1 pre-existing unrelated failure deselected, 2 skipped (gradio/integration)
• ruff check + ruff format --check — clean
• paperbanana sweep --manifest examples/sweep_manifest.yaml --dry-run — plans 16 variants correctly, writes sweep_report.json
• paperbanana sweep --manifest ... --vlm-providers gemini — correctly errors with the mutual-exclusion message

cc @dippatel1994 for review.

[eureka0928]

Hi @dippatel1994 — tagging you for review when you get a chance.

Quick summary:

• Scope: Add --manifest / -m to paperbanana sweep so users can store the full sweep plan (input, caption, axes, max_variants) in YAML/JSON instead of eight comma-separated CLI flags. Symmetric with how batch already works.
• Size: ~50 lines of functional code + ~160 lines of tests. One new core/sweep.py helper (load_sweep_manifest) and a small CLI wire-in. No changes to the sweep runner itself.
• Mutual exclusion: --manifest cannot be combined with the axis flags — errors fast with a clear message.
• CI: All 11 checks pass (lint, build, tests across Linux/macOS/Windows × Python 3.10/3.11/3.12).
• Smoke tested: paperbanana sweep --manifest examples/sweep_manifest.yaml --dry-run plans 16 variants correctly.

Happy to squash, rebase, or adjust based on your feedback.