Add gold_fidelity vision scanner (v0.6.2)

[MattFisher]

What this adds

The first vision LLM scanner, plus the image-input plumbing every future vision scanner builds on. gold_fidelity shows the model the rendered page image alongside the gold markdown and flags material fidelity errors — wrong numbers, dropped/invented table rows, garbled or mis-attributed text — while ignoring benign formatting differences.

Changes

• _llm.py → judge_batch_vision() — sends the page image as an inspect_ai ContentImage block ahead of the prompt text. The shared YES/NO parse is factored into _generate_judgment(), so the existing text-only judge_batch() path is unchanged.
• _artifacts.py → page_image() — resolves <artifacts-dir>/page.<ext> from the --files-root cache, the same opt-in mechanism the cross-artifact scanners use. No new extraction; it reuses the page.png the pipeline already writes.
• gold_fidelity scanner — registered as an LLM factory, so --model alone enables it. Skips any record without a page image.
• CLI + PLAN.md — --model help lists the scanner; PLAN.md marks the v0.6.2 plumbing + gold_fidelity done.

Verification

• 6 new tests + full suite green (196 passed, 15 skipped), mypy clean, ruff clean.
• Tests use a mocked model, so the suite needs no API key. One test asserts the image actually reaches the model as an image content block.

Credentials

No new credential plumbing — the key flows through the existing --model flag (inspect_ai's Anthropic provider reads ANTHROPIC_API_KEY). Recommended judge: Opus 4.8 (anthropic/claude-opus-4-8) for its high-res vision (2576px long edge) on dense tables/small text; Sonnet 5 for cheaper scale-out; not Haiku (no high-res vision).

---

📌 Handoff note (where we're up to)

This is a cloud (Claude Code on the web) session — nobody has interactive shell access to the sandbox, so this PR covers everything that could be done without a live API call.

Done: image plumbing (judge_batch_vision, page_image) + gold_fidelity, fully mocked-tested.

Blocked on credentials, not code — a live end-to-end run: ANTHROPIC_API_KEY must be added as a secret / env var in the environment's configuration (web app → environment settings); it's injected at container start, so it only appears in a new session. Once a fresh session has it (verify with env | grep -c ANTHROPIC_API_KEY, don't print the value), run the real pass against the benchmark's cached page images:

inspect-dataset scan <samples> --scanner-module <domain-module> \
  --files-root <cache> --scanners gold_fidelity --model anthropic/claude-opus-4-8

Not started (next up): gold_completeness (page regions absent from gold) and reading_order (gold block order vs visual reading order). Both ride the same judge_batch_vision path that's now in place — each is a new scanner module + prompt + mocked-model test, following gold_fidelity as the template. Neither needs a key to build or unit-test.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LLg5cbBcfNyyrRGJVPfsA6

---

Generated by Claude Code