reviewer3: compile FULL source + page-trim PDF (fix pdflatex failures)

[dangng2004]

Stacked on top of #85.

Problem

The reviewer3 perturbation run hit ~37% per-paper failure rate. Three distinct failure modes:

• pdflatex produced no usable PDF (4 cells across cs_CC paper_005, paper_008) — token truncation in `prepare_units` cuts the LaTeX-as-md source mid-environment, so pdflatex fails on the staged file.
• Some affected papers also had `\input{mypreamble.tex}` referencing files we don't have, which pdflatex fatal-errors on. The included preamble defined common shortcuts (`\bbC`, `\calA`, `\vvirg`, `\ootimes`, …), so even if the include is stripped the body fails on undefined commands.
• `UnicodeDecodeError: 'utf-8' can't decode byte 0xaa` (1 cell, cs_CC paper_009) — `subprocess.run(text=True)` blew up trying to decode pdflatex's own log bytes.

Fix

Switch the reviewer3 system to compile the full pre-truncation source and trim the rendered PDF to its first N pages — matches the `max_pages: 20` convention coarse already uses.

Three additional pdflatex robustness layers:

1. Strip orphan `\input`/`\include` whose target file isn't bundled.
2. Inject a defensive `\providecommand` preamble for common author shortcuts (`\bb[A-Z]`, `\cal[A-Z]`, `\bf[a-z]`, `\eps`, `\vvirg`, `\ootimes`, …) — `\providecommand` is a no-op when already defined, so it's safe blanket coverage.
3. `subprocess.run` switched to bytes mode so pdflatex's non-UTF8 accent bytes don't kill the Python side.

Files

• `benchmarks/perturbation/systems/reviewer3.py` — thread `src_corrupted` + `max_pages` into job payload
• `benchmarks/perturbation/systems/reviewer3_adapter.py` — `Reviewer3Job.source` + `.max_pages` fields; `_ensure_pdf` accepts and honors them; helpers `_strip_orphan_includes`, `_inject_rescue_preamble`, `_trim_pages_to`, `_maybe_trim_pages`
• `benchmarks/perturbation/run_benchmark.py` — `Config.max_pages: int | None = None`
• `benchmarks/perturbation/configs/full_*_reviewer3.yaml` × 8 — `max_pages: 20`

Test plan

• Smoke-validated end-to-end on three previously-failing cs_CC papers — all three now produce 20-page trimmed PDFs in 2–4s.
• Re-launch the perturbation runner against the same configs; the idempotent skip-completed logic preserves the cells that already succeeded and retries the failed ones.

🤖 Generated with Claude Code

[dangng2004]

Auto-closed: fast-forward push moved this commit onto reviewer3-integration (now part of #85). Branch deleted.