test(bulk): convert the fixture corpus in bounded parallel batches

[developer0hye]

File submission policy

Before attaching or committing files, read the submission policy.

• Any submitted sample files or attachments satisfy the submission policy, or none are submitted.

Summary

Last of three follow-ups from the CI review with Codex. After #1938 and #1939, a clean PR run takes 5.7 min wall. The Bulk Third-Party Fixture Test is the only job over 5 min (5.5 min); its gate test converted ~2,700 fixtures one at a time in 236.6s.

run_bulk_test now converts through a new map_in_batches helper:

• Workers: two by default; BULK_CONVERSION_WORKERS=N overrides it, and the module doc header documents the variable.
• Bounded batches: 32 files per batch, and each batch drains completely before the next starts. pdf.rs evicts Typst's memoization only when no compilation is active (TypstCacheState::begin_compilation). A pool that never went idle would therefore never evict and would keep every document's cache alive. Draining restores the idle point that the sequential loop had after every file.
• Order preserved: results come back in corpus order, and per-file progress lines print in that order after conversion. The report and the baseline comparison are unchanged.
• Panic isolation: each file still goes through convert_file's catch_unwind.

TDD: two new non-ignored tests.

• test_map_in_batches_preserves_input_order: results stay in input order across worker and batch sizes.
• test_map_in_batches_drains_each_batch_and_bounds_concurrency: no item starts before every earlier batch finished, and peak concurrency stays ≤ workers while still exceeding 1.

Related issue

Related: #1939

Testing

• cargo test --locked -p office2pdf --test bulk_conversion — 9 passed, including both new tests (first run red: map_in_batches did not exist)
• cargo clippy --locked --workspace --all-targets -- -D warnings and cargo fmt --all -- --check
• This PR's Bulk job runs the full gate against the baseline and records the new duration.

Visual impact

• No rendered PDF change
• Rendered PDF change or visual evidence added
• Reason: Only the bulk test harness changes; converter code is untouched.

🤖 Generated with Claude Code