docs(contract): document what a job must write to be rollout-browsable

[maharajamihir]

Nothing documented what a job has to write for its output to be browsable in the UI's rollout viewer, so every team reverse-engineers it from server.rs. Adds one ### Browsable rollouts subsection to docs/RECIPE_CONTRACT.md, at the end of "What labctl demands from your recipe". No restructuring of the document.

Facts covered, all verified in the code:

• type = "eval_result" + a marker is the entry ticket. runner.rs::register_outputs parses the marker file as JSON and stores it verbatim as the artifact's metadata.result; that blob is everything the viewer sees. Without a marker there is no result key at all.
• Browsable predicate (server.rs, eval-series builder): result contains traj_path, or gif_path, or a non-empty runs array.
• Two layouts (resolve_rollout_paths): multi (result.runs[].subdir → <artifact_path>/<subdir>/trajectory.jsonl + .../steps/) and single (absolute result.traj_path, steps/ beside it).
• The two hardcoded filenames, called out explicitly because they are where this goes wrong:
  • the multi shape looks for exactly trajectory.jsonl — nothing else;
  • frames must be step_%03d.png. get_artifact_frame builds steps_dir.join(format!("step_{n:03}.png")) from the requested index; it never lists or sorts the directory, so any other name is unreachable and a numbering gap is a hole the viewer can't skip.
• The frame_count trap: get_artifact_rollout counts every *.png under steps/, so the count and the fetchable index space agree only when there is exactly one PNG per step. The production OSWorld eval also writes step_NNN_after.png, so it reports ~2x its real step count and the back half of the scrubber 404s. Documented as the trap, not as the pattern.
• Trajectory fields the viewer reads per line: step_num, action, response (literal "<reset>" renders as a dash), reward. Lines must be in frame order — RolloutViewer.svelte indexes steps[frame] positionally for the frame caption and uses step_num for the click-to-jump target.
• Tab label is runs[].instruction, falling back to slug.

Worked example is the real artifact eval_logs/alfred.nguyen/osworld_cua_micro_ckpt_35k_run_01a01316fcde7a609d00ff87ac353ac9 (marker completed.json, 11 runs[] entries, task_000_.../trajectory.jsonl with 65 lines and steps/ holding 129 PNGs).

Two smaller things found while verifying, both written into the doc as one-clause caveats rather than as separate issues:

• gif_path alone makes a point look browsable in the eval series, but resolve_rollout_paths handles only runs and traj_path, so opening it 404s. The doc says to pair gif_path with traj_path.
• The UI sends ?task=<runs[].index> while the server resolves runs by array position, so an index that doesn't match its position selects the wrong episode. The doc says to keep them equal.

Docs-only change; cargo test --all-features unaffected (36 unit + 9 smoke pass).

🤖 Generated with Claude Code