feat(eval-view): add run comparison dashboard and visualize

[psiarkiewicz]

Summary

Part 3 of 3 (stacked on #<PR2_NUMBER>). Split from original monolithic PR #1010 per review feedback.

Adds interactive web UI for side-by-side run comparisons in eval-view.

• Compare UI: Interactive side-by-side viewer (/compare.html) showing trajectory divergence, code diffs, Playwright outputs, and metric charts.
• Guide View: Multi-run selection badges and compare modal in eval-view/guide.html.
• Server Endpoints: /api/compare, /api/download-run, and on-demand /api/trajectory-summary.

[micahjo7]

Our evals execute 1 run per task, and 1 task per guide. In the future, if we expand this (multiple runs per task, or multiple tasks per guide, or both), there are some things to consider:

What does the guide view show for these scenarios? IIUC it shows an aggregate of everything for that guide run on that day?
Given this^, should you be able to compare between different “runs” within the same suite? What about comparing guided vs unguided from the same run? In the guide view UI, you have to click on two separate data points (from separate suites), so not sure if it is possible.
For multiple tasks that run per guide, what will the guide comparison page show? Should you be able to select which task you want to compare (or which run?) similar to how you can select guided vs unguided? It looks like there might be some functionality for this currently, but from my testing it doesn’t look like it is working.

Another (less important) scenario is having multiple suite runs on the same day- I believe the guide view shows only the latest- should it aggregate instead? Or show two separate arrows?

In general, there is some aggregation that occurs for these scenarios when converting to the guide view- in the comparison view, we should be able to fully disaggregate into the specific permutations.

[micahjo7]

General comment for the PR:

there is a lot of code here, and several different components. I'm thinking it might be a better idea to separate this into multiple PRs- the first thought that comes to mind is that the trajectory parsing / summary is a feature in itself. then there is the gd compare workflow, then the UI component which presents and visualizes it all (and runs agent analysis). wdyt?

[psiarkiewicz]

General comment for the PR:

there is a lot of code here, and several different components. I'm thinking it might be a better idea to separate this into multiple PRs- the first thought that comes to mind is that the trajectory parsing / summary is a feature in itself. then there is the gd compare workflow, then the UI component which presents and visualizes it all (and runs agent analysis). wdyt?

That makes sense.

[psiarkiewicz]

Our evals execute 1 run per task, and 1 task per guide. In the future, if we expand this (multiple runs per task, or multiple tasks per guide, or both), there are some things to consider:

Added multi-guide view. For now, you can selected different versions on the same day, but the arrow shows the latest one. Can fix later when we know what's useful. Blending daily scores doesn't seem right.

[psiarkiewicz]

General comment for the PR:

there is a lot of code here, and several different components. I'm thinking it might be a better idea to separate this into multiple PRs- the first thought that comes to mind is that the trajectory parsing / summary is a feature in itself. then there is the gd compare workflow, then the UI component which presents and visualizes it all (and runs agent analysis). wdyt?

Agreed.
Splitting this into 3 chained PRs:

1. Trajectory Normalization Framework: trajectory-parser.ts, viewer updates, and agent hooks emitting trajectory_summary.json.
2. gd compare CLI & AI Variance Diagnosis: CLI command, lazy GCS downloader, prompt modules, and multi-stage subagent synthesis.
3. Dashboard Comparison UI: Side-by-side web comparison interface, guide view integration, and server endpoints.

[psiarkiewicz]

Split complete into 3 stacked PRs:

• Part 1: feat(harness): standardize trajectory parsing and auto-generate summary #1186 (feat/trajectory-normalization)
• Part 2: feat(cli): add gd compare workflow and AI variance diagnosis engine #1187 (feat/cli-eval-compare)
• Part 3: This PR (feat(eval-view): add run comparison dashboard and visualize #1010, rebased on Part 2)