feat(viewer): drawer-based run page — compact layout, click-to-inspect

[aktasbatuhan]

A UX pass on the run page, addressing concrete bothers with the old layout (modeled on AlphaEvolve's denser, more intuitive view).

What changed

• Chart + summary header. The trajectory chart now sits beside a compact summary panel (best score / steps / cost / AI calls / models), instead of spanning full width with dead space to its right. (fixes "too much empty space next to the graph")
• Full-width step table, no dead column. The old fixed two-column [step list | detail] left a tall empty gap on the left whenever the detail ran longer than the list. The step list is now a single full-width table. (fixes "empty space on the left after the table ends")
• Click-to-inspect side drawer. Clicking any step row slides in a right-side drawer with that step's notes + code diff + interactive solution viz, coupled in one view — no separate "see the solution" click. The viz loads in an inner <iframe> (so its <script>/document.currentScript and styles stay isolated) and height-syncs to the parent via postMessage. (the AlphaEvolve-style compact view; couples diff + viz)
• Programs browser rows open the same drawer.
• Back button at the top of every drilled-in page.
• Task description in the run header, pulled from the task's initial_program.py module docstring (e.g. "auto — First autocorrelation inequality").

Implementation

• New endpoints: …/step/{n}/detail (drawer HTML fragment) and …/step/{n}/viz (standalone per-step viz page for the inner iframe).
• solution.py gains render_step (renders an arbitrary step's stored code in the sandboxed subprocess, cached under .viz_cache/step_{n}.html) and task_blurb.
• Legacy …/step/{n} URL now 307-redirects to the run page at #step={n}, which the drawer auto-opens (deep links keep working).
• All drawer JS is dependency-free, lives in the base template (so run + programs pages share it).

Deferred from this pass (flagged for later): viewing/editing evaluators, prompts, and human-steer .md files through the viewer.

Verification

• Full suite green (337 tests). Tests updated for the fragment endpoints + drawer markup; added per-step render + task-blurb coverage.
• Verified live: chart+summary header, full-width table, row→drawer with notes/diff/viz (iframe auto-sized to 681px), programs-page drawer, ✕/Esc close, back button, blurb, legacy redirect.

🤖 Generated with Claude Code