Build hybrid browser worker loop and benchmark harness

[rsproule]

Stack

This PR is stacked on #86 and uses its structured personal profile and vault-backed benchmark fixtures.

Browser worker architecture

• expose deterministic Playwright execution as the primary surface for inspection, extraction, and related safe interactions
• add the Kernel Browser Loop semantic catalog for snapshot, text, find, wait, and relaxed browser_act fallback plans
• retain visual computer control as the final fallback for coordinate or rendering-dependent interactions
• preserve Kernel session ownership, secure vault autofill, and worker cleanup behavior
• currently pin the worker to zai/glm-5.2 with low reasoning while model candidates are evaluated

Benchmark and observability

• exercise eleven real transaction-boundary tasks, adding Yankees tickets, an Elsewhere concert, facial moisturizer shopping, a nonstop flight, and a weekend car rental to the original six
• inject a structured synthetic profile plus a fake Visa 4242 vault card into isolated benchmark databases
• run baseline and candidate variants concurrently against isolated Portless applications
• add a standalone shadcn benchmark dashboard with run history, per-task status, time/cost deltas, compact stacked timing breakdowns, durable trace links, and direct Kernel live-browser video links
• calculate time and cost comparisons only for paired tasks where both variants passed; mismatched outcomes show no delta and are excluded from averages
• persist model, Playwright, semantic DOM, visual CUA, setup, waiting, web, vault, and other wall-time categories
• make the gated LLM judge authoritative for task pass/fail; worker self-reported status is diagnostic only
• tell the judge that personal and payment data are synthetic fixtures, with an Apple-only rule for address-correction suggestions that preserves the required final ZIP
• persist judge score and rationale on the task trace page

Latest completed A/B

Run: terra-fast browser-act vs main 11-task suite

• baseline: origin/main at 490f709, before Browser Loop and browser_act, using the main default Sol Fast worker
• candidate: PR head plus the current Terra Fast model edit, isolated as a69ecd3 for the run
• task success: 5/11 baseline to 6/11 candidate
• recovered: Amtrak, Booking.com, and nonstop flight checkout
• regressed: Peek next-month selection and Target moisturizer checkout
• comparable set: 3 tasks passed by both variants
• comparable median time: 112.1s to 329.3s, a 193.7% regression
• comparable P95 time: 422.8s to 355.2s, a 16.0% reduction
• comparable LLM cost: $2.0574 to $3.2756, a 59.2% regression
• total LLM spend: $4.9694 to $8.8332
• judge score: 0.45 to 0.55
• model steps: 237 to 359; tool calls: 201 to 328
• artifacts: .eve/browser-ab/2026-09-01T00-24-12.632Z

Main predates structured profile storage, so it received the synthetic card but could not receive the structured profile fixture. Two judge passes also look overly permissive: main Target stopped at sign-in, and candidate flight stopped at required passenger gender before review/payment. Treat the headline success counts as provisional until those boundary judgments are tightened or manually reviewed.

Validation

• pnpm check — 48 files and 236 tests passed; lint, types, formatting, Knip, and boundaries passed
• pnpm build
• pnpm exec next build evals/browser/dashboard

knip.config.ts is unchanged from PR #86; dashboard discovery is owned through the package export and executable scripts.