Modernize pi-scraper for Pi 0.81

[brandonkramer]

Summary

• upgrade the Pi integration to the current 0.81.1 runtime and native extension/tool types
• resolve model selection, authentication, cancellation, provider errors, and usage through supported Pi APIs
• move resource ownership to Pi session lifecycle events and scope project manifests by trusted session cwd
• migrate repository checks and publishing metadata to Bun, strengthen CI/docs/security checks, and add a durable modernization dossier
• keep all seven public web tools registered while initially activating only common tools plus the web_tools discovery path
• replace repeated whole-batch progress recounts with incremental counters

Why

The extension still relied on pre-0.80 model globals, process-level cleanup hooks, trust-insensitive project caching, type-erasing registration casts, and an always-active tool contract. Batch progress also filtered the entire item list three times for every event, making large batches quadratic.

Impact

• Pi 0.81 model and lifecycle behavior is explicit and type-checked.
• Untrusted projects fail closed and cannot reuse another cwd's manifest cache.
• The initial active tool contract drops from 1,489 to 724 estimated tokens (51.4%) with deterministic selection and invocation metrics unchanged at 100%.
• A local same-process 5,000-transition micro-benchmark dropped progress accounting from 2,378.6 ms to 4.7 ms (about 501x); progress snapshots are also skipped when no update consumer exists.
• Legacy hosts without dynamic activation still receive the existing seven-tool set.

Validation

• bun install --frozen-lockfile --ignore-scripts
• bun run typecheck
• bun run format:check
• bun run lint (0 errors; existing warnings remain)
• focused lifecycle, progress, batch, crawl-adapter, and registration tests: 23/23
• bun run test:selection: 16/16
• bun run test:tools: 7/7
• deterministic contract benchmark: 51.4% reduction, 100% positive/negative/invocation metrics, 0 critical confusions
• bun audit --prod --audit-level=high: no vulnerabilities
• actionlint and package dry-run passed
• independent Cursor 3.12.17 read-only review found no actionable issues

Platform boundary

The broad local Windows suite is not claimed green: its last run retained 20 known failures in unchanged Windows-sensitive path/mode, drive-URL, and SQLite cleanup cases. A broader crawl-runner retry also reproduced the documented Windows EBUSY SQLite cleanup lock, while the changed crawl adapter tests passed. Linux full CI remains the authoritative full-suite gate.

No development server, production build, live provider request, or live browser smoke test was run.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.