Add Lightpanda as alternative to Chrome for the browser automation tool

[krichprollsch]

Summary

Adds Lightpanda as an opt-in alternative to Chrome for the browser automation tool. Lightpanda is a low-memory CDP-compatible headless browser, but its connection model differs from Chrome — it requires one CDP connection per tab (each connection is its own browser server-side) instead of multiplexing tabs/contexts over a single WS.

• pkg/browser/ — new Backend selector (chrome | lightpanda) with /json/version auto-detection. Lightpanda path mints a fresh CDP connection per tab (Target.createBrowserContext + Target.createTarget per conn). Tab tracking, shutdown, and the idle-page reaper all branch on backend via a shared closeManagedPageLocked helper. Tenant isolation on Lightpanda is implicit (every conn = fresh browser).
• Page.captureScreenshot guard — Lightpanda returns a placeholder image, so the screenshot tool action returns a clean error directing the agent to the snapshot (AX tree) action.
• Config — new BrowserToolConfig.Backend field + GOCLAW_BROWSER_BACKEND env var. Empty value triggers auto-detection. Chrome remains the default.
• docker-compose.lightpanda.yml — sidecar overlay running lightpanda/browser:latest, wires both env vars.
• docs/browser-backends.md — compatibility matrix vs Chrome (screenshot, multi-tab, cookie sharing, auto-reconnect) + minimum-Lightpanda-version note.
• Integration tests (//go:build integration, skipped unless LIGHTPANDA_CDP_URL is set) — golden path, AX-snapshot + Eval, multi-tenant isolation, backend reporting, screenshot guard.

Closes #223.

Lightpanda quirks worked around

Live testing surfaced three behaviors that needed code-side handling:

Quirk	Workaround in this PR
Lightpanda numbers targets per-browser; every conn's first target is FID-0000000001, so multi-tab maps collide	Synthesize globally-unique lp-N keys (via Manager.nextLpTabSeq); upstream targetID is only used inside openTabLightpandaLocked
page.Info() returns valid data once post-open, then errors on subsequent calls — ListTabs was silently dropping tabs	Cache URL + Title in pageInfos at OpenTab; ListTabs reads from cache
rod.Browser.Close() calls Browser.close which Lightpanda doesn't implement (UnknownMethod -31998); WS drops cleanly anyway	Swallow the error to silence the noisy log line

Minimum Lightpanda version

The AX-tree snapshot path requires Lightpanda with lightpanda-io/browser#2232 merged (fixes Accessibility.getFullAXTree to return nodeId as a string per CDP spec, instead of a JSON number). Earlier Lightpanda images cause Snapshot to fail with a typed-decode error.

Verified working with lightpanda/browser:latest (digest sha256:00aa5c68...).

Test plan

• go build ./... and go build -tags sqliteonly ./... pass.
• go vet ./pkg/browser/... clean.
• Existing browser unit tests pass (go test ./pkg/browser/...).
• docker compose -f docker-compose.yml -f docker-compose.lightpanda.yml config validates.
• Integration suite passes against a live Lightpanda sidecar:

  docker run -d --rm --name lp -p 9222:9222 lightpanda/browser:latest \
    lightpanda serve --host 0.0.0.0 --port 9222
  LIGHTPANDA_CDP_URL=ws://localhost:9222 \
    go test -tags integration -count=1 -run Lightpanda ./tests/integration/

  All five tests pass: SingleTenant_Golden, Snapshot_AndEval (AX tree returns refs + interactive nodes; Eval("() => document.title") returns the page title), MultiTenant_Isolation, Backend_ReportedCorrectly, Screenshot_BlockedByToolGuard.
• Manual: bring up the full stack with the Lightpanda overlay, verify the agent can open → navigate → snapshot → act → evaluate → close end-to-end, and that screenshot surfaces the guard error.
• Existing Chrome flow still works with docker-compose.browser.yml (Backend auto-detects to chrome from /json/version); covered by the existing browser unit tests.

Notes for review

• Pinned to lightpanda/browser:latest for parity with the Chrome overlay; consider pinning to a fixed version once Lightpanda publishes stable tags.
• The compose file invokes the lightpanda binary explicitly (command: ["lightpanda", "serve", ...]) because the image's default entrypoint isn't lightpanda.
• The compose file's healthcheck uses /dev/tcp via sh. If Lightpanda's image is scratch-based without a shell, drop the healthcheck and switch depends_on to condition: service_started.
• Eval requires a function form (e.g. () => document.title) — same as Chrome via go-rod's Page.Eval. Bare expressions fail on both backends because go-rod's wrapper does (USER_JS).apply(this, arguments).
• No upstream tab listing on Lightpanda — ListTabs reads the manager's local map. Anything that previously relied on m.browser.Pages() to discover tabs opened out-of-band won't find them on Lightpanda (acceptable: only goclaw should be opening tabs in its sidecar).
• No auto-reconnect on Lightpanda. A dead WS = that tab is gone server-side; the user/agent gets a clear "tab not found" / "tab closed" and reopens.
• MaxPages (default 5) on Lightpanda now bounds CDP connections per tenant, not just open tabs. Default is fine; bump if you have heavy parallel-tab agents on Lightpanda.

[krichprollsch]

Re-verified end-to-end with the upstream Lightpanda fix landed.

lightpanda-io/browser#2232 merged earlier today, which fixes Accessibility.getFullAXTree to return nodeId as a string per CDP spec (it was emitting a JSON number, which made go-rod's typed decoder choke). With that in place, the AX-snapshot path — the agent's primary "see the page" mechanism — now works on Lightpanda.

Updated the integration suite accordingly:

• Renamed TestLightpanda_KnownUpstreamGaps → TestLightpanda_Snapshot_AndEval and flipped it from "assert this fails" to "assert this returns refs + populated text".
• Compatibility matrix in docs/browser-backends.md: AX snapshot row is now ✅ on both backends. The former "Known upstream gap" section is replaced with a "Minimum Lightpanda version" note pointing at #2232.

Re-ran the suite against lightpanda/browser:latest (digest sha256:00aa5c68...):

PASS  TestLightpanda_SingleTenant_Golden
PASS  TestLightpanda_Snapshot_AndEval        — 2 refs, 1 interactive, 153 chars
PASS  TestLightpanda_MultiTenant_Isolation
PASS  TestLightpanda_Backend_ReportedCorrectly
PASS  TestLightpanda_Screenshot_BlockedByToolGuard

Sample snapshot output for https://example.com:

- rootwebarea "Example Domain"
        - heading "Example Domain" [ref=e1]
        - paragraph
        - paragraph
          - link "Learn more" [ref=e2]

This unblocks real agent-driven scraping on Lightpanda — open → snapshot → act → evaluate all work. Screenshot remains gated (Lightpanda returns a placeholder image; the tool layer routes the agent to snapshot instead, which is the better channel for the LLM anyway).

Ready for review.