fix: harden /browse lifecycle and ref safety

[morluto]

What This PR Does

• hardens the /browse daemon lifecycle so stop and restart complete deterministically, startup is serialized, recreated state files do not wedge shutdown, and concurrent shells do not race a replacement daemon
• switches snapshot refs from global live locators to tab-scoped frozen ElementHandles so refs do not leak across tabs or drift after SPA rerenders
• fixes snapshot correctness around filtered and skipped nodes, including contiguous ref numbering and stable nth() bookkeeping for -d and -c
• fixes browse command semantics for explicit-origin cookies, empty-string fill and select, quoted accessible names, and same-URL network attribution

Context

The original /browse implementation had a few correctness problems that were acceptable in a single-shell happy path but broke down under real agent usage:

• live locator refs could silently retarget a different element after SPA DOM changes or tab switches
• stop and restart could report success while the daemon was still in an ambiguous shutdown state
• concurrent CLI calls could race startup or replacement-daemon creation
• network attribution by URL could mix up concurrent requests to the same endpoint

The common theme is that wrong-target behavior is worse than an ordinary failure because it looks successful while mutating the wrong tab, element, or daemon state.

Implementation Overview

1. Snapshot refs and DOM targeting

• BrowserManager now stores snapshot refs per tab as frozen ElementHandles instead of one global Locator map
• stale or detached handles are normalized back into the same clear run snapshot guidance
• snapshot now resolves a concrete handle before storing a ref, so refs only exist for resolvable nodes
• ref numbering stays contiguous even when some accessibility nodes are skipped
• skipped nodes from -d, -c, or -i still advance same-role/name occurrence tracking when they affect later nth() disambiguation

2. Daemon lifecycle and startup serialization

• stop and restart return their HTTP response first, then schedule shutdown on the next event-loop turn
• the CLI waits for the old PID to actually exit before considering stop or restart complete
• startup is serialized with a lock file so parallel CLI invocations do not spawn duplicate daemons
• the daemon keeps its state file until actual exit so concurrent shells do not mistake a shutting-down server for a dead one
• recreated state files no longer block shutdown completion
• user-agent settings persist across restart via a settings file next to the state file

3. Command and capture correctness

• cookie accepts an explicit origin so it works before first navigation instead of failing on about:blank
• fill and select now accept explicit empty-string values
• snapshot parsing preserves escaped quotes in accessible names
• network capture uses Playwright Request identity instead of reverse-scanning by URL, so same-URL requests keep the right status/size/duration pairing

Reviewer Guide

Suggested review order:

1. browse/src/cli.ts, browse/src/server.ts, browse/src/meta-commands.ts
   Lifecycle, startup locking, stop/restart semantics, state ownership
2. browse/src/browser-manager.ts, browse/src/snapshot.ts
   Ref model migration, stale-ref handling, request-identity tracking
3. browse/src/write-commands.ts, browse/src/read-commands.ts
   Ref consumers and command-semantic fixes
4. browse/test/commands.test.ts, browse/test/snapshot.test.ts
   Regression coverage for each failure mode fixed here

Risk / Compatibility

• This intentionally changes /browse ref semantics: stale refs now fail instead of silently drifting to a different element
• The daemon lifecycle is stricter: stop and restart wait for a real shutdown boundary rather than treating socket drop as success

Verification

• bun test browse/test/commands.test.ts browse/test/snapshot.test.ts
• ./setup sanity check on the merged branch

[vasiliyk]

@dependabot close