feat: 'brow read' — clean text/content extraction (agents bypass snapshot for html+eval)

[detrin]

Problem

Analysis of 5,505 real brow calls across 11 agent projects shows agents overwhelmingly avoid snapshot for content extraction and fall back to raw html + hand-written parsers:

• brow html: 1,241 calls — 99% with no --search/--locator (full-page DOM dumps)
• Of those, 728 piped to python, 389 to grep
• brow eval: 1,173 calls — 418 do inner_text/textContent, 92 json.dumps, 172 write a temp file

Root cause

SNAPSHOT_JS (routes/browser.py) is an interaction map, not a content reader. Text is captured only for unnamed nodes and truncated to 80 chars; there's a hard NODE_LIMIT of 200–400 and repeated items are collapsed. So whenever an agent needs prose/prices/embedded JSON, snapshot drops it and they rebuild extraction by hand every time.

Real example (what agents actually do)

brow html -s 74 | python3 -c "
import sys,re,html
t=sys.stdin.read()
for m in re.findall(r'\"commentary\".*?\"text\":\"(.*?)\"', t):
    print(html.unescape(m)[:3000])
"

Comments in the wild: "look for the post commentary text", "events have name+start_at+city in embedded json", "find sentences with market size / CAGR / USD / billion".

Proposal

Add brow read [--locator <sel>] [--format text|markdown] returning readability-style clean main content (text or markdown), dropping nav/scripts/styles. This replaces the dominant ~2,400-call html+eval pattern and is a big token win.

Tasks to work on

• /browser/{sid}/read route + JS extractor (or Python-side html2text/readability)
• brow read CLI command, --locator, --format
• Tests: data-table page, article page, embedded-JSON page
• README: position alongside snapshot ("snapshot to act, read to extract")