Fixes a detection failure that affected most real prompts.

The extension read the composer with textContent, which concatenates block
elements with no separator between them. Every editor on the supported sites is
block-based, so a prompt written over more than one paragraph arrived as a
single unbroken run of characters, and any secret sitting at a block boundary
stopped matching. Measured on the markup these sites produce: two planted
secrets in a two-paragraph prompt, zero flagged, and no indication anything had
gone wrong. Anyone pasting a config file or a log, which is the case this exists
for, got no warning at all.

It reads the rendered text now, so line breaks are preserved. Redaction keeps
the same paragraph structure it did before.

Detection rules themselves are unchanged: 83 detectors, same behaviour, same
allow and redact lists.

Fixes the logo, which the previous release made invisible. The mark is parsed
as XML, and without an xmlns declaration every element lands in no namespace at
all, so the browser treats them as unknown tags and paints nothing. Size and
child count stay correct, which is why it went out: the only symptom was a
missing mark in the popup, the settings page and the in-page badge.

Detection is unchanged. Nothing about scanning, redacting or blocking differs
from the previous version.

First update since 0.1.0, and a large one.

Detection went from 41 to 83 detectors: AWS, GitHub, GitLab, Stripe, Slack,
OpenAI, Anthropic, Google, Azure, Twilio, SendGrid, Figma, Atlassian, Terraform
Cloud, Dropbox and more, plus PEM private keys, .env secrets, database
connection strings, JWTs, and personal data like Luhn-valid card numbers, IBANs
and national identifiers. Every match still explains why it was flagged.

Live detection now runs on ChatGPT, Claude, Gemini, Google AI Studio, Microsoft
Copilot, Perplexity and DeepSeek. Composer detection handles shadow-DOM editors
and nested focus, so a site redesign is less likely to break it silently.

Two fixes worth calling out. Overlapping matches used to be resolved by keeping
whichever started first, which could leave the tail of a longer secret visible;
overlaps now cover both spans. And a single malformed pattern in your allowlist
used to throw on every keystroke, which stopped detection entirely; bad patterns
are skipped now.

Block mode also treats a composer too long to scan in full as unresolved rather
than clean, instead of quietly letting it through.

Still zero network requests by default. Nothing is collected, nothing is sent.
The optional local stats endpoint is off unless you turn it on and refuses any
address that is not loopback.