Harvest inline <style> CSS from the live DOM for Baseline feature detection

[smsnedden]

Second slice of the --baseline CSS analysis feature. This PR adds extraction of inline <style> blocks by reading them from a page's live DOM; CSS feature detection (css-tree) and Baseline lookup (web-features) over the collected sources follow in later PRs. (External stylesheets landed in #317.)

Like the external extractor, harvestInlineStyles is intentionally not wired into the CLI or launchTest yet. It is designed to be driven by the Baseline detection pass — a live second-pass browser navigation (the same pass that performs JS API detection) — which is productionized in a follow-up PR. Reading inline CSS there keeps all live-DOM detection in one place and one navigation, and runs it once: Baseline analyzes browser-agnostic source CSS, so there is no need to repeat it across the browser matrix. Keeping this slice separate keeps each PR small and reviewable.

Why read the live DOM instead of parsing the HTML? So that <style> blocks a page injects at runtime (e.g. via JavaScript) are captured — static HTML parsing would miss them. The browser is already running in the detection pass, so it is the ground truth for what the page actually applied.

Changes

harvestInlineStyles(page): Promise<CSSSource[]> (src/baselineCssExtract.ts)

Reads inline CSS from a loaded page's live DOM:

• Collects the text of every <style> element in document order, returning one CSSSource per non-empty block (the same { css, file } shape as extractCSSFromHar).
• Skips empty / whitespace-only blocks.
• The read is strictly read-only — a single page.evaluate that reads <style> text with no DOM mutation and no layout-triggering reads — so it can run inside the shared detection pass without interfering with co-located instrumentation or page state.

Out of scope (documented in the JSDoc; may be revisited later): only <style> element text is read, so CSS with no <style> text — rules added via CSSStyleSheet.insertRule() or constructable adoptedStyleSheets — is not captured, nor are <style> elements inside shadow roots (querySelectorAll does not pierce shadow DOM) or <template> content.

Types (src/types.ts)

• Document CSSSource: css is the field that gets parsed; file is a provenance label for reporting — a URL for external stylesheets, a page-derived label (e.g. https://example.com/ (inline style #1)) for inline blocks. It is a display string, not a URL to parse.

Tests

Extends packages/telescope/__tests__/baselineCssExtract.test.ts with 6 tests parameterized per rendering engine (chromium/firefox/webkit; Firefox-only in CI), driving a real page via setContent:

• Single <style> block extracted as one source
• Multiple blocks preserved in document order
• <style> injected by JavaScript after load is captured (the case static parsing misses)
• No <style> blocks → empty array
• Empty / whitespace-only blocks skipped
• Read-only guarantee: harvesting leaves the document markup unchanged