Cli improvements + frontmatter - #2

Pick install destination by trying these in order, first writable wins:
  1. $INSTALL_DIR (env override)
  2. $(brew --prefix)/bin
  3. /opt/homebrew/bin (Apple Silicon Homebrew default)
  4. /usr/local/bin (Intel Homebrew / classic default; usually root-owned)

Fixes the previous script always copying to /usr/local/bin/ which on
Apple Silicon required sudo and didn't match where the existing binary
lived. With Homebrew installed on Apple Silicon, the script now lands
the binary at /opt/homebrew/bin/web-to-markdown and prints the chosen
path. Fails with a clear error and the candidate list when nothing is
writable.

Three new CLI flags for bridging the gap between page-load and SPA
hydration. All bounded by the existing --timeout.

  --wait <seconds>          Fixed delay after the page loads. Decimal
                            allowed (e.g. --wait 0.5). Default 0.
  --wait-for <selector>     Wait until at least one element matches the
                            CSS selector. Event-driven via MutationObserver,
                            no polling.
  --wait-for-text <text>    Wait until the body's visible text contains
                            this substring. Event-driven via MutationObserver.

The waits chain in order: load → fixed wait → wait-for selector →
wait-for-text → extract. Combine freely.

Implementation:
- Switched extraction from evaluateJavaScript to callAsyncJavaScript so
  the wait sequence and the chrome-stripping (--main) live in one async
  JS function with arguments passed safely (no string-escape gymnastics).
  As a small bonus, the implicit microtask hop seems to give SPAs like
  YouTube enough time to set document.title before extraction even
  without an explicit --wait.
- The JS uses MutationObserver to detect selector/text appearance —
  resolves on the first matching mutation, no polling loop.
- Outer Swift-side timeout (--timeout, default 30s) bounds all waits;
  if the JS never resolves, the timeoutTask fails the fetch.

Tests: 4 new WebPageFetcherTests cover status/final-url capture,
wait-for matching immediately, wait-for timing out cleanly, and fixed
wait actually delaying extraction. All 58 tests pass.

Two ergonomics improvements for library consumers:

1. FetchOptions struct in WebToMarkdown: bundles the optional knobs so
   call sites don't grow with every new feature. New overload
   `WebPageFetcher.fetch(from:options:)` accepts it. Old positional
   overload kept for backward compat — it just builds a FetchOptions
   and forwards.

2. Frontmatter type (new file Frontmatter.swift): holds (page,
   metadata, fetchedAt) as raw Swift values, with a `format()` method
   that emits the YAML block the CLI produces. Consumers who want the
   data without formatting can read the stored properties directly;
   consumers who want the YAML can call format(). yamlEscape is also
   exposed as a public static for callers who want the same escape
   rules on their own values.

CLI now uses both: builds a FetchOptions, calls
`Frontmatter(page:, metadata:).format()`. Old private formatFrontmatter,
yamlEscape, and truncate helpers in the CLI removed.

Tests: 11 new FrontmatterTests covering the format output, value
escaping (colons, quotes, backslashes, hashes, leading/trailing
whitespace, empty values), newline flattening, plain-value passthrough,
description truncation. All 69 tests pass.

Also: README updated to use plain "YAML" (it parses as real YAML on the
consumer side; the "fake" framing was misleading). Library example in
README updated to show FetchOptions + Frontmatter usage.

ISO8601 strings contain ':' so they need the same quoting as URLs and
other colon-containing values. Previously the timestamp was emitted raw,
producing technically-invalid YAML.

Frontmatter.format() quotes any value containing ':' (which both URLs
and ISO timestamps do). The README example previously showed unquoted
forms that didn't match real output.

Three test fixes:
- waitForExistingSelectorReturnsImmediately: replace hardcoded 'elapsed
  < 5' with 'elapsed < timeout - 1' parameterized against the configured
  timeout, so a slow CI/network doesn't flake the test.
- fixedWaitDelaysExtraction: loosen 'elapsed >= 1.0' to '>= 0.9' for
  ~100ms timer-precision/scheduling tolerance.
- formatsAllFields: update expectation to the quoted form now that
  fetched-at runs through yamlEscape.