refactor: collapse parser text/cdata handlers, centralize value checks (9.1.0)

[derduher]

What changed

XMLToSitemapItemStream maintained two near-identical SAX switch statements — the text handler (~40 tags) and a cdata handler covering only the 8 tags patched for issue #445, duplicating ~150 lines verbatim. Both events now route through a single handleCharData, and the parser's inline value checks are delegated to shared guards in lib/validation.ts.

• CDATA is now ordinary character data in every tag (e.g. <priority><![CDATA[0.5]]></priority> now parses). Previously anything outside the 8-tag list was logged as "unhandled cdata".
• <loc> is validated with the same validateURL the sitemap index parser uses, adding a URL-parseability check on top of length + protocol. A malformed URL that passed the old inline check but fails new URL() is now dropped with a warning.
• New guards isValidPriority, isValidVideoDuration, isValidVideoRating, isValidVideoViewCount, isValidISODate in lib/validation.ts; validateSMIOptions runs on the same guards, and the bounds (0..1, 0..28800, 0..5) now live once in LIMITS (lib/constants.ts). None of the new guards are added to the public index.ts surface.
• Version bumped to 9.1.0 with a changelog entry; decision recorded in docs/adr/0001-cdata-is-character-data.md; first entry in a new CONTEXT.md glossary ("Sitemap URL").

Net −44 lines (174 insertions, 218 deletions).

Why

This is the top candidate from an architecture review of the parser hot spots: the twin handlers meant every per-tag change had to be made twice (the "add a field" recipe in CLAUDE.md silently required editing both), and the inlined URL/bounds checks contradicted the repo's "all validation in lib/validation.ts" rule while enforcing a weaker loc check than the index parser.

Reviewer notes

• Behavior deltas (both documented in CHANGELOG under 9.1.0, hence minor not patch): CDATA parsing is more permissive; loc validation is slightly stricter, and invalid-loc warnings changed to the Invalid URL in sitemap: … form already used by the index parser. Test churn is confined to 4 message assertions.
• Field-level warn messages for everything except loc are byte-identical to before — the parser still composes its own messages and only the checks moved.
• Per-field append/truncation semantics and the errors/errorCount surface are intentionally untouched.
• <loc> keeps assign-last-wins for multi-chunk character data; accumulating at closetag was considered and deferred (see the ADR).
• New tests: CDATA in tags beyond the [ BUG ] CDATA cannot be used for loc tag #445 set, and rejection of an unparseable loc. The existing alltags.cdata.xml fixture still passes, so [ BUG ] CDATA cannot be used for loc tag #445 behavior is intact.
• npm run test:full passes (lint, ESM+CJS builds, 387 tests with coverage thresholds, xmllint schema validation).

🤖 Generated with Claude Code