feat(extract): opt-in --html-as-code to route .html through the AST path (#1230)

[Yyunozor]

Closes #1230.

Problem

.html is always classified as DOCUMENT (detect.py's DOC_EXTENSIONS), so it only goes through the semantic (LLM) extraction path — never the local AST path, even under --code-only, where it's simply skipped. For a project whose .html carries real logic (Angular templates, generated dashboards, anything with meaningful inline <script>), that logic is invisible to the graph:

Without the html templates, the graph loses a lot of its value. (@mfallouh-TS)

The workaround suggested in-thread (.graphifyignore the .html files, @LuisPeregrina) just removes them from the graph entirely rather than making them useful.

Fix

Opt-in --html-as-code flag on graphify extract, off by default (zero behavior change for existing users):

• detect.classify_file() takes an html_as_code param; when set, .html classifies as CODE instead of DOCUMENT, checked before the DOC_EXTENSIONS/paper-heuristic branch so an opted-in file isn't second-guessed by the paper sniffer.
• New extract_html() mirrors the existing extract_vue/extract_svelte/extract_astro extractors: mask everything outside inline <script> bodies (preserving newlines, so source_location stays correct in the host file), then feed the masked source to the plain JS grammar. No new dependency — reuses the masking technique already shipping for three other markup+JS host languages, instead of adding tree-sitter-html.
• <script src="..."> (external, no local body) and non-JS type="..." blocks (e.g. application/json) are blanked, not parsed — they contribute nothing rather than producing garbage nodes.
• The flag persists into update/watch/hook rebuilds via the existing .graphify_build.json mechanism (same pattern as --exclude/--no-gitignore, --exclude patterns only apply to the initial scan — update/watch/hook rebuilds silently re-include excluded paths #1886/Allow extracting git-ignored code: add --no-gitignore opt-out #1971), and is threaded so .html's registered extractor can't leak into the unrelated doc-AST-fallback path in watch.py when the flag is off.

Validation

• New tests: masking edge cases (no <script>, external src=, non-JS type=, multiple blocks, call-graph + line numbers), the opt-in contract (off → unchanged, on → extracted), persistence across update.
• Full suite, branch vs. parent HEAD (e32c9f4): 3395/3381 passed, 1/1 failed (same pre-existing, unrelated: missing optional build dev dependency), 137/137 skipped — delta is exactly the 14 added tests. ruff check clean.
• Real-world smoke test: a ~4,100-line HTML template with 113 embedded functions (canvas rendering + UI logic) produced 0 nodes by default and 113 correctly-attributed function nodes plus their calls edges under --html-as-code — e.g. updateCamera() calling matPerspective() — with line numbers matching the source.

Scope

Deliberately minimal: JS extraction, not full HTML/DOM structure. The CSS-out-of-scope position from discussion #1625 is untouched. Structural HTML parsing is a natural follow-up on this foundation, not bundled here.