Add plain.html, a Plain-native HTML template engine, and remove Jinja

[davegaeddert]

Stacked on #128 (plain-forms); retarget to master when that merges.

plain.html is an HTML-aware template engine that replaces Jinja across the repo. plain-templates and plain-elements are deleted. Every in-repo template is ported, plain.pages renders through it, and the checker runs in CI.

The engine

• Templates are HTML with {{ expr }} expressions and {% if %} / {% for %} / {% slot %} / {% fragment %} blocks. Expressions are real Python, not a filter language: .lower(), str(), or "", comprehensions. No {% set %}, no inheritance, no macros, no registry, no autoescape opt-out — components and Python cover those.
• Every template declares its names in YAML frontmatter (attrs:, components:, slots:). Nothing is ambient.
• Components are templates invoked as tags (<Pagination page_obj="{{ page_obj }}" />) with typed attrs: and named slots. Call sites are checked against the component's declaration.
• Compiled to Python ahead of time with a disk cache; tracebacks map back to template source; contextual autoescape by construction.
• {% fragment "name" %}…{% endfragment %} renders inline normally, and render(..., fragment="name") returns only that region. HTMX fragment requests are rebuilt on it, including status_code= forwarding. Fragments inside components are reachable.
• Tooling: plain html check [--typecheck], plain html format, plain html compile. --typecheck extracts every expression into a synthesized module and runs ty over it against the declared names, so a misspelled attribute or a Jinja-ism like is false fails before render. --template-dir runs both commands standalone with no app or database; a plain-html console script exposes that mode. ./scripts/html-check and ./scripts/html-format walk every package and every test app and run in the Postgres-less CI lint job. plain html format --check is part of ./scripts/fix --check.

Porting notes worth knowing

• Jinja's silent fallbacks (dict attribute access, StrictUndefined sentinels, is none, filters) all became caught errors under --typecheck; several had shipped as latent bugs.
• Logic that Jinja templates carried with {% set %} moved into Python. The admin list's three-state sort header is now get_column_headers() on the view, which also fixed a bug where an unsorted list rendered no sort links.
• lazy=True on HTMX fragments is not carried over: component slot content renders eagerly, so it would defer nothing. The README points at dedicated templates instead.
• plain-html/ROADMAP.md lists deferred follow-ups: global attribute pass-through, always-on component validation, :values enums, repeatable slots, the static/dynamic compiled representation as the optimization path for fragments, and lint tiers. {% let %} is recorded as the candidate if a second real case appears that Python can't absorb.

Review shape

The engine (plain-html/plain/html/: parser, compiler, typecheck, loader, format) carefully; plain-html/README.md for the authoring surface; the ported templates sampled, with admin/list.html and the oauthserver template as the two largest.

Verification

• ./scripts/fix --check clean (ruff, ty, oxlint, oxfmt, prettier, html format --check)
• ./scripts/html-check — 120 templates across every package and test app, 0 errors
• ./scripts/type-validate 25/25
• ./scripts/test — 3421 passed, 1 skipped, across 25 packages plus the example app