Translation follow-ups + TableDefault class attr widening

[mdon]

Summary

Consolidated follow-ups to the translation sweep (PRs #557→#560)
that all merged on 2026-05-21, plus one unrelated TableDefault
fix folded in to keep all phoenix_kit follow-ups in one review
cycle.

1. Parser marker leak (originally #562)

parse_response/2's field-capture boundary regex only terminated
at requested markers. A model that emits an unrequested marker
(e.g. ---TITLE---{{title}} when the caller asked only for
name + description) silently rolled the unrequested block's
content into the previous requested field.

Real-world trigger observed against deepseek-v3.2: a projects
prompt template referenced {{title}} literally (no title
variable was bound for a template resource that only has name +
description), so the AI dutifully emitted ---TITLE---{{title}}
between ---NAME--- and ---DESCRIPTION---. The parser, asked
for only name and description, set name to
"Mitarbeiter-Onboarding---TITLE---{{title}}" because nothing
in its boundary regex matched ---TITLE---.

Fix: bound captures at any ---[A-Z0-9_]+--- marker, not just
the explicitly-requested ones. Codex review then flagged that
the new boundary was not line-anchored — a literal ---WORD---
mid-paragraph (technical content, API docs) would prematurely
terminate. Final form requires \n---<NAME>--- so only markers
on their own line act as boundaries.

Same commit promotes handle_ai_response/2 from defp to
@doc false def so the unit tests can drive it directly, with
the inline OpenAI-shape match replacing the cross-module
PhoenixKitAI.Completion.extract_content/1 call.

2. Validation-order regression test (originally #563)

Backstop for the validation chain in Translation.translate_fields/6:
endpoint → prompt → non-empty → unique-markers → plugin-available.
Stacks multiple input violations and asserts which one wins,
including whitespace-only endpoint/prompt to pin the
trim-then-check contract.

3. Magic-link comment trim (originally #564)

Codex final-review CONCERN on PR #559: long inline comments
explaining the data-phx-skip / 183px / form-ownership root
causes belong in the PR description, not in templates. Trimmed
from 19 lines down to 6 short ones naming the root rules.

4. TableDefault class attr widening (NEW — originally #566)

attr :class, :string on the seven TableDefault.* components
rejected the Phoenix-idiomatic class={[if(...), if(...)]}
pattern at compile time, but every component already accepts
lists internally — they wrap @class inside another list that
Phoenix flattens. Surfaces as a warning in phoenix_kit_ai's
endpoint list table; widening all 7 declarations to :any
clears the warning and lets future consumers use the standard
idiom without relitigating per-component.

Codex review findings (round 2)

Five concerns from a second codex pass, all applied:

• Line-anchored boundary regex (above)
• Stale test comment referencing the removed extract_content/1
• Broader malformed-shape coverage (empty choices, non-binary
  content, missing message)
• Whitespace-only validation tests
• (PR Widen TableDefault class attr to :any so class={[...]} idiom works #566 fold-in)

Supersedes

• Fix parse_response/2 marker leak + drive handle_ai_response/2 via real test #562 (parser fix + handle_ai_response test)
• Pin Translation.translate_fields/6 validation order with a regression test #563 (validation-order regression test)
• Trim implementation-detail comments in magic_link.html.heex #564 (magic_link comment trim)
• Widen TableDefault class attr to :any so class={[...]} idiom works #566 (TableDefault class attr widening)

All closed with a back-reference to this PR.

Test plan

• mix compile --warnings-as-errors clean
• mix format --check-formatted clean on touched files
• mix test test/phoenix_kit/modules/ai/translation_test.exs — 23/23 pass
• mix credo --strict clean on touched files
• Manual: end-to-end SQ retranslation on the dev parent app — should produce clean output with no ---TITLE---{{title}} tail