Plan 87 extends MDS034 (flavor validation, plan 86) with a 13th
feature: GitHub Alerts `> [!NOTE]` blockquotes. Detected via AST
walk on the dual parser, no new goldmark extension needed. Fix
strips the marker line while preserving the blockquote body.

Plan 88 adds a new rule MDS035 that flags four renderer-specific
TOC directives (`[TOC]`, `[[_TOC_]]`, `[[toc]]`, `${toc}`) which
silently disappear on CommonMark / goldmark, and suggests the
existing `<?catalog?>` directive as a portable replacement.

These address the high-value items from the markdown-extension
survey comment. Items evaluated and deliberately skipped:

- Spoilers `||text||`: only Discord/Mattermost, severe ambiguity
  with table pipes, no linter in the ecosystem covers it.
- Generic attribute syntax beyond heading IDs: plan 86 already
  enables `parser.WithAttribute()`; heading IDs are feature 8.
- Pandoc `:::` fenced divs / MyST / MkDocs admonitions: five
  mutually incompatible variants, no standardization, goldmark
  renders them as plain paragraphs (not a lint-visible failure).

Address Copilot review:

- Error message: drop trailing period and capital leading 'For',
  matching the mdsmith CLAUDE.md convention (lowercase, no
  trailing punctuation). New message keeps both pointers — the
  missing heading-TOC generator and MDS019 for file indexes.
- Detection: `[TOC]` is a valid CommonMark shortcut reference
  link. The earlier design would false-positive when a matching
  `[TOC]: <url>` definition is present. Add an explicit
  link-reference map lookup (case-insensitive label) before
  emitting the diagnostic for this pattern; the other three
  patterns keep their simple regex match.
- Tasks: add a step for the link-reference suppression and a
  good fixture exercising it.

Address Copilot review:

- Failure description: earlier text said the directives "simply
  disappear from the rendered output" / render as "empty text".
  That was wrong — goldmark and CommonMark render the literal
  directive token (`[TOC]`, `[[_TOC_]]`, etc.) as paragraph
  text. Rewrite the Context section to describe the actual
  behavior: a visible failure where the author's intent is
  lost but the token itself is shown. Update the front-matter
  summary and Goal paragraph to match.
- Error message: the previous version started with `[TOC]`,
  which read as a capitalized leading token and conflicted
  with the note claiming a "Lowercase start". Reword to start
  with the lowercase word "unsupported": `unsupported TOC
  directive \`[TOC]\`; mdsmith has no heading TOC equivalent;
  use <?catalog?> for file indexes (MDS019)`. The backticked
  directive token is quotation, not prose capitalization.

Also addresses a Copilot comment on PLAN.md: the catalog table
rows for plans 87 and 88 were added by running `mdsmith fix
PLAN.md`, not edited by hand. Re-running `mdsmith fix PLAN.md`
after this change produces no diff, confirming the table is in
sync with the plan files.

Copilot review noted the error message backticked `[TOC]` but
left `<?catalog?>` as bare text, inconsistent with the following
explanation that says directive tokens are backticked. Add
backticks around `<?catalog?>` and update the explanation to
cover both tokens.

Two other comments in the same review batch — claiming the
comparison tables in plans 87 and 88 start each row with `||`
and have an empty leading column — are factually incorrect.
Both tables use a single leading `|`; replies posted on the
threads with the verbatim current content.

Copilot review: the example error message hardcoded [TOC],
which would misrepresent the diagnostic for the other three
patterns ([[_TOC_]], [[toc]], ${toc}).

Change: describe the message as templated on `{directive}` and
show the concrete message each of the four patterns would emit.
The template itself is unchanged; only the documentation makes
clear that the matched token is substituted per call site.

Also explicitly document the mdsmith error-message convention
(lowercase start, no trailing punctuation) inline so that
implementation follows it for every pattern.

Address two Copilot review points:

- plan 87: described the non-GFM render of `> [!NOTE]` as a
  "silent failure", but the literal marker text is visibly
  rendered — that is a visible failure. Reword to match.
- plan 88: the phrase "None are part of CommonMark, GFM, or
  goldmark" was ambiguous — `[TOC]` IS valid CommonMark syntax
  as a shortcut reference link. Reword to make clear that none
  of the four tokens are standardized as TOC directives, while
  `[TOC]` remains a valid (but differently-resolved) link
  shorthand.