`grammar`: render any grammar in Plotnik syntax, with single-rule retrieval

[zharinov]

Promote grammar rendering to a top-level grammar <locator> [rule] command, backed by a rewrite of the renderer. The current GrammarRenderer (crates/plotnik-cli/src/commands/lang.rs) is the v1 format this redesign explicitly replaces.

Why the v1 format is wrong

The current renderer descends into token internals as if they were children, shows inline-listed rules as queryable nodes, leaks hidden rules into supertype member lists, prints underscore-prefixed category names the query language can't even say, and is unconditionally one-item-per-line (a mid-size grammar → ~2,500 lines). It renders how text parses; query authors need how trees are shaped.

Design: three registers

Every definition body is in exactly one register, each lexically distinct so they can't be confused:

Register	Definitions	Body syntax	Example
Pattern	node shapes, hidden rules	real plotnik query syntax	pair = {key: (string) ":" value: (value#)}
Type	categories (supertypes)	bare names joined by |	value# = object | array | number | string
Text	tokens	string literal or regex	number = /-?(0|[1-9]\d*)(\.\d*)?/

This replaces v1's single overloaded notation that used the same marker ((_x ...)) for opposite things.

First-class categories

Tree-sitter supertypes become a named feature with suffix-# syntax (expression#, (expression#call_expression)), de-underscored names, and contextual-refinement semantics. Uses #, not the deprecated tree-sitter /. The dump's category section is the discovery surface for that feature.

Current state (audit)

Entirely unbuilt. The renderer uses a single uniform render_rule() (lang.rs:135-233), emits :: supertype = (not # / |-union), taps RawGrammar at the start of the pipeline (not ExtractedSyntaxGrammar + node-shapes), lives in plotnik-cli, and has no fragment API or single-rule retrieval.

Scope

• Render subsystem in crates/plotnik-core/src/grammar/render/ — a model + width-aware pretty-printer, tapping the existing pipeline (post extract_tokens/resolve_symbols, joined with extract_default_aliases + node-shapes closures).
• Three-register format + first-class categories.
• Diagnostics fragment API (Def::fragment depth-capped, kind_list length-capped, category member lists). analyze/link.rs:382 hand-rolls a format_list for field hints that belongs here; the grammar-based validation messages (Grammar-based query validation: structural checks from existing grammar metadata #443, Grammar-based query validation: StructureTable + SAT/THREAD engine for sequence & anchor impossibility #444) are the driving consumer.
• Top-level grammar <locator> [rule] (locator from -l as a grammar locator #429). Optional second positional narrows to one rule + the rules it references. No --rule subcommand — the format is engineered so grep is the focused view (rg '^call_expression =').

Acceptance

• plotnik grammar typescript and plotnik grammar typescript call_expression (rule + referenced shapes).
• Native plotnik syntax with # refinements; tokens in the text register; no hidden-rule leakage; compact (not one-per-line).
• Unknown rule → did-you-mean.
• Diagnostics consume the same fragment API.

Companion to the grammar-based-validation issues (diagnostics consumer) and the supertype work (#417 interim reject, #423 implementation). Locator support: #429.