Grammar wishlist from the docs: seven constructs the examples use that the parser does not implement #27

Filed so the grammar questions can be evaluated on their own, separately from getting the docs site rendering again. The examples are being rewritten onto today's syntax first, so this is not blocking anything.

How this was found

Every live <notation> element in docs/content was extracted and run through load() on v1.1.0.

live examples: 148 | failing: 39 | clean: 109
clean but containing ___ (renders as literal text): 5
pages affected: 8

NotationBlock.updatePreview() logs to the console and renders nothing when load() returns errors, so those 39 are blank boxes on the site with the explanation only in devtools.

The constructs

What each would take, and the question it raises

Bar separators (issue 19). | is excluded from IdentChar and never declared as a token, so it fails at the tokenizer. The cheap version is a BAR token plus an Atoms production that accepts and discards it, which makes the docs render and gives the reader nothing. The question is whether a bar should instead assert a cycle position, so Sw: S R | G M reports an error when the bar falls somewhere the cycle says it should not. That is a real feature with its own error reporting, and it is the reason people write bars in the first place.

Silent spaces (issue 17). The Space -> UNDER_SCORE { newSilentSpace } production already exists; %token UNDER_SCORE does not. _ is a valid IdentChar, which is why ___ currently lexes as an IDENT and renders as visible text. Declaring the token at a priority above IDENT fixes it. Two questions: does one _ equal one silent space regardless of run length, which is what the existing production implies, and does any real notation use underscores inside a literal name, which would change meaning under the new token.

Sharps. # is excluded from IdentChar. The cheap version is to stop excluding it, so M# becomes a literal named M# with no semantics, which renders correctly and means nothing to the layout engine. The real version models accidentals, which needs a decision about whether the library represents pitch at all. Today it deliberately does not, per src/ARCHITECTURE.md: literals carry no assigned meaning. \accidentals(M#) as a document-level declaration only makes sense in the second reading.

Free-time passages. \noTala is used for alapana and tanam sections, where there is no cycle and no beat grid. Adding the command is trivial; deciding what GlobalBeatLayout.processBlock does when there is no cycle is not. Worth checking whether the existing layout code can express an unmetered line at all.

Paren gamaka grouping. ((S R G) kampitam) is a function-call shape with a named gamaka, which is a different model from the ~ operator the grammar implements, where the embellishment attaches to a literal. Either the docs are describing a design that was never built, or [ ] grouping and ( ) gamaka application are meant to coexist. Worth settling before either is touched, since it affects how the embellishment reference reads.

Floating embellishment operators. Lit -> Lit POST_EMB requires the operator to attach to a literal. The docs write it standing alone between groups. Same underlying question as the row above.

Suggested order, if these get picked up

Issue 17 and the sharps row are the two where the cheap version is genuinely the right version and needs no design work. Issue 19 is worth doing properly rather than cheaply. The last three are design questions before they are grammar questions.