1.1.0: the demo cut every held note short, and the time matrix that makes alignment mean something

[SuperInstance]

Two things, plus a decision that had been left open.

The landing page was playing the music wrong

. and - sat in the demo's REST set where the compiler has them in
SUSTAIN. They are the two most common tokens in the notation and they hold the
note before them, so | Am . . . | played a one-beat chord followed by three
beats of silence instead of a chord lasting the bar. Every sustained note on
the front door was cut to a single subdivision — the first thing anyone
arriving from an announcement hears.

The parity test watched this the whole time and saw nothing. It pins the note
count each preset produces, and a rest and a sustain produce the same count:
they differ only in how long the note before them lasts. The counts matched
exactly, start to finish.

A second implementation checked by counting is not checked. Counts are the
property most likely to survive a divergence.

So the guard now measures what actually differed:

• tests/test_demo.py compares both token sets against notation/parser.py and
  pins the durations that follow from them. Restoring the old sets fails it —
  confirmed by doing exactly that.
• tools/demo_differential.py is new: it runs notation through the real page
  in a browser and through the compiler, comparing pitch, start and duration
  of every note across triplets, 3/4, 6/8, chord qualities, stacks, sustains,
  rests and repeated rows. CI cannot run it — the suite installs nothing and
  assumes no browser — so it is a documented manual check. Against this revision
  all eight cases agree; against the previous one six differ, every difference
  a duration.

The page also claimed 6/8 parity that nothing checked. It is checked now, and
the claim says how.

The demo also gained the missing parenthesised token forms ((hold), (rest),
(let ring), and the rule that any other (direction) holds what is sounding),
none of which it had at all.

The ah-ha

The playground had a text box and a piano roll side by side and never drew the
line between them: a visitor had to infer that the third A4 they typed made the
third block moving in the canvas. That inference is the entire idea of the
notation, and it was left as an exercise.

Each token now lights while the note it wrote is sounding — including the .
tokens the note is held across, so Am . . . illuminates as one thing and the
dots are visibly part of the chord rather than events of their own. A textarea
cannot hold a styled span, so the same text is painted behind it in identical
metrics and lit there.

A silent preview sweeps the playhead once on load, so the page is visibly a thing
that plays before anything is clicked. Canvas and CSS only — it never opens an
AudioContext, so it cannot make a sound nobody asked for — and it ends at the
first interaction and does not run under prefers-reduced-motion. On a phone the
editor used to push Play below the fold; the panel that makes sound now comes
first.

Verified in headless Chromium at three viewport sizes: 97 token spans built,
three rows lighting together on the beat, held sustains lighting with their note,
and no horizontal overflow introduced (485/500 both before and after).

Every token now has a position, sounding or not

Rows divide their bars independently, which is the point of the notation and also
means the columns a reader trusts mean nothing to the compiler:

Melody: | A4  .   C5  E5 |
Lyrics: | the tide came  |

came is written directly beneath C5 and sounds two thirds of a beat after it.
That was not merely unreported, it was inexpressible — there was no
coordinate in which the two could be compared, which is why a renderer had
nothing to lay lyrics against, a merge had no space to detect a conflict in, and
a linter had no way to say "this row disagrees with its neighbours".

notation/timegrid.py places every written token on one coordinate system.
bar and unit are derived in TimeGrid.add and nowhere else — a lyric and a
note cannot drift apart by being computed in two places. Rests and sustains are
recorded too, so a token that makes no sound still holds its column.

Three problems become one projection: rendering is x = unit * bar_width;
merging is set intersection on (row, bar, unit) with a disjoint row axis, so two
agents editing different rows provably cannot collide; and
grid.disagreements() names each bar whose rows divide it differently.

Nothing is reported and nothing is enforced — uneven subdivision is legal and
usually deliberate. The grid observes the arranger and does not steer it, which
the fingerprint holds rather than intent.

Phase 1 of proposals/02-the-voyage.md, exit criterion met.

One thing worth keeping from building it

The first version guarded the bar boundary twice — rounding the position and
nudging it before flooring — and neither mutation could be made to fail a test,
because each fixed the case alone. Two guards no test can tell apart are one
guard and one piece of dead code. Collapsed to the nudge, which then failed the
suite when removed.

The open question in the plan, settled

Every other format stacks repeated lyric rows as parallel verses; Plainsong runs
repeated rows on in time. The draft proposed inventing Lyrics 2: to have both.

That is the wrong shape. Run-on is a claim about time, and once lyrics bind to
notes they own no time for a second row to follow on into — so the rule does not
reach them and needs no exception clause.

Every Melody: row in a section concatenates into one melodic stream. Each
Lyrics: row is a verse sung over that whole stream.

No new syntax. Not one of the 6,321 .song files here has a section with more
than one Lyrics: row, and lyrics do not enter the fingerprint, so this can move
no note. It ships gated with the rest of lyric binding rather than on its own.

Verification

615 tests, 7 specs, ruff clean, ok 6334 file(s) checked, 3 warning(s), and
6321 file(s) compile exactly as recorded — the corpus is untouched by all of
it, which for the time matrix is the whole point.

Also drops server.pid, a stale pid file that had been committed, and ignores it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PBAjxy7cD6DzJ72NX8TJEc

---

Generated by Claude Code