[azfoo]

🤖 Claude

A restart marks the first measure of a new movement or song in the same file. It now starts the beat units again, as the timeline's first measure does, so a later movement whose first measures have no unit reads its own meter, not the previous movement's.

• meters_of: every movement starts with a unit, at the first row and at each row with restart. Where a movement's first measure has none, meters_of supplies one, assumed and with no beat_unit_id, since no beat carries it:
  • a copy of the movement's first unit, if it has one;
  • else of the unit in force before the movement;
  • else the default.
• A movement's first measure always has starts_here true.
• No unit reaches back across a restart: a first movement with no unit stays N/4, assumed, even when a later movement has a unit.
• Unchanged: a restart on the first measure changes nothing, and a timeline without restarts reads as before. units.py doesn't change.
• How: meters_of splits the rows into movements once, and runs the rule for the start of a timeline on each, with the unit in force before as the fallback. It stays linear in the number of rows: 20,000 one-measure movements take 0.06 s.

Before, a pickup without a unit in movement 2 took the unit from the end of movement 1, not assumed. In the test's example (3/4, then 6/8 tapped in two with a one-tap pickup), the pickup read 1/4 instead of 3/8, and every quarter stamp after it was half a quarter note early.

Tests

• test_timemap_meters.py, TestRestarts. Its first five tests were written first, and three of them failed before the change:

  • a pickup without a unit takes its movement's first unit, assumed, and movement 2's bar 1 starts at quarter 7.5;
  • a movement without a unit carries the one before, assumed;
  • no unit reaches back across a restart;
  • a restart on the first measure changes nothing;
  • a movement that starts with its own unit.

  After the code review, three more tests: a third movement that copies a supplied unit, a movement that carries the unit set in the one before it, and the measure table's rows for a restart (movement, time signature, assumed).

• Restart cases stay out of shared_examples.json: the app's format has no place for them.

• uv run --package tilia-core pytest packages/tilia-core -q: 874 passed, 21 skipped, 7 xfailed, on v1-prerelease as of 2026-10-19. python scripts/check_packages.py passes. black and ruff are clean.

Guarantees. I neutralised the code behind each one, ran the suite, then restored it:

Guarantee	Code neutralised	Tests failing
a restart starts a movement	_movements never splits	6
the supplied unit is assumed	the copy keeps its flag	8
a movement without a unit carries the one before	no fallback to it	3
no unit reaches back across a restart	the first unit taken from the whole timeline	4
a movement's first measure starts a unit	starts_here only on the timeline's first row	3
the supplied unit has no id	the copy keeps its id	5

Part of TimeLineAnnotator/project-management#323

🤖 Generated with Claude Code

[azfoo]

🤖 Claude

Code review of fad3da68: no correctness bug found. The review traced meters_of through these cases:

• a restart on the first row, and none on it;
• consecutive restarts;
• a restart row with its own unit;
• a movement with no unit;
• a pickup before a movement's first unit;
• empty input.

Each gives what the rule asks, and a timeline without restarts reads as before. What it raised, and what 8d10fa52 does about it:

• Changed:
  • Quadratic in the number of movements. Each movement start without a unit copied the rest of the table, so 20,000 movements took 0.46 s. meters_of now splits the rows into movements once, and takes 0.06 s.
  • The first unit of a movement was parsed twice, and one condition could never be true. Both went with the restructure.
  • Tests:
    • a third movement copying a supplied unit;
    • a movement carrying the unit set in the one before it;
    • build_rows's rows for a restart, so that the table, not only meters_of, is covered;
    • the restart on the first measure is now set on the row only, so the test shows which source it reads.
• Not changed here:
  • The app and restarts. feat: beat patterns, beat units, time signatures and quarter stamps #621's loader copies the timeline's first unit onto beat 1 whatever restarts lie between, so once the app reads files with restarts, it and the core would disagree. The app can't read them yet. This is noted for the work that switches the app to the core's rules, TimeLineAnnotator/project-management#10.
  • MeasureMeter's docstring lists where beat_unit_id is None without movement starts. units.py is feat: beat patterns, beat units, time signatures and quarter stamps #621's file, copied unchanged, so the wording changes there first. It's still right in general: None where no unit set on a beat governs.
  • Movement starts are worked out in three places: table.py, meters_of and _movements. Working them out once, in rows_of, would touch more than this task needs.
  • A restart on a row whose downbeat has no time is lost by rows_of. build_rows refuses such a timeline before meters_of runs.

802 tests pass, and check_packages.py passes. The guarantees table in the description is updated: removing each guarantee's code now fails 3 to 8 tests.