gui: the batch screen stops slowing down - a control is wired once, and the minimal set remembers each format's smallest size - #128

What was wrong

Measured in the real window (Windows, hardware OpenGL) with a probe that drives it through the toolkit's own queue and times a heartbeat on its one goroutine:

• The Start from a preset switch on Several batches got slower with every press: 0.7 s at the first, 6.1 s at the twentieth.
• With a preset switched on, every key typed took ~0.4 s and allocated ~379 MB.

Two causes:

1. Fields.listen and Fields.counter wrapped a control's change callback on every registration. The batch screen registers every field again on every rebuild (an address carries the batch's position) and keeps the controls (so typed text survives). After k rebuilds one change was reported k times, under every address the control had ever had, and a size box counted into every caption it had ever been drawn with.
2. Every report re-settled the form, and with a preset on that expanded empty-and-minimal, which works out the smallest size of each format by planning it - encoding the pictures. 50.4 MB allocated per expansion.

What changed

• wiredOnce / chainOnce (internal/gui/parts/fields.go): the handler is put into a control once and re-pointed at the latest address and caption on every registration. The state lives on the control (Entry, Chooser, Toggle), not in a map on Fields, which would keep a removed batch's boxes alive.
• preset.smallest remembers each format's smallest size per process, by id (format.Register refuses a second descriptor under one id). 0.51 MB per expansion.
• No change to recipes, manifests or bytes: the same number is returned, so the expanded recipe is the same.

Measured after (interleaved with main, two runs each)

Guards

• TestAControlRegisteredAgainReportsOnceUnderItsLatestAddress - a box, a menu and a switch registered five times report once, under the last address, and count into the last caption (red before the fix).
• TestNoControlIsRegisteredUnderTwoAddressesAtOnce - the assumption the fix rests on, asked of all three work screens, the batch screen with three batches, archive contents and a preset.
• TestTheMinimalSetIsWorkedOutOnceAndNotAtEveryExpansion - a second expansion stays under 5 MB.
• All new guards proven by mutation (5/5), three moved mutation entries re-proven (7/7). crowdedDepthFunctions ratcheted 50 -> 49.

Not in this PR: most of the window's remaining memory is fonts parsed once per theme-override scope (26 sets of ~7.4 MB). That is the next, separate change.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Updates and format selections on the Several batches screen no longer trigger repeated recalculations for every box, reducing work and memory use while typing.
  • Control changes and byte counts are now reported for the currently selected setting, avoiding duplicate or misplaced updates.