A node was a Lua table of ~625 bytes, of which 21 keys pushed it over a
power-of-two hash boundary and eight were style copies inheritance had
splattered down from its parent. A 400 node screen cost ~250 KB and could not
coexist with wifi's buffers.
The tree now lives in src/ui/layout.h as a 16 byte struct in a flat arena, and
splits by lifetime: Node holds what hit testing and repainting need forever,
Spec holds what only measure/place read and is dropped when layout ends. Style
is sparse and resolved by walking parents, so a node naming no colours costs
nothing. Re-layout rebuilds from Lua rather than retaining the inputs.
401 nodes: 8218 B steady, 21050 B peak
Lua: ~250000 B steady
sdcard/lib/ui.lua stays the toolkit and keeps every constructor signature, but
returns integer handles: 627 lines to 374. Composition, the palette and custom
painters are still Lua on the SD card; only primitives now need a reflash.
BREAKING CHANGE: ui constructors return handles, not tables. Use
ui.setText(id, text) and keep per-node app data in a table keyed by id.
The bar repainted itself whole every second. It now compares each field against what
it last painted, adds seconds and a memory percentage, and keys the cache on
gui.getRotation() and ui.themeName so rotation and theme changes still repaint it.
Invalidation lives entirely in Lua; the firmware's push flag and gfx/statusbar.h are gone.
Bindings follow getName/setName/isName, persisted preferences move from sys to a settings
table, and gui.setRotation takes degrees like settings does. A bar that dies mid-run now
keeps its rows reserved rather than silently resizing the running app.
The http table copies crosspoint-reader's signatures exactly -- get/head/delete/post/
patch returning (body|nil, status), download taking maxBytes/expectedSize/sha256, the
same 50000 byte body cap and the same -1 for a request that never left the device --
so a script that talks to a server runs on either firmware. docs/lua-api-parity.md
records that, and every other place the two APIs agree, differ for a reason, or differ
because nobody noticed.
Two crosspoint behaviours are deliberately not copied. It reinterprets a string in
argument 2 of a GET as a request body, which turns a mistyped headers table into a
silent protocol error. More seriously it calls setInsecure() on every request, so TLS
is encrypted but unauthenticated on the very path a firmware update would use; this
verifies against the root bundle already sitting in the framework, and the emulator
confirms expired.badssl.com is refused while a wrong sha256 deletes the file.
Downloading exposed two failures worth naming. A 2KB read buffer on the stack tripped
the loop task's canary because a TLS handshake had already spent it, and the
hand-rolled read loop spun forever on a stream that stopped producing -- HTTPClient's
own writeToStream handles both, so the loop is gone and the loop task gets 16KB.
scripts/gen_lua_stubs.py generates stubs/esp32lcd.lua in the same LuaLS format
crosspoint uses, reading annotations off the luaL_Reg tables so a module's docs sit
with its registration. make test runs --check, which crosspoint's copy never wired up.
lua_app.cpp had grown to 701 lines holding every binding, the module loader and the
app lifecycle, so new bindings landed wherever the cursor was. Each Lua table now has
its own file under bindings/, and the app is recovered from the lua_State's extra
space instead of a file-static, so a second state cannot reach the wrong app.
Three tests each redeclared the binding surface, which broke twice this session when
a binding changed; test/fake_device.lua is now the single stub. `make test` runs all
four suites and pins Lua 5.4, matching the vendored interpreter rather than the 5.2
the tests had silently been using.
evan