8382 -> 8004 B of flash, 725 -> 511 B of RAM on a part that has 2048.
The command names were a std::to_array of string_view, and on a Harvard
machine that is the worst of both: the characters land in .data and so does
the table's own pointer-and-length pair for each of them, so the firmware
carried 216 B of RAM for thirteen words that never change - and paid for them
in flash too, since .data is copied out of an initialiser image at startup.
They are one NUL-separated blob in flash now, walked with lpm. Separators
rather than an offset table, because an offset table is the RAM this exists to
give back; the names sit in it in match order, so the walk that finds a name
is the same walk that compares it and measures it. Flash falls further than
RAM does: the initialiser image and the two tables were 216 B of it, and the
blob is 87.
The header said the names "cannot" be in flash because they are matched at run
time. Being matched at run time is not a reason to be in RAM on a machine with
two address spaces - only being *written* is, and nothing writes these.
Two smaller things came with it. `reset`'s exact-match rule was a bool on
every entry to protect one; it is an index found by searching the list, so
reordering the commands cannot move the protection onto a different one. And
`version` was the last string_view left, holding its own characters and a
pointer to them.
The matching is now pinned rather than assumed: lookup() is constexpr and the
battery asserts the load-bearing order the README documents - `s` is show and
not statistics, `st` is statistics, no abbreviation of `reset` resolves, and
`helpful` is not `help`. Red-checked by claiming `s` is statistics.
Both modes byte-identical, ten tests green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
libavr's guidance binds this repo too, and until now nothing here checked it -
`ctest` runs `libavr_format_test()` over this tree's own sources now (rules 11
and 33), skipping rather than passing where clang-format is absent. It caught
drift on its first run: a file written this week and edited after formatting.
Where the README states a measured size, `libavr_size_claim_test()` holds it to
the image and holds the image to the prose: advancing the library pin moved
three of these across the fleet with nothing saying so, and re-recording one
now requires the sentence that quotes it to move too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`(quarters + 2) / 4` is round-half-up only for positive values: C truncates a
negative quotient toward zero, so -3.00 C read as -2, -1.00 C as 0 and -0.75 C
as 0 - nine of fifteen negative quarter-values off by a whole degree, always
toward zero. `(quarters + 2) >> 2` is an arithmetic shift, which floors, and
is right across the whole range. It is also smaller: the shift skips the bias
correction signed division needs.
The formula was written three times - thermistor::celsius(), which nothing
called, and inline at both live call sites - so the defect had three homes and
so would its fix. One `thermistor::whole_degrees()` now, called from both, and
the dead entry point is gone (rules 6, 27).
test/consteval.cpp is new and is what should have caught it: whole_degrees
across zero including both ties and both table limits, the curve's start and
saturation points and its monotonicity, and the thermistor table anchored
where the Beta equation fixes it - the count at which the divider reads the
thermistor's nominal resistance must read the nominal temperature - plus both
clamps and the fall across every step. Red-green: four assertions fire against
the old division.
Beside it: the cubic's three coefficients are named rather than inlined and
restated in prose (rule 5), the consteval table builders take explicit 32-bit
types (rule 25), the curve's clamp reads the table's own size (rule 36), and
the serial override says what expects the rate rather than what the board has
always done (rules 12, 13).
And bootloader::handle_reset()'s watchdog diversion is gone. Its own comment
called it a leftover "kept only because it is free and cannot hurt", and it
did not: pureboot peeks WDRF without clearing it and hands back on purpose, so
a watchdog reset arrives here with the flag still set and the diversion jumped
into the loader with MCUSR already cleared - opening the activation window
that policy exists to close. Clearing MCUSR is the whole job and stays.
8206 -> 8168 bytes, byte-identical between generated and reflect.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
