A power computed in float64 is refused unless its base and its result
are normal doubles, and so is a result other than zero that is not one:
below about 2.2e-308 a double keeps fewer digits, and at 0 or infinity
none. A literal that go/constant reads as 0 though it is not, such as
1e-999999999, is refused too.
A whole exponent up to 4096 either way is computed exactly and kept
when go/constant holds the result exactly, so 2^-1400 is exact.
ErrTooLarge becomes ErrOutOfRange, and its reply, "That needs a number
too large or too small for me.", is true of both.
Model: opus-5-5
internal/calc gets its own tokenizer and precedence-climbing parser in
place of go/parser, which cannot express a power operator; go/constant
still computes. ^ (also **) binds tighter than * / % and a sign on its
left, and groups to the right. % takes the sign of the divisor and is
exact on decimals. A whole exponent is computed exactly while the
result's numerator and denominator stay within 4096 bits; otherwise,
and for a fractional exponent, in float64. A remainder, or the sign of
a negative base's power, that depends on a value go/constant holds
rounded is refused as too large. The welcome text and the replies
mention the new operators; a negative base with a fractional exponent
gets its own reply.
Model: opus-5-5
Remove the template's HTTP service, database and fx wiring. Add exact
arithmetic on go/parser and go/constant, a client that runs simplex-chat
as a child process and drives its WebSocket API, and the bot, which keeps
an auto-accepting address and replies to each message. The image adds the
checksum-pinned simplex-chat v7.0.2 on Ubuntu 22.04.
Model: opus-5-5