Compute — two smooth summands

Independent checks for Green #59 / Erdős #334. Nothing here is a published-exponent improvement.

script	what it certifies
`trivial_cover.py`	$F(n)<2\sqrt{n}+1$ on a stated range, via the square-plus-remainder split
`obstruction.py`	Jacobi form of the negative-pseudosquare lemma; exact $F(131486759)=83$
`g_of_y.py` / `g_of_y.c`	first missing sum $G(y)$ by exact bitset coverage of $S_y+S_y$
`covering_search.py`	residue-class and factorization templates; prints holes
`f_exceptions.c`	exact exceptions to $F(n)\le n^{p/q}$ via $F^q>n^p$
`plot_exceptions.py`	the two figures under `../figures/`

Heavy $G(y)$ (y ≥ 37) should use the C bitset:

cc -O3 -std=c11 -o g_of_y_c g_of_y.c
./g_of_y_c 73 131486759 200000

Exact exception sweeps:

cc -O3 -std=c11 -o f_exceptions f_exceptions.c -lm
./f_exceptions 0.4 1000000 2 5

A hole-free prefix is not an asymptotic bound. See CONSTANTS.md.

The 2026-08-27 closed-form search lives in q1/. Replay:

cd q1 && ./run_all.sh

That folder does not improve the exponent. It records why the closed-form templates fail, including two infinite failure families.