#Design notes

The measurements come from examples/bench.j, best of several runs on one developer machine. Compare ratios, not milliseconds: on a machine about twice as fast every conclusion still holds, so a different absolute number is not a regression. jennifer run examples/bench.j shows yours.

#Byte loops, not string indexing

Jennifer strings are rune-indexed, so $s[$i..$i + 1] rescans from the start on every access and a per-residue loop goes quadratic. Sequences are ASCII, so every hot loop converts once to bytes, indexes that, and converts back.

#Accretion into a bytes buffer

lists.push returns a new list each time, so building a result with it copies the whole result per step:

#Plain maps beat clever encodings

Three implementations of translate, over 10 000 codons:

The packed index lost because its three helper calls per codon cost more than the lookup they replaced. So the genetic codes stay plain map of string to string, and a helper call inside a per-residue loop is avoided: case folding is written inline as if ($c >= 97 and $c <= 122).

#Alignment is the one quadratic surface

Everything else is linear in sequence length.

About 70 000 cells a second: linear in cells, so quadratic in sequence length. A few hundred residues a side is comfortable; a chromosome is out of scope. The 50 x 50 row is mostly fixed overhead with run-to-run spread wider than the measurement; the linear shape shows from 100 x 100 on. Memory and the maxCells guard are in Comparison.

The DP matrix is a flat list of int indexed i * width + j, and substitution scores for the two sequences are precomputed into a map keyed by the two bytes, so the inner loop is indexed reads and one lookup. The traceback recomputes which move produced each cell instead of keeping a direction matrix: same answer, half the memory.

#Helpers return, they do not fill

Jennifer passes arguments by value. A helper cannot fill in a list handed to it, because it appends to a copy; helpers return what they found and the caller concatenates.

#No filesystem, no network

Nothing under src/ opens a file, makes a request or runs a subprocess. The deck therefore runs on jennifer-tiny with no capability, and every test fixture is embedded in the test files. The examples that read testdata/ (formats.j, composition.j, search.j, manipulate.j, compare.j) do it with fs as the caller; formats.j is the one written to show that boundary.

#Generated data, hand-written API

The genetic codes, residue properties and enzyme table are wide, repetitive data where a transcription slip is invisible in review. tools/gentables.j and tools/genenzymes.j generate them and hold the data in reviewable form: each alternative code as a short diff from the standard one. They assert invariants before writing: BLOSUM62 symmetric across all 576 pairs, every enzyme site palindromic with its cut offsets inside the site.

Generated data and hand-written accessors live in separate files (tabledata.j and tables.j, enzymedata.j and enzymes.j), so a generator only overwrites files holding nothing but def const. The first version had no split, and regenerating deleted accessors appended to the output.

#Correctness

The tests distinguish three kinds of check:

• External anchors, values published independently of the deck. Bradykinin's average mass, a mass-spectrometry calibration standard: the deck computes 1060.22 against a published ~1060.2, which validates the residue masses and the residues-plus-water convention together. Translation is anchored on the human preproinsulin CDS and its published protein, so a slip in the codon table shows up as a wrong protein.
• Cross-implementation checks, values computed separately from the same published parameters: the nearest-neighbour Tm and the isoelectric points. They catch coding errors, not parameter errors, and are labelled as such.
• Property checks, invariants independent of parameters: charge falls monotonically with pH; Tm rises with GC, length and salt; each residue's synonymous codon fractions sum to 1.0; a digest's fragments reassemble to the original; an alignment's rows are the inputs with gaps added.

#Not included