DDIR: write the example programs as corgi lambdas

[frankmcsherry]

It also carries a fix that missed #934: a DDIR -- comment inside a lambda now reaches corgi as a # comment, which corgi reads (before, corgi rejected it).

Every scalar slot in DDIR's .ddp programs is now a corgi lambda: the AoC programs, examples/programs, the tests/programs fixtures, and the server and LDBC programs.
No program here uses DDIR's term language ($n, ^n, tuple, list, if, fold, case, infix operators, constructors).
And and or of conditions are written with corgi's (c1, c2, ..) all and any.

Each program computes what it did before, checked against the original on the corgi backend:

• all 33 AoC programs give identical inspect output, not only the checked answer;
• the corgi gate's programs match their originals at 1 and 3 workers;
• the server session scripts give identical output (times and thread ids aside);
• the LDBC smoke run and the SNB suite verify their answers.

Three fixtures lose what tested only the term language: if_literal is gone (a literal-condition if whose branches differ in shape), scalar_ops drops tuple len, and sum_ops drops istag on a non-sum.

binders, case_ops and tour take sums apart in lambdas, which the vec backend cannot run (it types a lambda's input from the row alone).
The corgi gate checks them against recorded output instead: <prog>.expected beside the program, one row per line, from the original programs on vec.
The gate uses a recorded file where one exists and the vec backend otherwise.

A second commit makes each program's shapes plain.
The originals wrapped nearly every single-field key and value in a 1-tuple, because DDIR's projections always built tuples, and the first commit kept those shapes so each rewrite could be checked row for row.
Keys and values are now bare where only the program sees them, and lambdas bind a value whole when they do not take it apart.
Inputs keep their declared shapes, count still emits (n), and the server and LDBC programs keep the shapes of their exports and imports.
Each program computes what it did before up to 1-tuple wrappers, by the same checks, and the three .expected files are regenerated after confirming they differ only in wrappers.

Two more commits: the spread_values fixture is gone (it pinned how a bare $n spliced fields, which lambdas do not do), and count emits a bare Int rather than (n), in both backends and shape inference.
The programs that read counts take a bare n; the server and LDBC programs re-wrap a count where it reaches an export, adding one operator to count_gen.
These were checked against the previous commit's programs run on a server built before the change.

The empty_batch fixture is gone: it needed a projection that fell back to a row-wise path, which no longer exists, and it passed with its guard removed.

The last two commits bump corgi to frankmcsherry/WIP#100 and use it: constant lists are literals (day16's table, day18's powers of two, day03's offsets), computed lists are one n-ary append, and a lambda builds Some(x) as x inject Some.

Not converted here: inline programs in Rust tests (mostly tests/explain.rs), and the SNB program generator server/bench/ldbc/snb/rel.py.

cargo test -p interactive -p ddir-server passes.

🤖 Generated with Claude Code