[nicolas-abril]

This PR edits bend2/bend.ts, which AGENTS.md marks as human-written. It was opened at Nicolas's request, as a proposal for the file's owner to review, rewrite or close.

Two hot spots in the parser, found while profiling why the generics_3200 checker cell sits at 1.12–1.21× its pin. Base plays no part: none of the checker benches imports Base, and swapping base.bend for the pin's version or an empty file leaves the time unchanged.

1. Operator lookup (parse_term_ops). After every term the parser sliced the next three characters and shortened the slice until INFIX knew it. At most positions no operator starts, so that was three slices and three missed lookups each time: about 29 ms of generics_3200, the parser's hottest lines (the pin's operator search took about 13 ms). INFIX_BY_HEAD now lists the INFIX keys by first character, longest first: the parser reads the next character, tries only the keys it heads with startsWith (no slices), and a term ends at once when it heads none, end of input included. A new operator must be added there too.
2. Name check (parse_lexeme). Every name was tested against the whole-name regex ^[A-Za-z_]\w*(\.[A-Za-z_]\w*)*$. The loop before it only consumes letters, digits, _ and . after a head character, so a name breaks the words-joined-by-dots rule exactly when a dot is not followed by a letter or _ (foo., a..b, Vec.2d). The check is now k.includes(".") && /\.(?![A-Za-z_])/.test(k): names without a dot never reach a regex, and the regex says the rule directly. It agrees with the old regex on every name the loop can produce up to 4 characters (14.1M names).

Both changes keep the parser's behaviour: the first picks the longest key that matches, as the old loop did, and returns exactly where it would have shrunk to ""; the second rejects exactly the names the old regex rejected.

Checker benches (local, min of 9 runs)

bench	main	this PR
generics_3200	0.397 s	0.380 s
defs_12800	0.316 s	0.288 s
proofs_3200	0.820 s	0.772 s
compute_1600	0.515 s	0.513 s
trees_400	0.631 s	0.625 s

On the cluster, generics_3200 read 1.09× in a verbose perf run under full gate load (it read 1.12–1.21× on main). Together with #2060 it should land about 1.03×.

Checks

• tests/parse and tests/check (458 files) print identical output, error messages included.
• INFIX_BY_HEAD holds exactly the 23 INFIX keys, each under its first character, longest first.
• bend.ts 45425 → 45539 ttok (cap 48000).
• Gates (_run.ts, 3 runs): repo 56/56; ping 81/83, failing 2 BendHub publish checks, as on main; test 1831/1838, failing 7 io/gfx tests with Permission denied, as on main: stale /tmp/bend_io_* files owned by another user on cluster nodes 58 and 72–78; perf 117–119/124, and main gives 118/124 on the same cluster at the same time: merkle and nbody's compile and GPU cells (1.23–1.54× on both) and noisy SEQ-CPU cells. The same base gave 124/124 earlier in the day.

🤖 Generated with Claude Code