[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.

Three small fixes from the complexity issues, each a few lines, none changing what the parser builds.

• [complexity, parser] O(n^2) — Each field name is compared with every earlier name #2005: field names. parse_tele rejected a duplicate field by comparing it with every earlier field (tele.some). A Set of the names seen now answers it: O(n) for n fields instead of O(n²).
• [complexity, parser] O(n^2) — Groups rescan call arguments to find array updates #2022: array updates and operator namespaces.
  • term_write collected every argument of a group's call to test whether it is a four-argument Array.set, once in each enclosing group. It now follows at most five links: a fifth already rules the form out.
  • parse_term_ns read the namespace type's head again at every operator it renamed. Its callers now read the head once and pass it in.
• [complexity, parser, flattening, checker] O(n^2) — Each use of an outer variable searches past every local #2046 (the match-lowering part only): matching a later parameter. When the matched variable was the k-th parameter, match_flatten wrapped one parameter at a time, searching and copying the parameter list again at each step. It now finds the parameter once and wraps the k parameters before it in one pass. The lambdas come out the same and in the same order, and the skipped steps could never take another branch: the only test before them does not depend on which parameter is first.

The rest of #2022 and #2046, and the other complexity issues, need larger changes. The parse_var scope scan from #2046 comes in a separate PR.

Load times (in-process book_load, best of 12, machine under load)

program	main	this PR
a call of 2000 arguments inside 2000 groups (#2022)	91–100 ms	12–22 ms
a constructor with 4000 fields (#2005)	142–237 ms	23–39 ms
a match on the last of 2000 parameters (#2046)	28–30 ms	10–11 ms

Checks

• Every .bend file under tests/ (1851 files) prints the same output with --check-only as on main, error messages included. Among them, 19 tests use x[i] <- v, 234 use namespaced operators, and 16 hit the match-lowering errors.
• bend.ts 45508 → 45590 ttok (cap 48000).
• Gates (_run.ts): repo 56/56, ping 83/83. Test 1833/1838 fails only io/gfx tests with Permission denied from stale /tmp/bend_io_* files on cluster nodes, as on main. Perf 119–121/124 fails the same cluster cells as main does today: merkle and nbody, compile and GPU time. Two earlier runs hit the 30 s cap while the local machine sat at load 20–28; ping passed 83/83 on its own in 8 s.

🤖 Generated with Claude Code