[nicolas-abril]

This makes room under comp.ts's 64,000 ttok cap for #1438, which adds about 810 ttok. comp.ts goes from 63,968 to 63,521 ttok. It also absorbs #1562. Every change is a plain simplification: no code moves out of the file, and no if is rewritten as an expression.

Compiler (TypeScript)

• bind_set, ty_adt and ck.tld are used where the code repeated what they do.
• quant_live and ty_clo fold into their callers (from comp.ts sheds 65 tokens with the same output #1562).
• seg_new takes its parameters' values instead of words and kinds flattened at each call site.
• emit_match names its tag test once, as js_match already does.
• js_call's tail and js_expr's ty0 default to the value most callers pass.
• F32's intrinsic reuses VIEW, and tag lists are a single string, as one site already did.

C runtime

• ERR_* is an enum.
• term_buf folds into term_blk.
• WL_ARGS takes the count directly and pushes with WL_PUSHN; WL_FRAME and FID_EXIT reuse WL_PUSHN and WL_POPN.
• A dead show_val prototype, redundant casts and single-use temporaries are gone.

IO layer

• io_wait's two walks over the parked ring share one io_parked loop.
• fcntl.h is included once, and IoWork is declared once.

JS runtime

• f32_show lets its shortest-text loop reach infinity and fixes only the spelling.
• io_strerror folds into io_fail, and io_addr checks a byte with one test.
• cmp_new is one line (from comp.ts sheds 65 tokens with the same output #1562).

Review fixes (second commit, +135 ttok)

• emit_clo sets a captured binding's count directly again. bind_set could sink the value and leave bind_pop reading a deleted binding.
• Tag lists test exact membership through ty_tag instead of a substring test.
• The IO_EMIT and CLO_APPLY entries take plain words (W64) again.
• js_expr captures each outer variable once.
• Metal keeps its own ulong/uint/uchar typedefs.
• WL_POPN, WL_ARGS and the term bit macros parenthesize their arguments.
• term_rfc tests RFC_BIT. io_again, io_wait and io_late say "no deadline" plainly instead of through unsigned wraparound.
• show_val writes fixed text with fputs, not a format string (-Wformat-zero-length on Linux GCC).
• io_addr names its byte test, and io_push adds a number.

The fixes change no emitted code on the compiler side. The emitted C and JS differ from the first commit only in the runtime lines above, identically across all 772 compiled tests. Building and running every compiled test (C and JS, 5 s each) gives the same pass set on both commits: C 756/772, JS 758/772. The failures are expected-error IO tests that the local harness doesn't model. A ! test's Metal shader builds with --gpu-build, and tsc reports no errors in comp.ts.

Perf fix (third commit, +20 ttok)

• err_post is a function again. As a macro, combined with the WL_ARGS rewrite, it pushed editdist's SEQ-CPU time from 1.90 s to 2.38 s (1.24x, over the gate's 1.15x limit) and its PAR-CPU time to 1.11x. Neither change slows editdist alone, so this looks like clang laying out the hot loop differently. Bisected locally by applying groups of runtime hunks to main's emitted C.

Gates (minis, at the third commit): repo 54/54, ping 57/57, test 1631/1631, perf 124/124. editdist reads 0.99x on SEQ-CPU and 0.96x on PAR-CPU.

Verification (macOS, M-series)

• Absorbing comp.ts sheds 65 tokens with the same output #1562 changes no emitted C. The only JS change is cmp_new's line join, the same in all 802 JS outputs.
• Emitted C and JS are byte-identical to main for all 1,631 tests/*/*.bend wherever only the compiler side changed. Where the runtime text changed, the diffs are exactly the runtime edits.
• Test runs on the C and JS lanes (rfc, run, reg, base, io, state, eval, halt, stats, show, printer) match main exactly: C 375 pass / 14 fail, JS 399 pass / 20 fail on both. The failures come from my local harness, not from this change.
• ! programs built and ran on Metal.
• tsc reports no new errors in comp.ts.
• Benches (tree-bitonic, hashmap, symreg, raytrace) are within noise of main.
• Linux (Arch, x86_64, GCC 16, a local port of the test gate that runs shards in bwrap): test passes 1631/1631 at the third commit and on its base. editdist's SEQ-CPU median is within 1% of the base's over 15 interleaved runs. The Linux GCC -Wformat-zero-length concern from the review doesn't reproduce: GCC warns about printf("") but not about the conditional printf(i > 0 ? ", " : ""), so the fputs revert is cleanup only.
• Not run: CUDA.

🤖 Generated with Claude Code