build: follow moonbit 0.10.13 and moonbitlang/async 0.22, update all deps

[mizchi]

Summary

The workspace no longer built on the current MoonBit toolchain (0.1.20260915 / moonc 0.10.13): the pinned moonbitlang/x@0.4.40 and bobzhang/toml@0.1.7 fail to compile, and moonbitlang/async changed its HTTP header and cancellation APIs. This PR updates every third-party dependency to its latest release and adapts the code.

Dependency bumps (modules/*/moon.mod, component/moon.mod.json)

package	before	after
moonbitlang/async	0.19.4	0.22.1
moonbitlang/x	0.4.40	0.5.5
bobzhang/toml	0.1.7	0.4.3
mizchi/x	0.2.0	0.6.1
mizchi/zlib	0.4.8	0.4.9
mizchi/simd	0.4.1	0.6.1
mizchi/llm	0.2.2	0.3.2

mizchi/tempfile, mizchi/libgit2 and mizchi/experimental_crypto were already at their latest release.

Code changes

• HTTP headers: @http now keys request/response headers by CaseInsensitiveString. bit keeps plain String maps in its transport-neutral interfaces and converts at the boundary (bit_io_native/http_client_native.mbt, cmd/bit/serve_http.mbt).
• @fs.read_dir returns an ArrayView; the three RepoFileSystem::readdir impls materialise it.
• moonbitlang/core/strconv is empty now; parse_int/parse_int64 come from @string. Dead imports removed.
• async cancellation is no longer an error and cannot be caught, and @async.is_cancellation_error is deprecated (it always returns false). The 123 err if @async.is_cancellation_error(err) => raise err guards are therefore dead code and were removed; the is_being_cancelled() checks stay.
• fragile_catch_all lint (new in this compiler, and fatal under pkf run check's --deny-warn): removed the 60 pure re-raise catch { err => raise err } handlers and moved the three cleanup-in-catch sites (bit cat, bit tree temp dirs, HTTP client close) to errdefer, which also makes them run on cancellation.
• npm CLI payload is now an ES module: moonc emits import statements for the #module("node:fs") FFI bindings that moonbitlang/x 0.5 uses, so the JS build of modules/bit can no longer be required as bit.cjs. It is now copied to npm/bit.mjs; npm/bit.js re-exports it and npm/bin/bit.mjs imports it. package.json exports/files, .gitignore and the pkf task (sync-npm-bit-mjs) follow.
• JS bundle guardrail rebaselined: the tree-shaken minimal lib bundle grew from ~163.5k to 174,458 raw / 40,265 gzip bytes with the new toolchain (same exported surface). Ceilings raised to 180k / 44k in tools/verify-lib-js-treeshake.mjs; the minimal/git-ops ratio guards are unchanged and still hold (0.56 raw, 0.54 gzip vs 0.70).
• CI's setup-moonbit action now pins 0.10.13+cbb11c36f (the old pin cannot compile the new dependency releases).
• moonc stack overflow on the cmd/bit debug build: moonc 0.10.13's link-core pass aborts with an ICE (Error: Stack overflow) when building modules/bit in the debug configuration under the default 8 MiB stack. The same step succeeds with ulimit -s unlimited, and the release build is unaffected. The cmd-native-test job now lifts the limit before building/testing, and AGENTS.md documents it.
• flake.lock refreshed for nix-build: moon-registry advanced to today's mooncakes index (the 2026-08-12 one lacked bobzhang/toml@0.4.3 and moonbitlang/async@0.22.1, failing with attribute '"0.4.3"' missing), and moonbit-overlay to its current HEAD (86fff168, ships the v0.10.13+cbb11c36f toolchain) together with the nixpkgs/treefmt-nix pins from the overlay's own lock. The overlay entry was hashed with nix flake prefetch over git because the sandbox blocks GitHub tarball downloads; the tree has no export attributes, so it is the hash the github: fetcher yields.
• Two pre-existing cmd-bit-1 test failures fixed (93e8778, 44b919f), so the shard is green for the first time in a while: bit grep --column -o gave every match after the first a column that summed the preceding matches' ends — the same bug git fixed upstream in "grep: fix --column --only-match for 2nd and later matches" (2.55, the release third_party/git pins) — and serve --http: a repository resolves under the root asserted a .git suffix strip that http_resolve_repo never did and must not do, since a bare repository is the directory acme/app.git.

Verification

• moon check --deny-warn --warn-list=-27-20-82 (what pkf run check runs) passes for native, js and wasm (also green in CI's test job).
• moon build --target native --release modules/bit succeeds; bit --version runs.
• moon build --target native modules/bit (debug) succeeds with ulimit -s unlimited; the resulting binary passes a smoke run (init, add, two commits, status, log) and real git fsck/git log read the repository cleanly.
• moon test --target native for every non-cmd workspace module plus mizchi/bit/tests: all pass (same loop as the CI full run; also green in CI's test job).
• moon test --target wasm for bit_runtime, bit_diff3, bit_repo, bit_grep: pass.
• moon test --target js -p mizchi/bit -p mizchi/bit_lib: 361/363 pass. The two failures are the ssh-ed25519 signing tests, which shell out to ssh-keygen; it is not installed in this sandbox (ssh-keygen failed with status null). They should pass on the CI runner.
• JS build pipeline reproduced locally (moon build --target js --release for bit_lib and bit, bun bundles, node --test tools/js-build.test.mjs tools/npm-lib.test.mjs tools/npm-cli.test.mjs): 5/5 pass; CI's js-build is green.
• nix flake metadata (with the overlay resolved over git) reports no pending lock changes. A full local nix build was not possible: the sandbox blocks the GitHub tarball downloads the other flake inputs need, so CI's nix-build is the check.
• node tools/check-layers.mjs and tools/check-layer-guards.sh pass.
• The grep --column -o fix was checked against the native debug binary with GIT_CONFIG_* unset (otherwise the shim delegates to the system git): foo_mmap bar mmap now reports columns 5 and 14, abmmapXmmap 3 and 8 — the true columns, where it printed 13 and 9 before.
• All 22 CI checks are green on 44b919f.

Not done / follow-ups

• Remaining deprecation warnings (@sys.get_env_var → @env, StringBuilder::new() → StringBuilder(), ~1.5k sites) are warning-only and left as is.

🤖 Generated with Claude Code

https://claude.ai/code/session_018UDsHQQ4RZzjsqix2HYGBv

[mizchi]

nix-build is red and I cannot fix it from this session.

The job fails inside the moonbit-overlay dependency resolver with

error: attribute '"0.4.3"' missing
Did you mean one of 0.2.3, 0.4.1, 0.4.2, 0.1.5 or 0.1.6?

flake.lock pins the moon-registry input (the mooncakes index) at a 2026-08-12 revision. It predates bobzhang/toml@0.4.3 (2026-08-16) and moonbitlang/async@0.22.1 (2026-09-16), which this PR now depends on. The moonbit-overlay input is from the same date and will need the newer toolchain as well.

Required fix, to be run by someone with nix available (it is not installed in the sandbox this PR was prepared in, and a git+https lock entry needs a narHash I cannot compute without it):

nix flake update moon-registry moonbit-overlay
nix build   # then push the refreshed flake.lock to this branch

Everything else on this PR is unaffected by the lock file.

---

Generated by Claude Code

[mizchi]

cmd-native-test (cmd-bit-1) is red on main too, with the same two tests.

The shard fails 2/242 on this branch:

• cmd/bit/serve_http_wbtest.mbt:41 "serve --http: a repository resolves under the root": http_resolve_repo("/srv/git", "/acme/app.git") returns /srv/git/acme/app.git, the test expects the .git suffix stripped (/srv/git/acme/app).
• cmd/bit/grep_wbtest.mbt:1481 "grep: command path cases stay compatible": git grep --column -n -o reports column 14 for the second match on a line where git reports 13.

The last CI run on main (run 34554739985, the merge of #178) has the identical failure in the same shard (Total tests: 242, passed: 240, failed: 2, same diff), on the old toolchain. Neither test touches code this PR changes, and no fix exists on any branch yet, so I am not widening this PR with one. Happy to open a separate fix if wanted.

---

Generated by Claude Code

[mizchi]

Update on nix-build: flake.lock is now refreshed on this branch (8930986). moon-registry points at today's mooncakes index and moonbit-overlay at 86fff168 (ships v0.10.13+cbb11c36f), with the overlay's own nixpkgs/treefmt-nix pins mirrored. A full local nix build is still not possible here (GitHub tarball downloads are blocked in the sandbox), so the CI nix-build run on this head is the verification.

---

Generated by Claude Code

[mizchi]

The two cmd-native-test (cmd-bit-1) failures are fixed on this branch (93e8778, 44b919f), so the earlier "left for a separate fix" note above no longer applies. Both were real, both pre-date this branch, and neither was a flake.

grep: command path cases stay compatible — git grep --column -o prints one line per match, each carrying the column where that match starts. bit summed the ends of the preceding matches instead, which double-counts the gap in front of every match after the first: on foo_mmap bar mmap the second match starts at column 14 and bit printed 13. Upstream git had the identical bug and fixed it in "grep: fix --column --only-match for 2nd and later matches" (René Scharfe, 2026-04), shipped in 2.55 — the release third_party/git pins — computing cno = bol - start + match.rm_so + 1. The test shells out to the system git, so its expectation had been captured from a pre-fix git and the shard started failing when the runners moved to 2.55. Both bit and the expectation now report the true column.

Verified locally with the native debug binary, GIT_CONFIG_* unset so the shim does not delegate to the system git (2.43 here, still pre-fix):

                        before        after / git 2.55
foo_mmap bar mmap       5, 13         5, 14
abmmapXmmap             3, 9          3, 8

serve --http: a repository resolves under the root — the expectation asserted that http_resolve_repo strips a .git suffix; it never did. The resolver hands the path to normalize_repo_root, which folds a trailing /.git (a working copy's git directory) to the work tree and leaves <name>.git alone — deliberately, since a bare repository is the directory acme/app.git, bit upload-pack/receive-pack/upload-archive read the suffix the same way, and docs/scalable-git-server.md records the route as verbatim. Neither function has changed since the test landed in dd09350, so this case had never passed. The expectation now matches the listener, plus the /<repo>/.git case that normalize_repo_root does fold.

moon check --deny-warn --warn-list=-27-20-82 --target native passes; CI on this head is the check for the rest.

---

Generated by Claude Code