fix: queue heredoc bodies per command line and drain in opener order

[378-kaiabot]

Note

Automated pull request by Pi, model aperture/synthetic/hf:zai-org/GLM-5.3-Flash. Implemented and validated empirically against aliou/sh on branch fix/heredoc-statement-boundary. See the local run log for the full investigation.

Summary

Second layer of the heredoc statement-boundary work, stacked on #17. Two failure groups remained after that fix — both empirically confirmed throwing Expected a command word on its tip:

1. Multiple heredocs on one command — cat << A << B with bodies foo/A/bar/B throws; the parser only lets a redirect consume a heredoc-body token that directly follows its opener, so the first body token sits between the second redirect and its target, and bodies bind to the wrong redirects.
2. A separator after a heredoc opener on the same line — cat << EOF; echo hi, cat << EOF && echo done, cat << EOF | grep foo: nothing on the right of the opener consumes the body token queued after the newline, so the body surfaces where a command is expected and throws.

Both are valid bash and are covered upstream by mvdan/sh's syntax/filetests_test.go (e.g. <<EOF1 <<EOF2 | c && d, foo <<EOF && { ... }, f1 <<EOF1; f2 <<EOF2). Several failing tests here are adapted from those cases.

Changes

• src/parser/heredoc.test.ts — commit 1/2: nine failing tests covering the cases above plus in-compound variants. Do not modify these to make the fix pass.
• src/parser/parser.ts — commit 2/2: the fix. parseRedirect queues heredoc redirects in pendingHeredocs instead of looking for the body inline; a drainPendingHeredocs step in skipSeparators/skipCaseSeparators attaches body tokens to their redirects in opener order as they surface (mirroring mvdan/sh's per-line heredoc queue). Orphan body tokens with no pending redirect are left in the stream and still error, as before.
• .changeset/legal-owls-check.md — patch changeset.

Behavior notes:

• Redirect-only statements like <<EOF1 <<EOF2 now parse with bodies attached instead of throwing (same shape as mvdan/sh's redirect-only statements).
• Verified empirically that bodies bind correctly when both sides of &&/| open heredocs (a <<A && b <<B: a gets x\n, b gets y\n).

Verification

• pnpm typecheck ✅
• pnpm lint ✅
• pnpm test ✅ (290 passed; the 9 new tests fail on fix: stop heredoc commands from swallowing the following statement #17's tip and pass here; all 281 prior tests untouched)
• Adversarial probes beyond the test file: heredocs across &&/|, multi-heredoc on a pipeline's left, case items, command substitution, EOF-terminated delimiters, empty bodies, comments after delimiters — all parse with correct body assignment.

Known limit (pre-existing, unchanged): redirects after compound commands (while ...; done <<IN) still don't attach to the compound — that needs compound-command redirect support first.

Stacked on #17

---

Run host: sh-d9e31b · Session: 01a062c1-c9e1-71cb-b132-aafe96837313 · Model: aperture/synthetic/hf:zai-org/GLM-5.3-Flash (validation), aperture/neuralwatt/kimi-k3 (implementation)