fix: attach trailing redirects to compound commands

[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-body-queue. See the local run log for the full investigation.

Summary

Third layer of the heredoc/redirect work, stacked on #18. PR #18's review surfaced a worse-than-throwing edge: redirects written after a compound command's closing keyword (done, fi, }, ), esac, ]]) were silently captured into a phantom word-less SimpleCommand statement instead of attaching to the compound node. Injury, error, and now insult: while read x; do echo $x; done < in | sort parsed < in as a redirect-only statement at the left of a new pipeline — the loop lost its stdin and the pipeline shape broke, with no error signal.

Bash and mvdan/sh attach these redirects to the compound node itself.

Changes

• src/ast.ts — optional redirects?: Redirect[] on Subshell, Block, IfClause, WhileClause, ForClause, SelectClause, FunctionDecl, CaseClause, TestClause, CoprocClause, CStyleLoop. Field stays absent when there are no redirects, so every existing fixture is untouched.
• src/parser/compound-redirects.test.ts — commit 1/2: fifteen failing tests (loops, if, block, subshell, case, select, function, [[ ]], coproc, heredoc-on-compound, multiple redirects, and the pipeline case). Do not modify these to make the fix pass.
• src/parser/parser.ts — commit 2/2: a shared parseTrailingRedirects runs after each compound's closing keyword and attaches redirects to that node. Two subtleties it handles: redirects after a function's block bind to the FunctionDecl (and coproc bodies to the CoprocClause), so capture is suppressed in the body scope and re-enabled for nested statement lists; and heredoc openers flow through the existing pendingHeredocs queue from fix: queue heredoc bodies per command line and drain in opener order #18, so done <<IN bodies still attach in order. time { a; } > out intentionally keeps binding to the timed block — TimeClause has no redirects field.
• src/test-helpers/ast-builders.ts — cStyleLoop now routes clause strings through the real arithmetic parser (the ArithLit placeholders never matched parser output).
• .changeset/fuzzy-hounds-tease.md — patch changeset.

Verification

• pnpm typecheck ✅
• pnpm lint ✅
• pnpm test ✅ (305 passed; the 15 new tests fail on fix: queue heredoc bodies per command line and drain in opener order #18's tip and pass here)
• Reviewer probes beyond the test file: nested loops inside function bodies, nested subshells, &&/background ordering, heredoc on a function, elif chains (redirect attaches once, to the outer if), unnamed coproc, dup-fd redirects (>&2), redirect-then-comment, position extents — all verified empirically.

Known limit (pre-existing, unchanged): time false > t2 still binds to the timed command, since TimeClause carries no redirects.

Stacked on #18

---

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)