[casselc]

Emit one last chunk for an empty known-length body

Summary

Make jolt.http.body/chunk-writer emit exactly one HTTP last-chunk when an
explicitly chunked, known-length response body has length zero.

before: 0\r\n\r\n0\r\n\r\n
after:  0\r\n\r\n

The ordinary three-stage writer emits a chunk header, body, chunk terminator,
and final last-chunk. At length zero, the chunk header and its following CRLF
already form a last-chunk, so continuing to the normal terminator emits a second
one. RFC 9112 permits exactly one last-chunk before the trailer section.

Fix

Use the existing five-byte empty-chunk writer for length zero. Positive body
lengths retain the existing partial-buffer state machine unchanged.

Add a deterministic regression that checks output capacities 1 through 5, so
the last-chunk must remain exact across every partial-copy boundary. Correct the
existing generated property's zero-length oracle, which previously repeated the
implementation's faulty header/body/end formula, and retain the independent
HTTP response decoder property.

Red to green evidence

Base: casselc/jolt-http@fb596ea1a7d1899fd437d21c9da4fbfab0e436a2

The deterministic test fails on the unchanged base at every capacity with ten
wire bytes instead of five. After the fix, that test and both chunk-writer
properties pass.

Validation

• focused deterministic framing regression and two chunk-writer properties:
  PASS
• full library gate on Jolt v0.8.1 aspect compiler: PASS, 338 checks / 0
  failures
• all Jolt commands ran through Chez Scheme 10.4.1

Provenance

Tracked as
jolt-aspect-packs#80.
The reproducer and fix require no aspect compiler or instrumentation.

CI propagation repair

The first hosted run exposed a stale compiler pin rather than a body-framing failure. Both POSIX jobs used c666a2d0175923cb7edeb36fb99c7e2c657af375, which predates the Jolt 0.8 scalar jolt.ffi/write call shape required by the already-pinned transport graph. The suite consequently produced cascading wrong-arity failures across both platforms before exercising the PR honestly.

The workflow now performs a full-history checkout of canonical aspect compiler 73b801e73e6441a2912fbc9ce22c529bd226c6f1 and fails closed unless its provenance lock derives jolt v0.8.1-34-g73b801e7. That compiler tip independently passed its tests, flake, and aspect-provenance workflows. Production HTTP and transport code are unchanged by this follow-up.

Updated exact head: 739728c929da6e4fd2715b6b393b9cce6d0ad2d0

Validation before push:

• actionlint: pass
• canonical version derivation from its full-history tree: v0.8.1-34-g73b801e7
• released Jolt v0.8.1 full suite through Chez 10.4.1: 338 checks, 0 failures
• exact-head hosted PR run: https://github.com/casselc/jolt-http/actions/runs/33982121496 — Linux and macOS passed