P2p: mapping inputs reach the CLI as temporary files (#6), and release 0.1.1

[npstorey]

PRODUCE-PY P2p: mapping inputs reach the CLI as temporary files, and release 0.1.1

Part of npstorey/typedstandards#135. G0 record: 5985856524, cards D14 = A and D17 = A. Fixes #6.

Head: 19eb4544de67ad3dda92def86fb153018118bfc4. Base: main 7513583.

Why

0.1.0's verify removes a served bundle's top-level trustRegistry (D9, npstorey/typedstandards#136)
and sends the result to the CLI on standard input. CLI 0.2.0 reads --input - with
readFileSync(0), which fails with EAGAIN once the document is larger than a pipe buffer holds
(npstorey/typedstandards#138). So any served bundle above that size failed with UsageError
(exit 2). A bundle that signs a notebook inline is above it. The same applied to any large mapping
given to sign, withdraw or attest.

Commits

Commit	What	By
7c1d8f0	tests only: the red. src/ is identical to 7513583	impl (Opus 5.5)
b956475	the fix: a mapping input is written to a temporary file, removed when the call returns or raises; run always gives the child stdin=DEVNULL; README transport sentences; CHANGELOG entry	impl (Opus 5.5)
cb7ca9a	CLAUDE.md: names scripts/publish.sh and its modes; says who pushes the rollback tags; says a branch that changes .gitleaks.toml is pushed from its own worktree (D17)	ORCH
19eb454	Release 0.1.1: __version__ 0.1.1, CHANGELOG dated 2026-10-05 (the publish day the owner named); uv.lock unchanged (the version is dynamic)	ORCH

$ git diff --numstat main...19eb454
8	0	CHANGELOG.md
11	3	CLAUDE.md
5	3	README.md
1	1	src/typedstandards/__init__.py
8	6	src/typedstandards/_cli.py
38	29	src/typedstandards/_commands.py
24	2	tests/conftest.py
6	3	tests/test_commands.py
12	2	tests/test_d9.py
2	0	tests/test_guards.py
171	0	tests/test_large_inputs.py

Blast zone: _commands.py, _cli.py, tests/, the README's transport sentences, CHANGELOG.md,
CLAUDE.md and __version__. uv.lock, .github/, scripts/, package*.json, the vendored
CLI and tests/fixtures/ are untouched.

Acceptance: red, then green

Red at 7c1d8f0. Driven by the impl and reproduced by the ORCH in a separate worktree on macOS,
Node 24.21.0, Python 3.12. The impl also reproduced it in a Debian bookworm container on Node
24.21.0. Result: 8 failed, 4 passed. Every failure is at the EAGAIN error or at its assertion:

E   typedstandards.errors.UsageError: typedstandards verify exited 2: typedstandards verify: --input - cannot be read: EAGAIN: resource temporarily unavailable, read
E   typedstandards.errors.UsageError: typedstandards sign exited 2: typedstandards sign: --input - cannot be read: EAGAIN: resource temporarily unavailable, read
E   AssertionError: the document reached the CLI on standard input
FAILED tests/test_large_inputs.py::test_verify_a_large_served_bundle_as_a_mapping
FAILED tests/test_large_inputs.py::test_verify_a_large_served_bundle_as_a_path
FAILED tests/test_large_inputs.py::test_sign_from_a_large_input_mapping
FAILED tests/test_large_inputs.py::test_no_mapping_input_reaches_the_cli_on_a_pipe
FAILED tests/test_large_inputs.py::test_temporary_files_are_removed_when_the_cli_fails
FAILED tests/test_d9.py::test_verify_drops_only_trust_registry
FAILED tests/test_d9.py::test_verify_drops_it_from_a_path_too
FAILED tests/test_commands.py::test_mapping_input_goes_in_a_temporary_file

Green at b956475 and again at 19eb454:

1. A served bundle over 2 MiB verifies, given as a mapping and as a path. The test signs a
   UTF-8 file over 2 MiB inline under a fresh seed, runs view, and inlines a trustRegistry in
   host-core's shape.
2. Signing from an input mapping over 2 MiB succeeds.
3. No input travels on a pipe. For sign, withdraw, attest and verify given mappings,
   --input names a file, the child's stdin is DEVNULL, and the file is gone after the call. It
   is also gone when the CLI exits 2, or exits 1 for verify.
4. D9 stays pinned. The CLI receives the bundle minus trustRegistry, with key order kept and
   the caller's mapping unchanged. test_the_cli_itself_refuses_the_key is unchanged and passes.
5. The guards pass, and still fail on a planted hashlib, a seed read and env= (the impl's
   mutation run: 5 failed). One check was added: the run-time guard now also fails if the seed
   appears in an input file.
6. The suite, lint and wheel pass: 243 passed (238 on main plus 5).

Cell	Who ran it	Result
py3.11, node 24.21.0	impl	243 passed
py3.12, node 24.21.0	impl, ORCH	243 passed
py3.14, node 24.21.0	impl, ORCH	243 passed
py3.12, node 22.23.1	impl	243 passed
py3.11, node 22.23.1	ORCH	243 passed
Linux container, py3.12, node 24.21.0	impl	243 passed

ruff check and ruff format --check are clean. The wheel job (clean clone, uv build, wheel in
a fresh environment, smoke_wheel.py) passed.

Release dry run

DRY_RUN=1 RELEASE_REF=HEAD scripts/publish.sh from a clean clone of 19eb454. It built and
checked both files, smoke-tested the wheel (typedstandards 0.1.1, CLI_VERSION 0.2.0, 232
vendored files), and ran uv publish --dry-run. PyPI still answers 404 for 0.1.1.

0d23dc9906ec2261d2264c492c8e1bcff43c299df9c1aa99fea587343ff499b2  typedstandards-0.1.1.tar.gz
55c650135c4531b628aa174f3db9919e0a74f9f81c7bc12f505206112990dcf3  typedstandards-0.1.1-py3-none-any.whl

The owner reruns the dry run from a clean checkout of main at the merge commit, then publishes
on 2026-10-05. publish.sh refuses the live run on any other local date.

gitleaks, sign-off, signatures

$ gitleaks git --log-opts="main..19eb454" --no-banner
INF no leaks found
$ git log --format='%h %G? %ae | %(trailers:key=Signed-off-by,valueonly)' main..19eb454
19eb454 G npstorey@users.noreply.github.com | Nathan Storey <npstorey@users.noreply.github.com>
cb7ca9a G npstorey@users.noreply.github.com | Nathan Storey <npstorey@users.noreply.github.com>
b956475 G npstorey@users.noreply.github.com | Nathan Storey <npstorey@users.noreply.github.com>
7c1d8f0 G npstorey@users.noreply.github.com | Nathan Storey <npstorey@users.noreply.github.com>

Fixtures

None added or changed. The large bundle is built at test time from a fresh random seed. The pinned
first-note.bundle.json check (cb11d2a2…) is unchanged.

Model

The impl ran on Opus 5.5: every turn in its metadata reads claude-opus-5-5, at its agent file's
effort: high. The ORCH, Opus 5.5 at xhigh, wrote cb7ca9a and 19eb454.

Flagged, not fixed

• Gap in the static seed scanner. tests/guards.py seed_references misses a name built with
  +. Only the run-time guard catches it. Out of scope here; for the cold read.
• README wording. One sentence names CLI 0.2.0's pipe failure. It needs revisiting when the
  pin moves to a CLI that fixes cli: --input - fails with EAGAIN on a large piped input (readFileSync(0)) typedstandards#138.

🤖 Generated with Claude Code

https://claude.ai/code/session_012qX8dRbkzJr3B24DbHS3wM