PRODUCE-PY P2f: the cold read's four fixes before 0.1.0

[npstorey]

This fixes four low findings from the cold read of P1 and P2 (npstorey/typedstandards#135, comment 5979743999): findings 1, 2, 3 and 4b. The owner ruled that these land before 0.1.0 is cut. No version heading is added and __version__ is unchanged; the release PR follows this one.

Branch: ts135/p2f-coldread-fixes · Head: 83c486853536777516b19da9741b71f18cd570fb · Base: main at 2a8ac5d
Size: git diff --numstat main...HEAD: 7 files, +246 −22.
Blast zone: this branch only. Files touched:

• tests/guards.py
• tests/test_guards.py
• tests/test_badge.py
• tests/test_readme.py (new)
• src/typedstandards/_badge.py
• README.md
• CHANGELOG.md

_cli.py, _commands.py, _node.py, errors.py and the CI workflow are unchanged.

The fixes

1. The digest guard catches re-import routes.
   • The static scanner now reports any route to a digest module outside pin.py: an imported alias (from .pin import hashlib), an attribute (pin.hashlib), a name, a string naming one, a sys.modules lookup that names one or uses a non-literal key, and import_module or __import__ with a non-literal name. The cold read's four routes are now in the offender tree.
   • A new run-time test records every hashlib constructor call by its calling frame while the five commands and every helper run. It fails on any typedstandards module other than pin; third-party frames are not counted. It also asserts that it saw pin's own call, so the recorder is shown to work.
   • hash_imports(PACKAGE, allow=frozenset()) still names only pin.py.
2. The README's Node sentence is now true. Only sign, withdraw, attest, view, verify, cli_version(), and show without a result run the CLI and need Node. pin, badge_cell, comparison_cell, sidecar and show(record, result) run without it. A test pins both halves, with TYPEDSTANDARDS_NODE pointed at a missing file. The code was already correct, so this test passes before and after; the fix itself is the text.
3. badge_cell accepts a host with a port. The date and time check now leaves out the URL's authority, so 192.168.1.10:8080 and records.example.org:8443 are accepted.
   • Measured on the way: the link carries the URL percent-encoded, which had hidden a time's colon. A time in the query (?t=12:30) or the fragment was accepted before this change, not refused.
   • The URL past its authority is now checked as written and percent-decoded, so those times are refused.
4. The README has no relative links. The CLAUDE.md link is now absolute, and a test fails on any relative Markdown link or image (#anchor links are allowed).

Acceptance

Each red below was driven locally.

#	Red	Green
1	The old scanner returned [] over the four routes, and the offender-tree test failed on the four missing files. Temporary offending calls in _sidecar.py failed both guards. One of them, a route with names built at run time, failed only the run-time test.	4 guard tests passed
2	Documentation fix; the grep of the callers of locate_node and run is in the phase report	test_the_helpers_need_no_node passed
3	https://192.168.1.10:8080/… was refused; a time in the query (raw and encoded) or the fragment was accepted	10 passed
4	The link test caught [CLAUDE.md](CLAUDE.md)	2 passed

Checks run locally on the head

From a clean clone at 83c4868, with UV_PYTHON set for each run:

• Python 3.11.15, Node 24.21.0: 238 passed
• Python 3.12.13, Node 24.21.0: 238 passed
• Python 3.14.6, Node 24.21.0: 238 passed
• Python 3.12.13, Node 22.23.1: 238 passed

ruff check and ruff format --check are clean. gitleaks git --log-opts="main..HEAD" scanned 3 commits and found no leaks. These runs were on macOS (arm64). CI has not run on this head yet: it runs once the branch is pushed.

There are three commits. Each is signed (G) and carries one Signed-off-by: Nathan Storey <npstorey@users.noreply.github.com>, equal to the author email.

Model: Claude Opus 5.5 (claude-opus-5-5).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ru4PYga7Zf8HANgotZKs4u

[npstorey]

GO — CAT PLAN (the program seat), bound to head 83c486853536777516b19da9741b71f18cd570fb.

Read from GitHub and disk, 2026-10-04:

• The PR API: head 83c4868…, base main 2a8ac5d (the P2 merge), 7 files, +246 −22, 3 commits, merge state clean; equal to the three-dot diff.
• Check runs on that exact head: 18, all success from GitHub Actions (15368), nine distinct names, the nine protect-main requires.
• Before the push: three commits, each signed (G) with one Signed-off-by read by the trailer parser; the rollback tag rollback/pre-produce-py-p2f annotated at 2a8ac5d; the guard's keyword scan over the 246 added lines found nothing, and gitleaks passed.
• The load-bearing change, _badge.py: the host and port are left out of the date-or-time check, so 192.168.1.10:8080 is no longer read as 10:80; the part of the URL after the host is checked as written and percent-decoded, which closes the ?t=12:30 case the implementer found at 2a8ac5d. The cold read's four findings (record 5979743999) are each addressed: the digest guard catches the re-import routes and records digests at run time, the README names the calls that need Node, and its links are absolute.

The owner merges with --match-head-commit 83c486853536777516b19da9741b71f18cd570fb through the seat's tested script, which also tags the merge. Next: P2r, the version cut dated 2026-10-04, the publish script's dry run from a clean clone, and the release PR.