Audit round three: the traceback my last fix missed, and six smaller holes

[Tippel-AI]

Follows #34, #35 and #36. This closes the remaining confirmed findings.

The one that got past a previous fix

The README's own demo still ended in a stack trace on a base install.

#22 moved evaluation inside the guarded block and listed the exceptions it knew
about. SynthIdUnavailableError was not one of them — and the demo config
declares text_marking: synthid, so the single command the README sends a
newcomer to was precisely the case the fix did not reach. Exit 1, which in
this tool means a rule failed.

I found it by building a wheel and installing into a fresh virtualenv rather than
by reading. That is the only way this class of defect shows up, and it is why the
audit's first-run axis was worth having.

before:  SynthIdUnavailableError: text marking needs the optional extra…   exit 1
         (plus 2 tracebacks)

after:   Install the optional detector: pip install 'markproof[synthid]'    exit 2
         (0 tracebacks — and the extra survives, which it would not have
          before the Rich escaping in #34)

The three optional-dependency errors now share a base the CLI catches by type,
so the next optional dependency cannot reintroduce this.

Six smaller ones

The manifest that decides a document verdict was recorded nowhere. A reader
could re-fetch the document, confirm its digest, and still not know which manifest
produced the answer. The finding carries the sidecar digest and whether it came
from a header or an element.

MPF-M-001 passed without saying who signed. The rulepack allows self-signed
chains, so "validly signed" without "by whom" is half an answer.

A <link> in a comment or a script string was taken as the manifest pointer —
so anything that can put text on a page could choose the bytes that decide the
verdict.

NOTICE and DISCLAIMER both claimed no component is named after a third-party
mark. That is false — checks/synthid.py, checks/c2pa_verify.py, the
[synthid] extra and the c2pa-verify check type all are. The narrower
parenthetical was true and the sentence around it was not. Both now state what is
actually the case: the marks appear where they describe what the code does, which
is nominative use, and nowhere as a product name or branding. Renaming them would
make the code harder to read and the claim harder to check.

Nothing acknowledged that evidence can contain personal data. Probes record
responses verbatim and reports record the URLs probed.

The determinism gate asserted nothing about rule coverage. It checked that all
four verdict values appear — which eleven cases exercising the same two rules
would satisfy. Every shipped rule must now appear in some golden.

530 tests, ruff + mypy strict clean.

🤖 Generated with Claude Code