fix: the Hypatia gate could never fire — the defects that made it unconditionally vacuous - #61

The Hypatia gate in this repo has never been able to fail

Static Analysis Gate is green here, and that green means nothing. Four defect classes, each
independently sufficient to make the gate vacuous. Measured in this repo: defects 1 and 4 are present and fixed here. Defects 2 and 3 were not present in this file — that code is already correct here, and is described below only to document the class.

1. 2>&1 folded the scan summary into the JSON payload

HYPATIA_FORMAT=json "$HOME/hypatia/hypatia-cli.sh" scan . > hypatia-findings.json 2>&1

Per Hypatia's own contract (hyperpolymath/hypatia, lib/hypatia/cli.ex:82-87) findings go to
stdout and a one-line summary always goes to stderr. Folding them together makes the file
invalid JSON, so jq empty fails, the guard concludes "the scan did not run", and [] is written.
Every count then reads 0 and Fail on critical findings cannot fire on any input.

Fixed: stderr stays on the log; --exit-zero is passed so exit 1 ("findings exist") is no longer
mistaken for a crash; the payload is validated with jq -e 'type == "array"'.

2. The availability probe tested for a directory that does not exist

if [ -d "$HOME/hypatia/scanner" ]; then

hyperpolymath/hypatia has no scanner/ directory, so this is unsatisfiable. The scan step was
skipped and a Create stub findings step wrote [] — a second, independent route to permanent
green, invisible at the check level because the check still reported success.

Fixed: probe $HOME/hypatia/mix.exs, which is what a successful clone actually leaves behind. The
"unavailable" notice is promoted from ::notice to ::error so a missing scanner is visible.

3. The clone used ${REPO_OWNER}, which 404s outside hyperpolymath

metadatastician/hypatia does not exist. In those repos the clone silently failed
(2>/dev/null || true), which is indistinguishable from "unavailable" — see defect 2.

Fixed: clone hyperpolymath/hypatia explicitly.

4. Every annotation said null, on a path GitHub cannot anchor

The jq emitted \(.message), but findings have no message key — the real keys are
action, file, line, reason, rule_module, severity, type. And .file is an absolute runner path.

Positive control on a real finding from the hybrid-automation-router artifact:

Fixed: .reason // .message // .type // "finding", and .file made workspace-relative with
ltrimstr($ws + "/"). The fallback chain means this is correct whether or not a message key is
ever added.

What this changes in practice

The gate can now fail. Threshold is unchanged and remains critical-only
(steps.scan.outputs.critical > 0); high/medium/low continue to annotate without blocking.

If this PR turns the gate red, that is the fix working — the finding was always there and the gate
could not report it. Do not merge a red one by overriding the gate. Either the finding is real
and wants fixing, or it is a false positive that wants filing upstream.

Provenance

Same four-defect repair, applied identically across every repo carrying this workflow. The transform
is a byte-exact block substitution with post-conditions asserting the defect is gone and the cure is
present; it refuses to write a file that fails any of them. Each post-condition is scoped to a live
shell construct, never to a comment, so the explanatory comments above cannot satisfy their own
assertions.