fix(#3841): parse the identity payload before consulting the probe exit code

`classifyIdentityProbe` short-circuited on a non-zero exit BEFORE it looked at
stdout, so a tool that had already proved its identity and merely exited
non-zero was classified `no_identity_verb`. The launcher preamble that this
module speaks for reads stdout only -- its command substitution discards the
status -- so the two surfaces disagreed on exactly that input: shell `ok`,
classifier `no_identity_verb`. Measured against the real snippet, not inferred.

That matters because the classifier is the engine for the announced hard-fail
phase and has no production caller yet. An install verified by today's warn
phase would have been refused the moment hard-fail landed, in the phase where
that stops the run rather than printing a line. Nothing recorded or tested the
difference.

The classifier moves rather than the shell: the shell is the shipped path with
observable dependents, the classifier has none. The predecessor defence is
untouched -- a usage screen yields no usable payload, so it still falls through
to the exit-code branch.

Adds the cross-surface parity test the gauntlet requires for two surfaces
implementing one decision: ten probe behaviors driven through both the real
snippet and the classifier, asserting they agree with each other.

Surfaced by re-running the feature-implementation directive's design and QA
steps against the code merged in #3848, which shipped without them.

Refs #3841

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.changeset/sturdy-jaguars-munch.md

@@ -0,0 +1,5 @@
+---
+type: Fixed
+pr: 0
+---
+**The identity classifier now agrees with the launcher preamble about a tool that proves itself and then exits non-zero** — the two surfaces implement one decision and disagreed on exactly that input, so the announced hard-fail rollout would have refused an install the current warn phase verifies. (#3841)