ci: second review pass — recurse examples, reject duplicate exceptions

Six more findings from CodeRabbit and Copilot on #934. All valid; four were
my own docs lagging the write-checking change in d0626f5.

check_forward() now discovers example packs recursively, so a pack nested
under examples/<group>/ can't slip past the "every example pack" contract.
Taken WITHOUT the suggested is_file() filter, which would have broken it: a
feedpak is dual-form — a zip (foo.feedpak) or a directory (foo.feedpak/) —
and the spec's own examples ship as directories, so is_file() would have
matched zero packs. Suffix matching covers both forms.

load_exceptions() rejects duplicate keys instead of silently keeping the
last one, which would quietly retarget the tracking issue for a piece of
debt this file exists to track.

The sloppak import is wrapped so a missing dependency produces a CI-legible
::error:: rather than a bare traceback.

Docs caught up with the code: the exceptions file header, its stale-entry
rule, and the changelog all said "reads" when the gate checks reads AND
writes.

Signed-off-by: topkoa <topkoa@gmail.com>

feedpak-spec-exceptions.yml

@@ -1,4 +1,9 @@
-# Manifest keys core reads that the feedpak spec does not (yet) define.
+# Manifest keys core reads OR WRITES that the feedpak spec does not (yet) define.
+#
+# Both directions are gated, and both are allowlisted here. A key core *writes*
+# (`manifest["k"] = v`) is spec surface pointed outward — it lands in every pack
+# we emit — so an undeclared one needs an entry here just as much as one core
+# reads.
 #
 # This file exists so the spec-conformance gate (tools/check_spec_conformance.py)
 # can be honest instead of being switched off. A gate with no legitimate way to
@@ -10,8 +15,10 @@
 #   - Every entry needs a tracking issue. No issue, no exception.
 #   - Entries are debt, not policy. The fix is to land the key in the spec
 #     (github.com/got-feedback/feedpak-spec) and delete the entry.
+#   - One entry per key. A duplicate is an error, not a last-one-wins.
 #   - The gate fails if an entry goes stale — i.e. the spec caught up, or core
-#     stopped reading the key. The allowlist must never become a hiding place.
+#     no longer reads or writes the key. The allowlist must never become a
+#     hiding place.
 #
 # For a key that is genuinely experimental and not yet ready for the spec,
 # prefer the reserved `x-` prefix (e.g. `x-my_new_key`) over an exception: the