EVIDENCE — jsonpatch (RFC 6901 pointer + RFC 6902 patch engine)
Date: 2026-10-21. Every number below is tool output, not an estimate.

ORACLE
  Vendored upstream suite, fetched 2026-10-21 from
    https://raw.githubusercontent.com/json-patch/json-patch-tests/master/tests.json
    https://raw.githubusercontent.com/json-patch/json-patch-tests/master/spec_tests.json
  sha256 pins (checked by stage 1 of the battery on every run):
    tests.json        18707 bytes  de3dce3d0d5029fed83007e50b54607750dd3d1478d3c59ca35fdc18fb1a04ae
    spec_tests.json    4031 bytes  a26b050292207033e5cccc5d6102b7bd6f8add7db0d0680e5d46a7ecf40a8c7b
  95 + 17 = 112 cases; 4 are marked "disabled" upstream and are skipped.

COMMANDS AND RESULTS (run from /root/troupe-standup/wizard-work/null/jsonpatch)

1. Full battery, teeth included:
     $ python3 test_jsonptr_patch.py
     official suite: 108 passed, 0 failed, 4 skipped (disabled)
     == 54 passed, 0 failed ==            (engine selftest: RFC 6902 App. A + RFC 6901 + equality)
     [atomicity] 118 checks, 0 failed     (doc untouched after a failed patch; input never mutated on success)
     fuzz: 2000 patches, 823 ok, 1177 rejected
     fuzz: identity 402 ok / 0 failed / 98 skipped
     fuzz: move out and back is identity   -> pass
     fuzz: 3000 bad patches rejected as expected
     fuzz: test op 439 accepted / 393 rejected / 0 wrong
     teeth: 8 killed, 0 survived
     TOTAL: 126 checks, 0 failed
     VERDICT: PASS
   exit code 0.

2. Same battery from the wizard-outputs copy (the deliverable path):
     $ python3 /root/.hermes/wizard-outputs/null/test_jsonptr_patch.py
   exit code 0. (This is the CHECK line.)

3. Repo copy, teeth skipped:
     $ python3 test_jsonptr_patch.py --no-teeth
     TOTAL: 125 checks, 0 failed / VERDICT: PASS

4. CLI smoke test (measured):
     $ python3 jsonptr_patch.py --doc cli_doc.json --patch cli_patch.json
     {"foo": ["bar", "qux", "baz"], "obj": {"k": "~x/y", "n": 1}, "tail": "qux"}   exit 0
     $ python3 jsonptr_patch.py --doc cli_doc.json --patch <failing test op>
     error: patch[0] ('test'): test failed at '/n': expected '999', got '1'          exit 1
     $ python3 jsonptr_patch.py --nope
     usage: ...                                                                      exit 2
     $ python3 jsonptr_patch.py --selftest | tail -1   ->  "== 54 passed, 0 failed =="  exit 0
   Also measured: `copy` with `from: "/foo/-"` is rejected with exit 1 — "-" is an
   add-only token, so the engine refuses it (correct per RFC 6901 §4).

5. Mutation teeth: 8 source mutants were generated by single-string edits of
   jsonptr_patch.py and the whole battery was re-run against each. All 8 were killed
   (battery exited non-zero), so the suite has teeth:
     add index==len allowed ............ killed
     missing parent tolerated .......... killed
     bool is a number .................. killed
     unescape ~0 wrong ................. killed
     move into descendant allowed ...... killed
     pointer slash not enforced ........ killed
     leading zeros allowed ............. killed
     remove root tolerated ............. killed

WHAT IS MEASURED vs CLAIMED
  Measured: everything above (exit codes, counts, sha256s).
  Claimed (spec reading, not machine-checked): that RFC 6902 §5's "the entire patch
  document MUST NOT be applied" reading matches this implementation's atomic
  behaviour. The vendored suite has no multi-op error case that pins this down, so
  the atomicity stage asserts it directly instead of relying on the suite.

KNOWN LIMITATIONS
  * `test` equality is implemented structurally (1 == 1.0); the official suite
    contains no 1-vs-1.0 case, so that behaviour is asserted only by the engine's
    own selftest.
  * Duplicate keys in a patch object (upstream case A.13) are marked "disabled"
    upstream and are skipped; Python's json.loads keeps the last one, so a duplicate
    `op` is not detectable here.
  * `disabled` cases: 4 skipped, by upstream's own marking.