@@ -220,6 +220,8 @@ The histories of landed packages moved verbatim to [fix-round-history.md](fix-ro
 
 **T1.24b.** E49 probes before the first push: (1) the route changes (every edit of a core-built schema is the core's `edit_schema`; `schema_edited` the core's), so an old-against-core differential over 102 generated DuckDB sources (52 and 50, two seeds, the scripts kept outside the repository as T1.26b's were), each with 2 to 7 edits (keys set, cleared and re-set as a name or a one-column tuple, time columns set and cleared, meanings as text and as a `Stype`, columns left out, relationships replaced, appended and popped, a missing column, a non-joinable link, a time column on a non-time type), then `materialize` and `predict` against the scripted endpoint, on the reference, the old route and the core: all 574 steps (218 refused) alike in outcome and public schema; `materialize` and `predict` the corrected reference's (P1B-13) but five `keyless_neighbors` (P1B-3) and one DT-12; one difference found and fixed (a key named on a link's own foreign key, which the reference's engine refuses; `check_links`), one registered (an unhashable value naming a column, P1B-16) and one new refusal registered (an edit after a change past the description, P1B-17); (2) the edit payload's readers (`edit_schema`'s three attachments, `Materialize.held`): 200, 5,000 and 100,000 levels and 100,000-wide lists and objects, each refused by name or read (`sdm-core`'s `tests/edit_schema.rs`); (3) the new comparators fail closed, seeded: the edit pin's check on both hosts (missing, empty, retyped, differing) and the core-agreement rule on a perturbed `edit_schema` input (by hand, `compare.py` naming `core.inputs`); a setter that skips the core fails the client's count of `edit_schema` runs; a core that took an invalid meaning fails `sdm-schema`'s tests (by hand). Rebased onto T1.18's landing (b455b94), then onto part 1's landing (de247a3, the freeze's reading fix, split out by the orchestrator so the gate could read this package's E66 allowances); the T1.26b history moved to [fix-round-history.md](fix-round-history.md).
 
+**T1.28w.** Part 1, the L0 guards' bug fix (E56(a)), measured before its push: `gate.sh --base 1471563e --run python-core,python-old,rust-host` passes in `ref-pandas3` with the protected paths named (observations, ratchet, compare with no differences, l0); protected paths changed (3): `l0_deps.py`, `l0_scan.py`, `selftest.py`. Seeds by hand, on real trees: the rename (T1.6b's kept-name commit reverted, which applies cleanly on top of part 1) fails under 1471563's guards (`sdm-pql` with no class, the edge, the allowlist's three paths) and passes under part 1's, its own copies passing too; the renamed tree with its guard data set back to `sdm-pql-interim` fails under part 1's guards; the post-landing state, part 1's tree judged against itself, passes under 1471563's guards (based at 1471563 and at itself) and under its own (based at itself); the renamed tree's `selftest.py --no-clippy` passes. `selftest.py` seeds the same on a fixture (14 cases, each guard: a control, the rename passing, crates or data alone, the copies missing, the name not a literal, the name of another crate); with the fix undone (each guard reading its own name) the rename and the five refusals fail. E49 probes: no route changed and no reader of outside JSON added (the guards read their own copies' Python as a syntax tree, a literal only); the reading fails closed on a copy missing, unparsed, not a literal, or naming another crate.
+
 ## Size-trend reviews
 
 The design's "Harness scope" asks the checkpoint reviews for the size trends; the status page was cut (E53(a)), so they are recorded here and nowhere else.