@@ -59,7 +59,7 @@ A task's checks are active from the moment it lands, as listed in the design's t
 | T1.26b Sampling | 1b | done | Landed at c52f280; candidate run 38191557540 green at 195624e; landing runs 38193268457 (lane) and 38193268537 (phase) green; `BASE_SHA` and `CHECKPOINT_SHA` 30d27a2; both approve at 195624e (3 review rounds, 2 fix rounds) | Contract version 16 (E64): the subgraph walked as the reference's SQL sampler walks it (`sdm-core`'s `subgraph`: foreign keys looked up by key and example, the tables referring read up to each anchor, asked together; `deeper_hop` and `parent_lookup` lifted), a text feature sent as the reference's words (`sdm_engine::words`, held by `words.json`; `text_column` lifted), both `neighbors` forms held on DuckDB to the reference's rows by L1 vectors (`neighbors.json`, ORD-2, the zoned case per DT-13); the sampler's requests and null order refused before any fetch (T1.26a's review P3); a null where the walk tells rows apart read as no link (P1B-13, the owner's decision (b), E66); E52's state value (`SamplerState`, left done or failed, handed back; H9's counts on a second call 6 then 4 on `python-core`, the old route and `rust-host`), the coverage warning once in a process in `sdm-host` (`context_coverage` lifted); the ledger's two rows gone, the bridge's changes declared moved; the AdventureWorks map's T1.26b entries gone, its journey reaching `predict` without `inference_config` on both hosts; 112 generated sources against the reference; the gate passing with no transition; as built in [as-built.md](as-built.md#t126b) |
 | T1.27 Labels | 1b | done | Landed at 81001a8 (part 1 at fed8174, the freeze's bug fix (E56(a)): an `allowance` declaration the merge base already holds, of an entry unchanged against it, is spent and tolerated as a spent `moved` one is); candidate run 38302822919 green at 6de5e11; `BASE_SHA` and `CHECKPOINT_SHA` 7314f1e; both approve at 6de5e11 (4 review rounds, 3 fix rounds). Measured before the push: both hosts' label tests (`sdm-testkit`'s `tests/labels.rs`, the client's `tests_test_ops/test_the_labels.py`) hold `base.regression`, `base.binary`, `base.multiclass` and `decimal.target-38-18` to the reference's context exactly; `sdm-engine`'s `tests/labels.rs` holds 5,601 vectors recorded from the reference's executor and pandas; the three seeded faults (Kahan removed, the month-end clamp flipped, `Gt` flipped) each fail tests; `contract/tools/differential.py` over 100 sources (seeds 27 and 28): no UNCLASSIFIED, 54 agree, 46 proposed under P1B-3 with the core's labels and `_by_time` seeds the reference's; the maps rewritten from both runners. Reach: AdventureWorks unchanged (`inference_config`, T1.28a, then `predict`'s result); `base.regression` to `predict`'s result; `base.binary` and `base.multiclass` to `classification_task` (T1.28a) after their labels. Decision rows E72 (T1.27) and E73 (the owner's decisions of 2026-10-12); register P1B-3 rewritten, P1B-18, P1B-19 and SMP-1 to SMP-3 proposed, P1B-1's and P1B-13's notes; fix rounds 3 (1: the differential's edits and evidence, a time column of an identifier refused as the reference refuses it; 2: a decimal average no longer refused as the caller's error; 3: a decimal sum's total alone checked, its terms added wrapping; see [Fix-round histories](fix-round-history.md)) | In Rust over fetched Arrow, compared exactly (E19); a tolerance rule goes to the owner only if exactness proves costly. Scope from T1.21b (E51(e)): the refusal map has no T1.27 entry, so, from T1.21e (E57), the family maps' T1.27 entries (`static_target`, with `base.binary`'s condition target, `COUNT(*)`, `orders.*` and boolean label masked behind it; `target_type`); a rule no map needs stays refused (P1B); from T1.24b's review, it moves T1.24b's eight edit cases from `contract/fixtures/edit-cases` into `contract/cases`, with their reference lines into the reference file, once its checkpoint's case loader takes the format's new edits, and takes T1.24b's code P3s (a time column given as `np.str_` kept as given; a cross-host case keying a VARCHAR time column) |
 | T1.28a Wire and payload rules | 1b | done | Landed at 447af89; candidate run 38331955521 green at 180b272; `BASE_SHA` and `CHECKPOINT_SHA` 8c06176 (2 review rounds, 1 fix round). Measured before the push: the test author's 46 python-core and 4 rust-host cases pass; `contract/tools/differential.py` over 250 sources with no UNCLASSIFIED. Decision row E76; contract version 19; register P1B-3 and P1B-4 rewritten, P1B-19 settled, P1B-20, P1B-21 and DT-18 proposed; the evidence as it stood is in [fix-round-history.md](fix-round-history.md) ("T1.28a, its task row as it stood at T1.28b.") | From checkpoint 7 (E71): the payload rules only (context leakage, statistics, mode, limits and notices, the seven ledger rows, the classification payload exit), through T1.28w's types, on lane 1b after T1.27, starting only after T1.28w lands. Split from T1.28 at T1.12b (checkpoint 4, subagent P3); own landing and stop-rule counter row: `sdm-wire` and the payload's rules in `sdm-request` (the context drop, the statistic, the limits, the run mode, the notices; seven ledger rows). Scope from T1.21b (E51(e)): the refusal map's `inference_config` and its gap plus, from T1.21e (E57), the family maps' T1.28a entries and gap (`wire_dtype`; the empty indices, the reference's own refusal in the host) and the classification payload masked behind `static_target`; a rule no map needs stays refused (P1B); the `StatePayload` lock derived from the typed `sdm-wire` structs, or at the owner's option by its version only (E52); carried from T1.27's review (subagent P3-1, P3-4): a static boolean target the reference takes as multiclass sends its labels as the strings 'False' and 'True' (its classes `['False', 'True']`, dtype string), where the `labels` operation gives booleans; `AVG(...) = NULL` and `!= NULL`, which the reference fails INTERNAL_ERROR and the core labels and refuses (`classification_task`), stays a named refusal when the classification layout is lifted; carried from T1.27's fix round 3: the reference's decimal sum fails at the column's declared precision (DECIMAL(4, 0): 9999 + 1 refused INVALID_REQUEST), where the core's fails only past 38 digits, whatever the precision, a difference the decimal feature's `wire_dtype` refusal hides today and which surfaces when a decimal feature has its wire dtype |
-| T1.28b Predict-side rules and the result | 1b | in progress | Started on lane `rust-core-phase-1-lane-1b` at 447af89 (T1.28a's landing); in two landings: part 1, the guard change alone (E56(d), E78(a) to (d)), rebased onto T1.15's landing (758e16f) and landed at 75620f8 (candidate 207cd9d; `BASE_SHA` and `CHECKPOINT_SHA` 758e16f); part 2 from 75620f8, rebased onto T1.6c's landing, `BASE_SHA` and `CHECKPOINT_SHA` 09b0d49: the predict-side rules, the result's layout, the boundary-vector matrix and the freeze's baseline in format 3; its acceptance tests from the independent test author cherry-picked unchanged (`test_t128b_predict_result.py`, `t128b_predict_result.rs`, and the amendment building `Predict` from its named fields); candidate complete, for review; lands by its landing commit, whose SHA and CI run the next package records (E25). Fix round 1 after the review of b7d29b9 (Codex: one P1, two P2s; the subagent: Codex's P1, with P3s): integer predictions kept, a missing one null, probability text as `float` reads it, the P3s folded in (E78(o)); the test author's 8 approved vectors moved to `divergence` and 34 cases recorded for the findings. Measured before fix round 1's push: the test author's python-core cases 144 of 144; rust-host 4 of 6 tests, the two failing on `absent.prediction-named` and `boundary.evidence.prediction-null`, recorded before the reference's items were, so held to the frame's NaN where the item is `None` and the core's cell now null, for the test author; `gate.sh --base 09b0d49c --run python-core,python-old,rust-host` (in the push's text); `contract/tests` 348 passed, the 9 of T1.15's gate-mode tests failing on this machine as at 758e16f; the client's tests 2,473 passed; `cargo test --workspace` passing but those two; `contract/tools/differential.py` over 240 sources (seeds 27 to 30): no UNCLASSIFIED, 209 agree, 24 approved P1B-11, 6 proposed P1B-3, 1 proposed DT-18; the protected paths 25 by the merge base's matcher; the freeze's self-tests 140; part 1's copy of the guards passes against 09b0d49 and as landed
+| T1.28b Predict-side rules and the result | 1b | in progress | Started on lane `rust-core-phase-1-lane-1b` at 447af89 (T1.28a's landing); in two landings: part 1, the guard change alone (E56(d), E78(a) to (d)), rebased onto T1.15's landing (758e16f) and landed at 75620f8 (candidate 207cd9d; `BASE_SHA` and `CHECKPOINT_SHA` 758e16f); part 2 from 75620f8, rebased onto T1.6c's landing, `BASE_SHA` and `CHECKPOINT_SHA` 09b0d49: the predict-side rules, the result's layout, the boundary-vector matrix and the freeze's baseline in format 3; its acceptance tests from the independent test author cherry-picked unchanged (`test_t128b_predict_result.py`, `t128b_predict_result.rs`, and the amendment building `Predict` from its named fields); candidate complete, for review; lands by its landing commit, whose SHA and CI run the next package records (E25). Fix round 1 after the review of b7d29b9 (Codex: one P1, two P2s; the subagent: Codex's P1, with P3s): integer predictions kept, a missing one null, probability text as `float` reads it, the P3s folded in (E78(o)); the test author's 8 approved vectors moved to `divergence` and 34 cases recorded for the findings. Measured before fix round 1's push: the test author's python-core cases 144 of 144; rust-host 4 of 6 tests, the two failing on `absent.prediction-named` and `boundary.evidence.prediction-null`, recorded before the reference's items were, so held to the frame's NaN where the item is `None` and the core's cell now null, for the test author; `gate.sh --base 09b0d49c --run python-core,python-old,rust-host` passes at f909e41, judged by 09b0d49's harness (observations, ratchet, compare, l0, coverage), the 25 protected paths named; `l0_scan.py`'s vocabulary check flagged the numeric-text reader's first name, so it is named for what it reads (`numeric_text`), whether the port belongs on the emulation allowlist (H21) left to the owner; `contract/tests` 348 passed, the 9 of T1.15's gate-mode tests failing on this machine as at 758e16f; the client's tests 2,473 passed; `cargo test --workspace` passing but those two; `contract/tools/differential.py` over 240 sources (seeds 27 to 30): no UNCLASSIFIED, 209 agree, 24 approved P1B-11, 6 proposed P1B-3, 1 proposed DT-18; the protected paths 25 by the merge base's matcher; the freeze's self-tests 140; part 1's copy of the guards passes against 09b0d49 and as landed
 | T1.29 Baseline cases | 1b | not started | | `base.regression` joins the acceptance corpus here at the latest; owner check-in after it (H12); E27: an AdventureWorks-style relational dataset in DuckDB is one of the four baseline cases, in addition to the 40-case inventory (E31), its reference observation recorded on `python-ref`, and gates T1.29. From checkpoint 4 (T1.12b): joint acceptance, not first discovery; T1.21e adds the inventory cases and baselines, and every package from T1.23 to T1.28 reruns the unstripped AdventureWorks journey, the baselines and the family maps; the pin stays until here under T1.11f's condition |
 | T1.30 DuckDB SQL goldens (L4 part) | 1b | folded into T1.32a to T1.32e | | Folded by the owner (E53(c)): each family adds the SQL goldens for its own statements |
 | T1.31 Arrow-to-pandas presentation (L5 part) | 1b | folded into T1.32a to T1.32e | | Folded by the owner 2026-10-11 (E66(f)), as T1.30 was: each family adds the presentation table's rows for its own types |