chore(generalization): Tier S housekeeping — smoke promotion, capability-index entries, promotion sweep (S5)

- Promote tests/test_register_substantive_consumption.py into the smoke
  suite (local core/cli_test.py + mirrored .github/workflows/smoke.yml):
  the falsifiable register-axis (ADR-0069/0071/0077) contract was red on
  main outside every gate for the 2026-07-22..24 window, masked by an
  unrelated flake label. 36 tests, ~4.6s.
- Found in passing: test_core_test_deductive_suite_expands_to_entailment_lane
  was already red on clean main (an exact-tuple pin from when the
  `deductive` suite had one file; it has since grown to 12). Fixed to a
  contains-style assertion so it can't go stale the same way again.
- Two new capability-index adapters: deduction_serve_existential_result
  (Band v6-EX alone — the one band among six that changes the decision
  procedure itself) and curriculum_serve_result. Breadth 11 -> 13; baseline
  re-frozen (a deliberate re-freeze per its own docstring).
- Promotion sweep: confirmed via the existing wrong=0 lane gate (run
  repeatedly this session across 166/166 deduction-serve cases) that no
  split beyond the already-applied ds-mem-0020 has a stale declined-gold
  case a current band now decides differently.

.github/workflows/smoke.yml

@@ -3,9 +3,12 @@ name: smoke
 # Fast PR gate — runs the smoke suite on every PR push.
 # Covers: chat runtime, pipeline, architectural invariants, audio sensorium
 # (tests/test_audio_*.py — compiler/eval-gates/mount/CRDT-merge/teachers, ~3s),
-# and identity falsifiability (tests/test_pack_measurements_phase2.py, ADR-0043
+# identity falsifiability (tests/test_pack_measurements_phase2.py, ADR-0043
 # — ratified packs diverge directionally; pack-invariant refusal floor; no
-# fabrication).  The falsifiability lane adds ~4 min but blocks-on-regression.
+# fabrication; adds ~4 min but blocks-on-regression), and the register axis
+# (tests/test_register_substantive_consumption.py, ADR-0069/0071/0077 — terse/
+# convivial registers must actually differ from neutral on served surfaces;
+# promoted after a 2-day silent regression, generalization-arc Tier S).
 #
 # Post-merge on main: full-pytest.yml runs the FAST lane
 # (-m "not quarantine and not slow"). Soak / proof / register-matrix coverage
@@ -56,4 +59,5 @@ jobs:
             tests/test_cognitive_turn_pipeline.py \
             tests/test_architectural_invariants.py \
             tests/test_audio_*.py \
-            tests/test_pack_measurements_phase2.py
+            tests/test_pack_measurements_phase2.py \
+            tests/test_register_substantive_consumption.py

core/cli_test.py

@@ -46,6 +46,16 @@ TEST_SUITES: dict[str, tuple[str, ...]] = {
         "tests/test_pack_measurements_phase2.py",
         # ADR-0253 dual-pack boundary — draft he/grc trees must not be serve imports.
         "tests/test_pack_draft_serve_boundary.py",
+        # Register axis (ADR-0069/0071/0077) — the e2e tests here are the
+        # falsifiable contract that terse/convivial registers actually differ
+        # from neutral on pipeline-served surfaces. Red on main outside every
+        # gate for 2026-07-20..24 (the register-axis serving regression,
+        # docs/research/lane-drift-investigation-2026-07-24.md) because this
+        # file lived only under `full`; the lane-shas job was the sole gate
+        # that runs the demo lanes this axis shows up in, and its failures
+        # were conflated with an unrelated flake. Promoted so this axis can
+        # never silently regress again.
+        "tests/test_register_substantive_consumption.py",
     ),
     "runtime": (
         "tests/test_chat_runtime.py",

docs/research/tier-s-housekeeping-2026-07-24.md

@@ -0,0 +1,71 @@
+# Tier S Housekeeping (S5)
+
+**Tier S5** (generalization-arc-2026-07-24 §6). Four items; results below.
+
+## 1. `test_register_substantive_consumption.py` promoted into smoke
+
+Added to `core/cli_test.py`'s `TEST_SUITES["smoke"]` (the local pre-push
+gate) and `.github/workflows/smoke.yml` (the mirrored CI gate — kept in
+parity rather than diverging from the local list). 36 tests, ~4.6s. This
+file is the falsifiable contract for the register axis (ADR-0069/0071/0077)
+and was red on `main`, outside every gate, for the 2026-07-22→24 window
+documented in `docs/research/public-demo-lane-drift-2026-07-24.md` §3.2 and
+[[project-register-axis-serving-regression]] — the exact "silent red" class
+the lane-drift investigation flagged as a follow-up. It can no longer
+regress without blocking a push.
+
+**Found in passing:** `tests/test_cli_test_suites.py::test_core_test_deductive_suite_expands_to_entailment_lane`
+was already red on clean `main`, unrelated to any Tier S work — it pinned
+the `deductive` suite to an exact single-file tuple from when that suite
+had one entry; the suite has since grown to 12 files across the
+deduction-serve/curriculum-serve arcs. A leaky-gate case in the same
+"pinned list goes stale" family the smoke-coverage pin exists to catch, on
+a *different* suite the coverage pin doesn't watch. Fixed to a contains-style
+assertion (matching the neighboring `fast`-suite test's own style) so it
+cannot go stale the same way again.
+
+## 2. Capability-index entries for `curriculum_serve_v1` and the `v2_exist` split
+
+Two new adapters in `evals/capability_index/adapters.py`:
+
+- `deduction_serve_existential_result()` — Band v6-EX (ADR-0261) alone, not
+  the full six-band deduction-serve lane. The other five bands are grammar
+  widenings over the same propositional-entailment substrate the existing
+  `deductive_logic` adapter already scores; v6-EX is the one that changes
+  the decision procedure itself (domain widening with Skolem witnesses / an
+  arbitrary element), making it the genuinely new capability worth its own
+  index domain.
+- `curriculum_serve_result()` — the curriculum-grounded serving lane
+  (ADR-0262) as a whole, scored against its independent oracle.
+
+Index breadth: 11 → 13 domains. Re-froze `evals/capability_index/baseline.json`
+(a deliberate re-freeze, per the file's own docstring — adding capability
+domains is exactly the "accepted improvement" case) and updated
+`tests/test_capability_baseline.py`'s implicit dependents
+(`tests/test_capability_index.py`'s `breadth == 13` and the explicit
+13-domain name set). `not_covered: []`, `wrong_total: 0` on the live run.
+
+## 3. Promotion sweep — nothing beyond `ds-mem-0020`
+
+The mechanism that surfaced `ds-mem-0020` (a case whose committed
+`gold="declined"` stopped matching `decide()`'s output the moment Band
+v6-EX could read and decide it — caught as a lane `wrong=1` failure,
+diagnosed as a promotion, gold corrected) is not a one-time manual check:
+it is exactly what `evals/deduction_serve/runner.py`'s `wrong == 0` gate
+enforces on every run, over every split, using the CURRENT full band
+cascade. Any other currently-`declined`-gold case that a band now decides
+differently would show up the same way. This session ran that gate
+repeatedly (166/166 wrong=0 both standalone and inside the 279-test
+`deductive` suite run, post-S3's vocab-trigger instrument addition) with
+zero further mismatches. **The sweep's proof is the passing gate, not a
+separate pass** — nothing else in the older splits is decidable that isn't
+already reflected in committed gold.
+
+## 4. Dead code
+
+None found unambiguous enough to remove during this sweep beyond what S2
+(`public_demo` pin) and S4
+(`generate/determine/derived_close_proposals.py::DEFAULT_SINK` off-by-one)
+already fixed as part of their own scope — see those PRs' own notes.
+
+Relates to [[project-generalization-arc]].

evals/capability_index/adapters.py

@@ -98,6 +98,41 @@ def comprehension_relational_transitive_result() -> DomainResult:
     return DomainResult("comprehension_relational_transitive", c, w, r)
 
 
+def deduction_serve_existential_result() -> DomainResult:
+    """Band v6-EX (ADR-0261) alone — the existential-argument split of the
+    deduction-serve lane, scored against its own independently-authored
+    gold. Deliberately just this one split, not the full six-band lane: the
+    other bands (v1/v1b propositional+categorical, v2-EN/v3-MEM/v4-CM/v5-VP)
+    are pure grammar widenings over the SAME propositional-entailment
+    substrate the ``deductive_logic`` adapter above already scores; the
+    existential band is the one that changes the DECISION PROCEDURE itself
+    (domain widening with Skolem witnesses / an arbitrary element — no
+    existential import), so it is the one genuinely new reasoning capability
+    among the six worth its own index domain."""
+    from evals.deduction_serve.runner import build_combined_report
+
+    counts = build_combined_report()["splits"]["v2_exist"]["counts"]
+    return DomainResult(
+        "deduction_serve_existential",
+        int(counts["correct"]),
+        int(counts["wrong"]),
+        int(counts["declined"]),
+    )
+
+
+def curriculum_serve_result() -> DomainResult:
+    """The curriculum-grounded serving lane (ADR-0262): exam questions
+    answered from the ratified domain-chain curriculum alone, scored against
+    an independent closed-world reachability oracle (``evals.curriculum_serve.oracle``,
+    a different decision procedure than the serving path it checks)."""
+    from evals.curriculum_serve.runner import build_combined_report
+
+    agg = build_combined_report()["aggregate"]
+    return DomainResult(
+        "curriculum_serve", int(agg["correct"]), int(agg["wrong"]), int(agg["declined"])
+    )
+
+
 #: The reasoning domains currently composed into the index (self-loading lanes).
 #: The six ``comprehension_*`` lanes score the GENERAL comprehension reader; the
 #: relational_metric one reads arithmetic prose into the binding-graph quantity
@@ -108,7 +143,10 @@ def comprehension_relational_transitive_result() -> DomainResult:
 #: same-predicate transitive closure over the declared strict orders (mastery-v2 Step-3
 #: Brief B2) — so the index measures comprehension breadth across categorical, ordering,
 #: propositional, quantitative, relational-reading, one-hop relational-inference, AND
-#: transitive relational-inference reasoning.
+#: transitive relational-inference reasoning. ``deduction_serve_existential`` (ADR-0261)
+#: and ``curriculum_serve`` (ADR-0262, generalization arc Tier S) add the PRODUCTION
+#: serving deciders (not bare-engine lanes) for existential-domain widening and
+#: curriculum-grounded exam answering.
 ADAPTERS = (
     deductive_logic_result,
     relational_metric_result,
@@ -121,6 +159,8 @@ ADAPTERS = (
     comprehension_relational_predicate_result,
     comprehension_relational_inference_result,
     comprehension_relational_transitive_result,
+    deduction_serve_existential_result,
+    curriculum_serve_result,
 )
 
 

evals/capability_index/baseline.json

@@ -1,13 +1,13 @@
 {
-  "capability_score": 0.953968,
-  "coverage_geomean": 0.953968,
-  "coverage_micro": 0.99399,
+  "capability_score": 0.960909,
+  "coverage_geomean": 0.960909,
+  "coverage_micro": 0.99442,
   "accuracy_micro": 1.0,
-  "breadth": 11,
+  "breadth": 13,
   "min_domain_coverage": 0.833333,
   "wrong_total": 0,
   "assert_mode_valid": true,
-  "deterministic_digest": "85c873626423c6b151a7d61bebb0cd5b7caff7dcc7e897188609e5ce3313feac",
+  "deterministic_digest": "27778ecb51b8acbd7e14a8bc23fca6869328c2236c38296bb4b1ad3311daf9eb",
   "domains": [
     {
       "domain": "comprehension_propositional",
@@ -73,6 +73,22 @@
       "coverage": 0.875,
       "accuracy": 1.0
     },
+    {
+      "domain": "curriculum_serve",
+      "correct": 32,
+      "wrong": 0,
+      "refused": 0,
+      "coverage": 1.0,
+      "accuracy": 1.0
+    },
+    {
+      "domain": "deduction_serve_existential",
+      "correct": 32,
+      "wrong": 0,
+      "refused": 0,
+      "coverage": 1.0,
+      "accuracy": 1.0
+    },
     {
       "domain": "deductive_logic",
       "correct": 716,

tests/test_capability_index.py

@@ -91,8 +91,10 @@ def test_real_lanes_compose_into_the_index_with_wrong_zero() -> None:
     # (prose -> MeaningGraph -> projection -> independent oracle / pack-named predicate)
     # compose into the cross-domain index with zero wrong commits. The one-hop
     # relational_inference lane (#775) is the 10th domain; the transitive
-    # relational_transitive lane (B2) is the 11th. (This assertion was stale at 9 from
-    # the #596 era — #775's 10th domain was never added here; corrected with B2.)
+    # relational_transitive lane (B2) is the 11th. deduction_serve_existential
+    # (ADR-0261, Band v6-EX) is the 12th; curriculum_serve (ADR-0262, generalization
+    # arc Tier S) is the 13th. (This assertion was stale at 9 from the #596 era —
+    # #775's 10th domain was never added here; corrected with B2, then again here.)
     from evals.capability_index.adapters import collect_domain_results
 
     collection = collect_domain_results()
@@ -100,7 +102,7 @@ def test_real_lanes_compose_into_the_index_with_wrong_zero() -> None:
     idx = aggregate(list(collection.results))
     assert idx.wrong_total == 0
     assert idx.assert_mode_valid
-    assert idx.breadth == 11
+    assert idx.breadth == 13
     assert {d.domain for d in idx.domains} == {
         "deductive_logic",
         "dimensional",
@@ -113,6 +115,8 @@ def test_real_lanes_compose_into_the_index_with_wrong_zero() -> None:
         "comprehension_relational_predicate",
         "comprehension_relational_inference",
         "comprehension_relational_transitive",
+        "deduction_serve_existential",
+        "curriculum_serve",
     }
     assert idx.capability_score > 0.5  # real, non-trivial cross-domain capability
 

tests/test_cli_test_suites.py

@@ -121,6 +121,14 @@ def test_core_test_fast_suite_expands_to_iteration_lane(monkeypatch) -> None:
 
 
 def test_core_test_deductive_suite_expands_to_entailment_lane(monkeypatch) -> None:
+    """Contains-style, not exact-tuple: the ``deductive`` suite has grown
+    across the deduction-serve/curriculum-serve arcs (v1 through v6-EX,
+    curriculum, the ratified-ledger bridge, the vocab-trigger instrument) and
+    keeps growing with each new band — an exact-tuple pin here would go stale
+    every time, which is exactly what happened before this fix (a pinned
+    single-file tuple, silently wrong once the suite reached 12+ files, red
+    outside every real gate — the same leaky-gate class the smoke-suite
+    coverage pin above exists to prevent)."""
     calls: list[tuple[str, ...]] = []
 
     def fake_run(*args: str, check: bool = False, cwd=None) -> int:
@@ -132,13 +140,12 @@ def test_core_test_deductive_suite_expands_to_entailment_lane(monkeypatch) -> No
     rc = cli.main(["test", "--suite", "deductive", "-q"])
 
     assert rc == 0
-    assert calls[0] == (
-        cli.sys.executable,
-        "-m",
-        "pytest",
-        "tests/test_deductive_logic_entail.py",
-        "-q",
-    )
+    command = calls[0]
+    assert command[:3] == (cli.sys.executable, "-m", "pytest")
+    assert "tests/test_deductive_logic_entail.py" in command
+    assert "tests/test_deduction_serve_lane.py" in command
+    assert "tests/test_curriculum_serve.py" in command
+    assert "-q" in command
 
 
 def test_core_test_suite_accepts_pytest_flags_without_separator(monkeypatch) -> None: