feat(pre-push-safety-net): enforce evidence eligibility and record blocked qualification

.kiro/specs/contributor-scan/pre-push-runtime-followup.md

@@ -41,3 +41,26 @@ and its compressed full evidence bundle. All 21 deadline-bound transactions time
 Every measured artifact matches the frozen installed semantics. No source, engine,
 query, fact or ranking change was made during this run. The follow-up remains necessary
 before rollout; the evaluation task does not authorize weakening its gates.
+
+## VAmPI qualification follow-up, 2026-09-13
+
+Task 8.3's packaged frozen regression (139,164 total snapshot bytes read per replay)
+retains raw findings at `models/user_model.py:73:get_user`,
+`api_views/users.py:187:update_password`, and `api_views/books.py:51:get_by_title`.
+The SQLI target has no selected exact function question; the two authorization target
+questions are selected at positions 93 and 63 but have `change_context_incomplete`.
+Head scope retains 138 selected and 20 deferred questions. Raw findings do not prove
+completed target questions or a transitive source-to-target witness.
+
+The first lab run took 112.415s; same-snapshot cached replays took 6.816s and 6.684s,
+all incomplete. This is diagnostic reuse, not changed-code performance qualification.
+The 600-second budget is distinct from the hook acceptance profile. Full artifacts are
+in `2026-09-13-push-qualification.json.gz` under the regression population.
+
+Reconcile evidence normalization and the unsupported context projection at the shared
+scan boundary before downstream admission: retain the actual query-origin/path witness
+when available, resolve supported first-party context, and keep unresolved dynamic/vendor
+boundaries explicit. Do not fabricate transitivity from a raw finding at the target.
+Revalidate ordinary VAmPI scan, exact question accounting, push regressions and independent
+qualification after an approved owner fix. This note records the prerequisite; it does
+not silently approve an upstream implementation or change detector semantics.

.kiro/specs/pre-push-safety-net/implementation-group-8.md

@@ -106,3 +106,98 @@ retune the ranker or weaken the population/gates to compensate.
 - CLAIM: Task 8.2 completed its prescribed measurement and owner routing; rollout remains NO-GO.
 - EVIDENCE: Versioned scorecard and 116-file verified bundle; seven scorer controls and full regression suite pass.
 - GAPS: No qualified warm graph path or independent capability eligibility established. These failed gates cannot authorize rollout.
+
+## Task 8.3 brief
+
+Requirements 3.1, 3.2, 6.1–6.5 and 8.1–8.4; evaluation and actionability integration.
+Implement a trusted installed eligibility resource whose evidence is recomputed against
+exact core/engine/runtime/facts/query/policy/semantic identities. Require reviewed
+operation/advice scope, vulnerable/fixed/benign populations, nonzero actual positive
+controls and reviewed alerts, untouched per-capability evaluation and all joint gates.
+The normal runner consumes only eligible declarations; stale or inconsistent resources
+admit nothing. This is integrity checking of trusted release evidence, not signing or
+permission to load repository-authored eligibility.
+
+Freeze the current packaged candidate before the six PMPro/Node development replays.
+Run readable VAmPI regression snapshots under a separate 600-second laboratory profile;
+keep their missing independent pairs unresolved. Execute libpng as an unsupported
+30-second envelope control. Do not inspect or tune on Ghost before its reviewed
+population exists. Missing required evidence leaves task 8.3 qualification blocked,
+even when the reproducible NO-GO artifact and admission implementation are delivered.
+
+### Review scope
+
+Synthetic controls prove that a matching reviewed population can reach the real runner
+admission seam, while stale facts/runtime, missing positive observations, absent reviews,
+wrong untouched workloads and changed consequence/repair semantics cannot. They are
+contract controls, not real capability qualification. The installed NO-GO resource must
+recompute identically and enable zero capabilities. The source has no repository- or
+language-specific rank changes, frontend adaptations or new abstract domains.
+
+### Task 8.3 measured outcome
+
+Packaged candidate `9464e229ccfdc873a02e3a0db7c1e83a6d6ec6d3e8234ae7b97677b4aed836fd`
+executed offline as UID 1000 with a 3 GiB memory limit. Each profile was written before
+its runs. All six development pushes saved artifacts, timed out in 30.559–30.678s and
+retained all 11 declared cases. Source snapshot receipts read 57.9–58.1 MB per PMPro
+comparison and 5.95 MB per Node comparison. These are snapshot bytes, not graph census.
+The supported introducing-change recall is 0/2, Wilson 95% interval [0, 0.658]; actual
+reviewed alert precision is undefined (0/0), never 100%. Neither fixed case has a
+completed witnessed negative. The two benign pushes emitted no alerts; that silence
+with incomplete coverage does not establish a useful safety net.
+
+VAmPI's 600-second laboratory profile executed all three pinned regression targets,
+reading 139,164 snapshot bytes per replay. First run 112.415s; subsequent identical
+comparisons 6.816s and 6.684s, each with two graph hits and 403 query hits. Raw findings
+retain the known SQLI and two authorization sites, but the SQLI exact function question
+is unlocated and authorization questions are `change_context_incomplete`. All three
+remain unresolved; no transitive witness or completed regression detection is credited.
+There are 138 selected and 20 deferred head questions. Context/witness normalization
+is routed to the shared scan owner. A registry-generation provenance probe overlapped
+startup of the first lab run; these timings are diagnostic, not runtime-gate samples.
+
+The initial instrument labeled all regression replays cold. The two proved identical
+cache replays were relabeled `identical_tip` and rescored without changing measurements,
+profile or population; original run/scorecard/instrument and correction are retained.
+Future instrument output distinguishes these identical comparisons. None is a warm edit.
+
+The libpng unsupported envelope saved a partial artifact after 30.590s. Ghost's checkout
+and reviewed independent cases remain missing and were not scanned. Independent PHP
+and Python admission populations also remain missing. No ranker, fact, adapter or query
+retune occurred. The 59-file compressed evidence bundle is verified against originals:
+`benchmarks/measurements/2026-09-13-push-qualification.json.gz`, SHA-256
+`79226b4e54042559ba5cc2b94b69b1dcddb550872fd0783b2380368be72f90bd`.
+
+The installed eligibility resource is reproducible NO-GO, digest
+`2ddd94d4dc3fd6b84e7de38e5f211d2653fb993c64e4e43cedf7e06062c3265c`, with zero enabled
+capabilities. Task 8.2 runtime evidence remains failed and predates the eligibility
+policy addition; it is correctly rejected as admission evidence, not relabeled current.
+
+## Status Report
+- STATUS: BLOCKED
+- TASK: 8.3
+- RED_PHASE_OUTPUT: `/tmp/group8-3-red.log`: eight failures with eligibility disabled; ON passed, switch removed; final synthetic and installed-resource contracts verified separately.
+- IMPLEMENTED: Versioned eligibility generation/loading, production admission integration, real offline replay and NO-GO artifact.
+- BLOCKER: Required independent Node/PHP/Python reviewed populations and verified VAmPI transitive question witness are absent. Runtime/completion gates fail.
+
+## Review Verdict
+- VERDICT: REJECTED
+- TASK: 8.3 completion claim
+- MECHANICAL_RESULTS: Eligibility and runner implementation inspected; fixed semantic/advice scope, observed controls and exact provenance required. No language-specific ranker changes, placeholders or secrets. RED phase verified.
+- FINDINGS: Implementation and reproducible NO-GO decision are concrete, but task 8.3's mandatory independent transfer and transitive-regression evidence is not established. Main-context fallback does not provide independent review.
+- REMEDIATION: Repair the shared context/witness and census/runtime prerequisites; obtain reviewed untouched paired populations, refreeze and rerun all joint gates. Keep capabilities disabled meanwhile.
+- SUMMARY: Retain the useful implementation, leave task 8.3 unchecked and feature rollout NO-GO.
+
+## Verification Result
+- STATUS: NOT_VERIFIED
+- CLAIM_TYPE: TASK
+- CLAIM: Full task 8.3 qualification complete.
+- EVIDENCE: Real frozen replay and installed decision explicitly preserve failed gates and missing evidence.
+- GAPS: Independent populations, completed positive/fixed controls, VAmPI transitive witness and representative runtime/coverage.
+- NOTES: Proceed with authorized task 8.4's explicit NO-GO acceptance reconciliation; it does not depend on pretending qualification passed.
+
+Task 8.3 implementation verification: full suite **1317 passed, 9 skipped in 115.80s**
+(`/tmp/group8-3-complete-suite.log`, exit 0). After the measurement-label correction,
+14 focused eligibility/replay-instrument tests passed in 0.95s. Ruff, format, mypy
+(116 source files) and diff checks pass. These verify the delivered code and recorded
+NO-GO behavior, not the rejected full qualification claim above.

.kiro/specs/pre-push-safety-net/spec.json

@@ -132,7 +132,10 @@
       "implementation-group-8.md"
     ],
     "next_group": 8,
-    "blocked_tasks": [],
-    "feasibility": "NO-GO: all 21 frozen NodeGoat transactions timed out; 0/24 declared target checks completed. Cancellation/reporting remained below 2 seconds. Priming did not qualify graph reuse; named census and engine lifecycle work routed to contributor-scan."
+    "blocked_tasks": [
+      "8.3"
+    ],
+    "feasibility": "NO-GO: all 21 frozen NodeGoat transactions timed out; 0/24 declared target checks completed. Cancellation/reporting remained below 2 seconds. Priming did not qualify graph reuse; named census and engine lifecycle work routed to contributor-scan.",
+    "qualification": "NO-GO: eligibility implementation and evidence delivered; required independent populations and verified VAmPI transitive witness remain missing. Zero capabilities enabled."
   }
 }

.kiro/specs/pre-push-safety-net/tasks.md

@@ -234,6 +234,7 @@ independent capability qualification. Groups 7–8 remain pending; rollout remai
   - _Boundary: Evaluation harness, Delta and Actionability Policy integration_
   - _Depends: 1.2, 1.3, 5.2, 7.3, 8.2_
   - _Requirements: 3.1, 3.2, 6.1, 6.2, 6.3, 6.4, 6.5, 8.1, 8.2, 8.3, 8.4_
+  - _Blocked 2026-09-13:_ Eligibility implementation and reproducible NO-GO artifact delivered; independent reviewed Node/PHP/Python populations and verified VAmPI transitive witness remain missing. Runtime/coverage gates fail. See `implementation-group-8.md` and contributor-scan owner follow-up.
 - [ ] 8.4 Verify the packaged end-to-end outcome and acceptance coverage
   - Run the required regression checks and a local packaged smoke over supported security/fixed/benign pushes; verify no-model startup, preservation of hook behavior and truthful partial results.
   - Reconcile every acceptance criterion with fresh evidence and record unresolved capability/performance blockers separately from completed plumbing.

benchmarks/measurements/2026-09-08-module-audit.json

@@ -1,10 +1,10 @@
 {
   "counts": {
-    "load_bearing": 105,
+    "load_bearing": 106,
     "standalone_capability": 9
   },
-  "modules": 114,
-  "note": "Refreshed for group 7: push runner reaches optional bounded witness selection; registry remains empty.",
+  "modules": 115,
+  "note": "Group 8: runner reaches recomputed versioned eligibility and its frozen scorer.",
   "orphaned": [],
   "rows": [
     {
@@ -140,6 +140,7 @@
         "model.scan",
         "model.taint",
         "push.cache",
+        "push.eligibility",
         "push.policy",
         "push.reuse",
         "push.runner"
@@ -152,6 +153,7 @@
       "importers": [
         "cpg.backend",
         "push.cache",
+        "push.eligibility",
         "push.policy",
         "push.reuse",
         "push.runner"
@@ -704,6 +706,7 @@
       "importers": [
         "cli",
         "push.assistance",
+        "push.eligibility",
         "push.policy",
         "push.report",
         "push.reuse",
@@ -746,6 +749,15 @@
     {
       "classification": "load_bearing",
       "importers": [
+        "push.runner"
+      ],
+      "module": "push.eligibility",
+      "reason": "reachable from gate, map_gate, pair_gate or the CLI"
+    },
+    {
+      "classification": "load_bearing",
+      "importers": [
+        "push.eligibility",
         "push.report",
         "push.runner"
       ],
@@ -794,7 +806,9 @@
     },
     {
       "classification": "standalone_capability",
-      "importers": [],
+      "importers": [
+        "push.eligibility"
+      ],
       "module": "push_scoring",
       "reason": "pre-push-safety-net task 1.3: frozen-profile diagnostic outcome scorer; python -m openultrasast.push_scoring"
     },

benchmarks/push/README.md

@@ -188,3 +188,27 @@ reports nearest-rank p50/p95 with sample counts, and never treats no answers as
 Exit 1 with a complete scorecard is measured NO-GO; a missing scorecard is an instrument
 failure. Three samples per workload on development NodeGoat cannot independently qualify
 a production language/framework/family or establish broad multi-project performance.
+
+### Capability qualification (task 8.3)
+
+The same `measure.py` replay instrument accepts `--selection development` (the
+six PMPro/NodeGoat changes), `--selection regression` (three VAmPI targets), or
+`--selection envelope` (libpng). Each profile retains the full declared population.
+Readable regression snapshots may execute while their missing independent evaluation
+prerequisites remain unresolved. `--deadline 600 --cache-dir /results/lab-cache`
+is a separately frozen laboratory experiment, never a replacement for the 30-second
+hook profile. A ranked or directly detected VAmPI target does not prove a transitive path.
+
+`qualify_registry.py --evidence DIR --runtime SCORECARD --out NEW_FILE` recomputes
+eligibility from the frozen profile, validated inputs, actual run and runtime scorecard.
+Without `--declarations` it creates explicitly unreviewed, experimental declarations.
+A reviewed declaration must match its frozen capability key, operation symbols, advice,
+review identity, untouched workload and exact installed analysis semantics; all quality,
+completion and runtime gates must pass. Synthetic positive controls in the unit suite
+verify this contract and do not qualify real capabilities.
+
+The installed `openultrasast/push/eligibility.json` is a trusted release input, not
+repository configuration. The runner recomputes its decision and rejects stale or
+modified evidence. This integrity check is not a cryptographic reviewer signature.
+The versioned NO-GO registry enables no capability. Regenerate evidence after changes
+to core, engine, facts, queries, policy or semantics; do not relabel old measurements.

benchmarks/push/measure.py

@@ -163,31 +163,44 @@ def main() -> int:
     parser.add_argument("--manifest", type=Path, required=True)
     parser.add_argument("--cache", type=Path, required=True)
     parser.add_argument("--out", type=Path, required=True)
+    parser.add_argument("--selection", choices=("development", "regression", "envelope"), default="development")
+    parser.add_argument("--deadline", type=float, default=30)
+    parser.add_argument("--cache-dir", type=Path)
     args = parser.parse_args()
+    if args.deadline <= 0:
+        parser.error("deadline must be positive")
     args.out.mkdir(parents=True, exist_ok=False)
     manifest = json.loads(args.manifest.read_bytes())
     validation = validate_manifest(args.manifest, cache_root=args.cache, fetch=False)
     write(args.out / "input-validation.json", validation)
     settings = ScanBudget(max_model_calls=0, max_regions=500, order_by_evidence=True)
     initial = _provenance(args.manifest.parent, settings)
-    keys = ("ranking_mode", "engine", "facts", "queries", "policy", "core")
+    keys = ("ranking_mode", "engine", "engine_runtime", "facts", "queries", "policy", "core", "semantics")
     config = {key: initial[key] for key in keys}
     config.update(
-        version="first-replay-v2-2026-09-13",
+        version="qualification-replay-v1-" + args.selection,
         scope="full tracked tree; declared first-party graph/target exclusions; evidence ranker",
         hardware=platform.platform(),
-        budget_seconds=30,
+        budget_seconds=args.deadline,
         targets=json.loads((args.manifest.parent / "diagnostic-config-v1.json").read_text())["targets"],
     )
     profile = freeze_profile(args.manifest.read_bytes(), config)
     write(args.out / "profile.json", profile)  # Freeze on disk BEFORE any measurement.
     run = {"profile_sha256": profile["profile_sha256"], "context": dict(config), "records": []}
     evidence = {"cases": [], "core_changes": [], "adapter_changes": [], "freeze_after_pmpro": None}
-    ready = {c["id"]: c["status"] == "ready" for c in validation["cases"]}
+    valid_snapshots = {s["id"] for s in validation["snapshots"] if s["status"] == "ready"}
+    ready = {
+        c["id"]: c["status"] == "ready" or (c["label"] == "regression" and c["base"] in valid_snapshots and c["tip"] in valid_snapshots)
+        for c in validation["cases"]
+    }
     snapshots = {s["id"]: s for s in manifest["snapshots"]}
     # Fixed declared order; reserved, regression and unsupported populations stay
     # in the joint scorer without scanning or tuning on the holdout.
     cases = [c for prefix in ("pmpro-", "nodegoat-") for c in manifest["cases"] if c["id"].startswith(prefix)]
+    if args.selection == "regression":
+        cases = [c for c in manifest["cases"] if c["label"] == "regression"]
+    elif args.selection == "envelope":
+        cases = [c for c in manifest["cases"] if c["label"] == "unsupported"]
     with tempfile.TemporaryDirectory(prefix="ousast-experiment-") as temporary:
         bare = Path(temporary) / "objects.git"
         bare.mkdir()
@@ -209,6 +222,7 @@ def main() -> int:
             args.out / "replay-pins.json",
             {"snapshots": pins, "provenance": "Full upstream trees; declared authored edits in a private bare Git repository."},
         )
+        seen_comparisons = set()
         for case in cases:
             if not ready[case["id"]]:
                 evidence["cases"].append({"case_id": case["id"], "status": "missing_input_prerequisite"})
@@ -231,14 +245,16 @@ def main() -> int:
                 "--artifact",
                 str(artifact_path),
                 "--deadline",
-                "30",
+                str(args.deadline),
                 "--cancellation-allowance",
                 "2",
                 "--max-regions",
                 "500",
             ]
+            if args.cache_dir is not None:
+                command += ["--cache-dir", str(args.cache_dir)]
             start = time.monotonic()
-            process = subprocess.run(command, capture_output=True, text=True, timeout=62)
+            process = subprocess.run(command, capture_output=True, text=True, timeout=args.deadline + 5)
             elapsed = time.monotonic() - start
             (args.out / (case["id"] + ".stdout")).write_text(process.stdout)
             (args.out / (case["id"] + ".stderr")).write_text(process.stderr)
@@ -257,7 +273,11 @@ def main() -> int:
                 # Unmeasured provenance caused by preparation timeout is explicit;
                 # never replace it with a fictitious runtime measurement.
                 entry["runtime_provenance_available"] = all(artifact["provenance"].get(key) == initial[key] for key in keys)
-                record = execution_record(case["id"], artifact, elapsed=elapsed, target=config["targets"][case["id"]])
+                record = execution_record(case["id"], artifact, elapsed=elapsed, target=config["targets"].get(case["id"]))
+                comparison = (pins[case["base"]], pins[case["tip"]])
+                if args.cache_dir is not None and comparison in seen_comparisons and artifact["timings"].get("graph_hits", 0) >= 2:
+                    record["cache_state"] = "identical_tip"
+                seen_comparisons.add(comparison)
                 entry["coverage_reasons"] = artifact["admission"]["coverage_reasons"]
                 entry["stage_timings"] = artifact["timings"]
                 entry["snapshot_bytes_read"] = sum(s["bytes_read"] for s in artifact["snapshots"])
@@ -291,13 +311,15 @@ def main() -> int:
         "decision": "NO-GO",
         "capabilities_enabled": False,
         "scorecard": "scorecard.json",
-        "reason": "First cold diagnostic experiment; independent eligibility and joint quality/warm-latency gates remain unmet.",
+        "reason": "Qualification evidence run; independent eligibility and joint quality/warm-latency gates remain unmet.",
+        "selection": args.selection,
+        "analysis_deadline_seconds": args.deadline,
         "recorded_cases": len(run["records"]),
         "declared_population": len(manifest["cases"]),
         "core_unchanged": True,
         "adapter_changes": [],
         "limits": [
-            "No warm-cache measurements",
+            "No representative warm changed-code measurements; identical-comparison reuse is reported separately",
             "Untouched Ghost population not scanned",
             "PHP/Python admission populations missing",
             "libpng retained as unsupported envelope; no C/C++ detection claim",

benchmarks/push/qualify_registry.py

@@ -0,0 +1,73 @@
+"""Recompute versioned eligibility from frozen results; default declarations are experimental."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+from openultrasast.model.scan import ScanBudget
+from openultrasast.push.eligibility import create_decision
+from openultrasast.push.policy import CapabilityAdmission, CapabilityKey
+from openultrasast.push.runner import _provenance
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--evidence", type=Path, required=True)
+    parser.add_argument("--runtime", type=Path, required=True)
+    parser.add_argument("--out", type=Path, required=True)
+    parser.add_argument("--declarations", type=Path)
+    args = parser.parse_args()
+
+    def read(name):
+        return json.loads((args.evidence / name).read_bytes())
+
+    profile, validation, run = read("profile.json"), read("input-validation.json"), read("run.json")
+    current = _provenance(args.evidence, ScanBudget(max_model_calls=0, max_regions=500, order_by_evidence=True))
+    if args.declarations:
+        declarations = tuple(CapabilityAdmission.from_payload(d) for d in json.loads(args.declarations.read_bytes()))
+    else:
+        declarations = []
+        for capability in sorted({c["capability"] for c in profile["cases"]}):
+            language, framework, family = capability.split("/")
+            key = CapabilityKey(
+                language,
+                "unspecified",
+                framework,
+                family,
+                "dominance" if family == "access_control" else "taint",
+                "unreviewed",
+                current["semantics"],
+            )
+            declarations.append(
+                CapabilityAdmission(
+                    key,
+                    "unqualified-population-v1",
+                    str(args.evidence / "scorecard.json"),
+                    "experimental",
+                    "Unreviewed consequence for {operation} in {context}.",
+                    "Unreviewed repair for {operation} in {context}.",
+                )
+            )
+    decision = create_decision(
+        profile, validation, run, current=current, runtime=json.loads(args.runtime.read_bytes()), declarations=declarations
+    )
+    args.out.parent.mkdir(parents=True, exist_ok=True)
+    with args.out.open("x") as stream:
+        json.dump(decision, stream, indent=2, allow_nan=False)
+        stream.write("\n")
+    print(
+        json.dumps(
+            {
+                "verdict": decision["verdict"],
+                "enabled": sum(e["declaration"]["enabled"] for e in decision["entries"]),
+                "artifact_sha256": decision["artifact_sha256"],
+            }
+        )
+    )
+    return 0 if decision["verdict"] == "PASS" else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

src/openultrasast/push/eligibility.py

@@ -0,0 +1,171 @@
+"""Versioned, locally shipped capability eligibility from frozen evaluation evidence.
+
+This verifies integrity and recomputes gates; it is not a signature verifier. Registry
+files are trusted release inputs, never repository configuration or model responses.
+"""
+
+from __future__ import annotations
+
+import json
+import math
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass, replace
+from pathlib import Path
+from typing import Any
+
+from openultrasast.cpg.artifact import digest_value
+from openultrasast.push.policy import CapabilityAdmission
+from openultrasast.push_scoring import _metrics, score
+
+_BINDINGS = ("core", "engine", "engine_runtime", "facts", "queries", "policy", "semantics")
+_WORKLOADS = ("identical-tip", "function-edit", "dependency-edit", "configuration-edit", "cold", "growth", "multi-ref")
+_DEFAULT = Path(__file__).with_name("eligibility.json")
+
+
+@dataclass(frozen=True)
+class EligibilityLoad:
+    capabilities: tuple[CapabilityAdmission, ...]
+    status: str
+    reasons: tuple[str, ...]
+
+
+def fingerprint() -> str:
+    try:
+        return digest_value(json.loads(_DEFAULT.read_bytes()))
+    except (OSError, ValueError):
+        return "eligibility_unavailable"
+
+
+def _runtime_pass(runtime: Mapping[str, Any], current: Mapping[str, str]) -> bool:
+    if runtime.get("runtime_verdict") != "PASS" or any(runtime.get("provenance", {}).get(k) != current.get(k) for k in _BINDINGS):
+        return False
+    try:
+        for name in _WORKLOADS:
+            row = runtime["workloads"][name]
+            values = [
+                row[k]
+                for k in (
+                    "population",
+                    "observed",
+                    "missing",
+                    "latency_samples",
+                    "p95_seconds",
+                    "completion",
+                    "complete_coverage_cases",
+                    "cancellation_overrun_seconds",
+                    "timeouts",
+                )
+            ]
+            if any(type(v) not in (int, float) or not math.isfinite(v) or v < 0 for v in values):
+                return False
+            population = row["population"]
+            if (
+                population < 1
+                or row["observed"] != population
+                or row["latency_samples"] != population
+                or row["missing"]
+                or row["timeouts"]
+                or row["completion"] < 0.95
+                or row["complete_coverage_cases"] / population < 0.95
+                or row["cancellation_overrun_seconds"] > 2
+                or (name != "cold" and row["p95_seconds"] > 30)
+            ):
+                return False
+    except (KeyError, TypeError, ZeroDivisionError):
+        return False
+    return True
+
+
+def create_decision(
+    profile: dict[str, Any],
+    validation: dict[str, Any],
+    run: dict[str, Any],
+    *,
+    current: Mapping[str, str],
+    runtime: dict[str, Any],
+    declarations: Sequence[CapabilityAdmission],
+) -> dict[str, Any]:
+    scored = score(profile, validation, run)
+    bindings = {k: current.get(k, "") for k in _BINDINGS}
+    common = []
+    if any(not bindings[k] or profile["config"].get(k) != bindings[k] for k in _BINDINGS):
+        common.append("evaluation_provenance_mismatch")
+    if not _runtime_pass(runtime, current):
+        common.append("runtime_gate_or_provenance_failed")
+    entries = []
+    for declaration in declarations:
+        key = declaration.key
+        capability = f"{key.language}/{key.framework}/{key.family}"
+        rows = [r for r in scored["cases"] if r["capability"] == capability and r["label"] != "regression"]
+        metrics = _metrics(rows)
+        reasons = list(common)
+        labels = {r["label"] for r in rows}
+        if not {"vulnerable", "fixed", "benign"} <= labels:
+            reasons.append("paired_positive_fixed_benign_population_missing")
+        if any(r["label"] in ("unsupported", "unreviewed") or r["input_status"] != "ready" or r.get("prerequisites") for r in rows):
+            reasons.append("unresolved_input_or_unsupported_population")
+        if not declaration.untouched_workload or any(
+            r.get("role") not in ("independent", "reserved") or r.get("workload") != declaration.untouched_workload for r in rows
+        ):
+            reasons.append("independent_workload_missing")
+        semantic = profile["config"].get("capability_semantics", {}).get(digest_value(key.to_payload()), {})
+        if (
+            semantic.get("key") != key.to_payload()
+            or semantic.get("operation_symbols") != list(declaration.operation_symbols)
+            or semantic.get("review_id") != declaration.evaluation_id
+            or semantic.get("consequence_template") != declaration.consequence_template
+            or semantic.get("repair_template") != declaration.repair_template
+            or not declaration.operation_symbols
+            or key.semantics != bindings["semantics"]
+            or key.context in ("unreviewed", "unspecified", "")
+            or not declaration.calibration_artifact
+        ):
+            reasons.append("semantic_scope_unreviewed_or_mismatched")
+        for name, threshold in (("precision", 0.95), ("actionable_recall", 0.90), ("completion", 0.95), ("query_completion", 0.95)):
+            if metrics[name]["value"] is None or metrics[name]["value"] < threshold:
+                reasons.append(name + "_gate_failed")
+        if metrics["fixed_side_silence"]["value"] != 1 or metrics["benign_interruption"]["value"] != 0:
+            reasons.append("fixed_or_benign_gate_failed")
+        warm = metrics["latency"]["warm_changed"]
+        if not warm["samples"] or warm["p95"] > 30:
+            reasons.append("independent_warm_latency_missing_or_failed")
+        positive = metrics["recall"]["numerator"]
+        reviewed = metrics["precision"]["numerator"]
+        if positive < 1 or reviewed < 1:
+            reasons.append("positive_controls_or_reviewed_alerts_missing")
+        accepted = not reasons
+        resolved = replace(
+            declaration, verdict="PASS" if accepted else "NO-GO", enabled=accepted, positive_controls=positive, reviewed_alerts=reviewed
+        )
+        entries.append(dict(declaration=resolved.to_payload(), metrics=metrics, reasons=sorted(set(reasons))))
+    result = dict(
+        schema_version=1,
+        bindings=bindings,
+        verdict="PASS" if any(not e["reasons"] for e in entries) else "NO-GO",
+        entries=entries,
+        inputs=dict(profile=profile, validation=validation, run=run, runtime=runtime, declarations=[d.to_payload() for d in declarations]),
+    )
+    result["artifact_sha256"] = digest_value(result)
+    return result
+
+
+def load_registry(*, current: Mapping[str, str], path: Path | None = None) -> EligibilityLoad:
+    try:
+        with (path or _DEFAULT).open("rb") as stream:
+            content = stream.read(2 * 1024**2 + 1)
+        if len(content) > 2 * 1024**2:
+            raise ValueError("registry_size_limit")
+        data = json.loads(content)
+        if data.get("schema_version") != 1 or data.get("bindings") != {k: current.get(k, "") for k in _BINDINGS}:
+            return EligibilityLoad((), "stale", ("eligibility_provenance_mismatch",))
+        inputs = data["inputs"]
+        declarations = tuple(CapabilityAdmission.from_payload(d) for d in inputs["declarations"])
+        recomputed = create_decision(
+            inputs["profile"], inputs["validation"], inputs["run"], current=current, runtime=inputs["runtime"], declarations=declarations
+        )
+        if recomputed != data:
+            raise ValueError("registry_evidence_mismatch")
+        capabilities = tuple(CapabilityAdmission.from_payload(e["declaration"]) for e in data["entries"] if not e["reasons"])
+        return EligibilityLoad(capabilities, data["verdict"], tuple(sorted({reason for e in data["entries"] for reason in e["reasons"]})))
+    except (OSError, ValueError, TypeError, KeyError, AttributeError):
+        return EligibilityLoad((), "unavailable", ("eligibility_invalid_or_unavailable",))

src/openultrasast/push/runner.py

@@ -26,6 +26,7 @@ from openultrasast.model.shipped import declared_sources
 from openultrasast.preprocess import build_file_target, enumerate_source_files
 from openultrasast.push.cache import ArtifactCache, SemanticKeys
 from openultrasast.push.contracts import ComparisonAnalysis, PushComparison, PushResult, SnapshotManifest
+from openultrasast.push.eligibility import fingerprint, load_registry
 from openultrasast.push.policy import (
     ActionableDefect,
     AdmissionCandidate,
@@ -151,11 +152,22 @@ def _provenance(repository: Path, scan_budget: ScanBudget) -> dict[str, str]:
         ),
         "ranking_mode": "evidence",
         "model": "disabled",
-        "capabilities": "empty_registry_pending_task_8.3",
+        "capabilities": fingerprint(),
         "cache": "cold_no_reuse",
     }
 
 
+def _admit(candidates: tuple[AdmissionCandidate, ...], provenance: dict[str, str]) -> AdmissionResult:
+    registry = load_registry(current=provenance)
+    result = admit_candidates(candidates, capabilities=registry.capabilities)
+    if candidates and registry.reasons:
+        result = replace(
+            result,
+            coverage_reasons=tuple(dict.fromkeys((*result.coverage_reasons, *("eligibility:" + reason for reason in registry.reasons)))),
+        )
+    return result
+
+
 def _discovery_identity() -> str:
     root = Path(__file__).resolve().parents[1]
     # Discovery invokes the shipped mapper, preprocessing, layout and family specs.
@@ -326,7 +338,7 @@ def _analyze(
                     )
                     for c in delta.candidates
                 )
-                admission = _prepare(lambda: admit_candidates(candidates), budget)
+                admission = _prepare(lambda: _admit(candidates, provenance), budget)
                 timings[stage + "_seconds"] = time.monotonic() - stage_started
                 stage, stage_started = "cleanup", time.monotonic()
         timings[stage + "_seconds"] = time.monotonic() - stage_started

tests/test_push_eligibility.py

@@ -0,0 +1,163 @@
+"""Synthetic eligibility contracts, not real capability qualification."""
+
+import copy
+import json
+from dataclasses import replace
+
+import pytest
+from test_push_admission import control
+from test_push_scoring import inputs, record
+
+from openultrasast.push.eligibility import create_decision, load_registry
+from openultrasast.push_scoring import freeze_profile
+
+
+def fixture():
+    old, validation = inputs()
+    current = {key: "synthetic-" + key for key in ("core", "engine", "engine_runtime", "facts", "queries", "policy", "semantics")}
+    from openultrasast.cpg.artifact import digest_value
+
+    _, declaration, _ = control()
+    declaration = replace(declaration, key=replace(declaration.key, semantics=current["semantics"]), untouched_workload="untouched-control")
+    config = {
+        **old["config"],
+        **current,
+        "capability_semantics": {
+            digest_value(declaration.key.to_payload()): {
+                "key": declaration.key.to_payload(),
+                "operation_symbols": list(declaration.operation_symbols),
+                "review_id": declaration.evaluation_id,
+                "consequence_template": declaration.consequence_template,
+                "repair_template": declaration.repair_template,
+            }
+        },
+    }
+    cases = [{**row, "role": "independent", "workload": "untouched-control"} for row in old["cases"][:3]]
+    raw = json.dumps({"schema_version": 1, "cases": cases}).encode()
+    profile = freeze_profile(raw, config)
+    validation = {**validation, "manifest_sha256": profile["manifest_sha256"], "cases": [{**c, "status": "ready"} for c in cases]}
+    positive = record(
+        alerts=[dict(review="actionable_true", review_evidence="synthetic-review", question_id="q-handler", site="api.js:4")],
+        cache_state="warm_changed",
+    )
+    fixed = record(
+        "fixed",
+        queries=[
+            dict(
+                id="q-handler",
+                status="answered",
+                outcome="negative",
+                target=config["targets"]["fixed"],
+                witness="bound parameter discharges target",
+            )
+        ],
+        delta="removed",
+        cache_state="warm_changed",
+    )
+    benign = record("benign", delta="unchanged", cache_state="warm_changed")
+    run = {"profile_sha256": profile["profile_sha256"], "context": config, "records": [positive, fixed, benign]}
+    _, declaration, _ = control()
+    declaration = replace(declaration, key=replace(declaration.key, semantics=current["semantics"]), untouched_workload="untouched-control")
+    runtime = {
+        "runtime_verdict": "PASS",
+        "provenance": current,
+        "workloads": {
+            name: dict(
+                population=3,
+                observed=3,
+                missing=0,
+                latency_samples=3,
+                p95_seconds=3,
+                completion=1,
+                complete_coverage_cases=3,
+                cancellation_overrun_seconds=0,
+                timeouts=0,
+            )
+            for name in ("identical-tip", "function-edit", "dependency-edit", "configuration-edit", "cold", "growth", "multi-ref")
+        },
+    }
+    return profile, validation, run, current, runtime, (declaration,)
+
+
+def decision(parts):
+    profile, validation, run, current, runtime, declarations = parts
+    return create_decision(profile, validation, run, current=current, runtime=runtime, declarations=declarations)
+
+
+def test_matching_reviewed_synthetic_population_can_be_loaded(tmp_path):
+    parts = fixture()
+    data = decision(parts)
+    assert data["verdict"] == "PASS"
+    path = tmp_path / "eligibility.json"
+    path.write_text(json.dumps(data))
+    loaded = load_registry(current=parts[3], path=path)
+    assert len(loaded.capabilities) == 1
+    assert loaded.capabilities[0].enabled
+    assert loaded.capabilities[0].positive_controls == 1
+    assert loaded.capabilities[0].reviewed_alerts == 1
+
+
+@pytest.mark.parametrize("mutation", ["no_alert", "runtime", "stale_runtime", "no_independent", "no_positive", "wrong_semantics"])
+def test_missing_joint_evidence_never_enables_advisory(mutation, tmp_path):
+    parts = list(copy.deepcopy(fixture()))
+    if mutation == "no_alert":
+        parts[2]["records"][0]["alerts"] = []
+    elif mutation == "runtime":
+        parts[4]["workloads"]["function-edit"]["p95_seconds"] = 31
+    elif mutation == "stale_runtime":
+        parts[4]["provenance"]["policy"] = "other"
+    elif mutation == "no_independent":
+        parts[5] = (replace(parts[5][0], untouched_workload="wrong-repository"),)
+    elif mutation == "no_positive":
+        parts[2]["records"] = parts[2]["records"][1:]
+    else:
+        parts[5] = (replace(parts[5][0], key=replace(parts[5][0].key, semantics="wrong")),)
+    data = decision(parts)
+    assert data["verdict"] == "NO-GO"
+    path = tmp_path / "eligibility.json"
+    path.write_text(json.dumps(data))
+    assert not load_registry(current=parts[3], path=path).capabilities
+
+
+def test_stale_or_tampered_registry_is_rejected(tmp_path):
+    parts = fixture()
+    data = decision(parts)
+    path = tmp_path / "eligibility.json"
+    path.write_text(json.dumps(data))
+    stale = {**parts[3], "facts": "new-facts"}
+    assert not load_registry(current=stale, path=path).capabilities
+    data["entries"][0]["declaration"]["repair_template"] = "invented advice"
+    path.write_text(json.dumps(data))
+    assert not load_registry(current=parts[3], path=path).capabilities
+
+
+def test_runner_consumes_only_matching_registry(tmp_path, monkeypatch):
+    import openultrasast.push.eligibility as registry
+    from openultrasast.push.runner import _admit
+
+    parts = fixture()
+    data = decision(parts)
+    path = tmp_path / "eligibility.json"
+    path.write_text(json.dumps(data))
+    monkeypatch.setattr(registry, "_DEFAULT", path)
+    candidate, _, _ = control()
+    candidate = replace(candidate, capability=parts[5][0].key, delta=replace(candidate.delta, semantics=parts[3]["semantics"]))
+    assert len(_admit((candidate,), parts[3]).defects) == 1
+    assert not _admit((candidate,), {**parts[3], "facts": "changed"}).defects
+
+
+def test_changed_semantic_advice_needs_a_new_frozen_review():
+    parts = list(fixture())
+    parts[5] = (replace(parts[5][0], repair_template="Unreviewed advice at {operation} for {context}"),)
+    assert decision(parts)["verdict"] == "NO-GO"
+
+
+def test_shipped_registry_is_reproducible_no_go():
+    from openultrasast.push.eligibility import _DEFAULT
+
+    data = json.loads(_DEFAULT.read_bytes())
+    loaded = load_registry(current=data["bindings"])
+    assert loaded.status == "NO-GO"
+    assert not loaded.capabilities
+    assert data["entries"]
+    assert all(not row["declaration"]["enabled"] for row in data["entries"])