sourcecode 4.5.3__tar.gz → 4.6.0__tar.gz

{sourcecode-4.5.3 → sourcecode-4.6.0}/CHANGELOG.md

@@ -2,6 +2,133 @@
 
 ## [Unreleased]
 
+## [4.6.0] — 2026-08-04
+
+Field evaluation #11 (7,5/10), by the evaluator who measured 4.5.2 twelve hours
+earlier, on the build that closed their previous list. **Eight complaints, eight
+reproductions, eight fixes** — the first evaluation where every complaint
+survives verification and none re-derives a closed row.
+
+The headline is not in the list below and is worth stating: the four false
+positives measured on 4.5.2 are **gone**, verified line by line in the field, and
+the run reports **0 false positives across 19 findings** — precision 9/10,
+epistemic honesty 10/10 (*"no he visto nada mejor en ninguna herramienta de
+análisis, comercial o no"*). E-2 is closed in the field, not only in the battery.
+
+Minor, not patch: `risk` publishes a sixth reachability value and three new
+fields, `posture` publishes two, `impact-chain` publishes two, and the test/
+production split moves published populations. **No field was removed and no exit
+code moved.**
+
+### Fixed
+
+- **`impact-chain` spent 2 300× longer answering than building the model it
+  answers from** (C3-42, **high**). The field published `query_time_ms: 1 187 063`
+  beside `model_build_time_ms: 516` and abandoned two more invocations past 25
+  minutes — the axis this evaluation scored **3/10**, with the verdict that
+  decides it: *"un gate de CI que tarda 45 min no es un gate"*. The seed
+  expansions (CH-001a/b) answered *"every symbol of this class"* by scanning the
+  whole symbol table once per subtype, with the seed set rebuilt for every symbol
+  tested. Both now read one class→symbols index memoized on the CIR, which
+  matters twice: `RiskComposer.reach` calls that path **once per defect**, and
+  that is where the field's ~45-minute `risk` comes from. Measured A/B on
+  keycloak (5 525 files, 59 410 symbols), identical results in all four queries:
+  `Provider` **8.15 s → 2.60 s**. Second half, and the reason the rest is flat:
+  `query_time_ms` was never the walk — it includes a repository-wide TX +
+  security audit, now published as `audit_time_ms` with a `query_time_basis`
+  stating that the walk is the difference between the two and that
+  `model_build_time_ms` is not comparable with either.
+- **`risk` did not read the other authority that answered its own access axis**
+  (C1-26, **high**). It asked `posture` only when `--profile` was given, so the
+  default invocation weighed an endpoint at `unknown` (×1.4) that `posture`, in
+  the same process and over the same CIR, decided was reachable without
+  authentication (×2.0). The resolution is now read always; with no set named it
+  is the one Spring performs with none given — `default` alone, which CL-8
+  records as the permissive case, not a neutral one — and every row publishes
+  `profile_set`, `profile_set_basis` (*"a default, never a guarantee: the process
+  environment outranks every file"*) and `profile_decisions`. It cannot
+  manufacture a verdict: `no_rule_matched` stays unknown coverage, the
+  combination is per endpoint, and on a fixture where both chains are
+  profile-gated the answer stays `unknown`. Measured on the field's shape: **8.06
+  → 11.52**, exactly the ratio of the two access weights, every other factor
+  unmoved.
+- **`risk` degraded a repository-wide control defect with a sentence that is
+  false in effect** (C1-25, configuration half; second field witness). *"No HTTP
+  endpoint in this repository reaches it"* — about the file that decides whether
+  any endpoint is reachable at all. Reach for such a defect is **declared, not
+  walked**: `posture` already knew which configuration file decided which
+  request and now publishes it as `chains.decides_endpoints`, `risk` reads it as
+  the reachability value `request_chain`, and the two units stay apart —
+  `endpoints_reached` remains the walk, `decides_requests` is the scope.
+  Evidence-bound both ways: only a file declaring readable chain rules qualifies,
+  and an unresolved servlet path clears the attribution entirely. Fixture: **low
+  2.8 → medium 4.48**.
+- **`posture --resolve-environments` truncated the open-endpoint list without
+  declaring the cap** (C2-25). 50 entries under a count of 2 635, with no
+  `total`, no `omitted`, no `direction: under_reports` — in the module that
+  defines the cap helper and applies it to four other lists. *"Rompe la garantía
+  que el resto del tool cumple escrupulosamente, y es precisamente en la lista
+  que un auditor necesita completa."* The helper grew the one parameter it was
+  missing (the limit that actually bit) and every other slice in the module was
+  checked.
+- **`explain` answered a configuration class with every section empty**,
+  `found: true` and no warning (C3-41, **high**) — on the command the same
+  evaluation calls the product's best asset, for the class an auditor opens
+  first. One model error with three faces: a factory member is a method
+  (`symbol_kind: "bean"` was excluded by kind, so **every** `@Configuration` in
+  every repository published no methods); a configuration class wires through its
+  members' signatures and bodies, not through an injection graph (both are atoms
+  already published, read only when the class really declares factory members);
+  and its constraints are in a DSL, so the section now reads `chain_rules` — the
+  authority `posture` uses on that same file — with `file:line` per rule. Fourth,
+  and the reason the report was silent: every section was wrapped in a bare
+  `except: []`, so a crash and a measured emptiness were one output. A failed
+  section now says which one failed.
+- **`cache freshness` called a snapshot STALE over a file no analyser reads**
+  (C1-27) — with `RIS HEAD == current HEAD` and `Delta: 0` beside it. C3-40
+  (4.5.3) replaced a false *fresh* with a false *stale* by counting any porcelain
+  line. One authority now decides which paths qualify
+  (`path_filters.analysis_reads`) and both consumers ask it: the freshness
+  boolean and the **tree signature**, which is what actually decides the verdict
+  and is also the cache key — so an agent writing its own settings file was
+  discarding every warm answer. The excused set is state nothing opens and is
+  **not** every dot directory: `.github` is read as tooling evidence and stays a
+  change. The default is inclusive, because a missed change is a false *fresh*.
+- **The test/production split still had two authorities** (C1-28).
+  `path_filters.is_test_path` was never retired after C1-2 made `test_sources`
+  the authority in 3.2.1, and the two disagreed on five path shapes in both
+  directions. Field symptom: *"`--compact` counts 1 test file that is actually in
+  `src/main`"* — verdict right, numerator wrong. The second module is now a
+  passthrough, after absorbing everything it knew that the authority did not; and
+  a declared **main** source root now outranks a naming convention, the same rule
+  the directory half always applied, in the other direction. Fleet effect,
+  measured cold: keycloak 5 525 → 5 486 Java files with endpoints **676 → 678**,
+  openmrs-core 866 → 861, Broadleaf / alfresco / petclinic unchanged.
+- **`archetype` never read the endpoint surface for the dimension the endpoint
+  surface decides** (C3-43). A 3 574-endpoint CRUD REST monolith was labelled
+  `engine` at medium confidence on a fan-in Gini, while the legacy classifier it
+  supersedes answered `api` and was right. The density `_primary_interface`
+  already reads is now positive evidence for `application` and **named negative
+  evidence** against `engine` — the shape the library de-bias uses — scaled by
+  the measured density, and emitted only when the surface was measured.
+  Measured: Broadleaf and keycloak `engine/high → application/low`, petclinic
+  `engine/low → application/medium`, jobrunr `engine/high → engine/low`,
+  openmrs-core unchanged (1 endpoint, no density, no discount — the negative
+  control).
+
+### Notes
+
+Two rows from this evaluation are **not** fixed here, on purpose: **P-9** is a
+pricing decision (a fifth pass, and the fourth to argue itself off the seat axis
+— the ceiling is conditioned on this battery's performance work and on a tier
+promotion, not on any missing capability), and **CL-13** is the narrow rule
+catalogue (7 SEC + 6 TX), queued as **F-F** with the *test surface in
+production* rule first, because C1-28 just gave it its admission surface for
+free. Full record in `.planning/FIELD-EVAL-11-2026-08-04.md`; every row is in
+`docs/DEFECT-LEDGER.md`.
+
+Suite: 6 468 passed, 6 skipped.
+
 ## [4.5.3] — 2026-08-04
 
 Patch: the fixable half of field evaluation #10 (7,5/10, CI reliability **4/10**), in one

{sourcecode-4.5.3 → sourcecode-4.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sourcecode
-Version: 4.5.3
+Version: 4.6.0
 Summary: Persistent structural context and ultra-fast repeated analysis for AI coding agents
 License-File: LICENSE
 Keywords: agents,ai,codebase,context,developer-tools,llm
@@ -42,7 +42,7 @@ Description-Content-Type: text/markdown
 
 **Context · Impact · Migration · Architecture · Review — everything from one structural model.**
 
-![Version](https://img.shields.io/badge/version-4.5.3-blue)
+![Version](https://img.shields.io/badge/version-4.6.0-blue)
 ![Python](https://img.shields.io/badge/python-3.9%2B-green)
 
 > **ASK Engine** is the product. The CLI command is **`ask`**. The legacy **`sourcecode`**
@@ -97,7 +97,7 @@ brew tap haroundominique/sourcecode && brew install sourcecode
 # pip / pipx
 pipx install sourcecode        # or: pip install sourcecode
 
-ask version                    # ask 4.5.3 — and, on a build that has aged,
+ask version                    # ask 4.6.0 — and, on a build that has aged,
                                # how many releases have probably shipped since
 ```
 

{sourcecode-4.5.3 → sourcecode-4.6.0}/README.md

@@ -4,7 +4,7 @@
 
 **Context · Impact · Migration · Architecture · Review — everything from one structural model.**
 
-![Version](https://img.shields.io/badge/version-4.5.3-blue)
+![Version](https://img.shields.io/badge/version-4.6.0-blue)
 ![Python](https://img.shields.io/badge/python-3.9%2B-green)
 
 > **ASK Engine** is the product. The CLI command is **`ask`**. The legacy **`sourcecode`**
@@ -59,7 +59,7 @@ brew tap haroundominique/sourcecode && brew install sourcecode
 # pip / pipx
 pipx install sourcecode        # or: pip install sourcecode
 
-ask version                    # ask 4.5.3 — and, on a build that has aged,
+ask version                    # ask 4.6.0 — and, on a build that has aged,
                                # how many releases have probably shipped since
 ```
 

{sourcecode-4.5.3 → sourcecode-4.6.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "sourcecode"
-version = "4.5.3"
+version = "4.6.0"
 description = "Persistent structural context and ultra-fast repeated analysis for AI coding agents"
 readme = "README.md"
 requires-python = ">=3.9"

{sourcecode-4.5.3 → sourcecode-4.6.0}/src/sourcecode/__init__.py

@@ -4,4 +4,4 @@ ASK Engine is the product. ``ask`` is the canonical CLI command; ``sourcecode``
 the legacy compatibility alias and the Python/PyPI package name. See
 docs/PRODUCT_IDENTITY.md (normative)."""
 
-__version__ = "4.5.3"
+__version__ = "4.6.0"

{sourcecode-4.5.3 → sourcecode-4.6.0}/src/sourcecode/archetype.py

@@ -86,6 +86,14 @@ _MIN_CONFIDENT_SCORE = 0.5
 # discriminating. Tuned so a clear library (no app/server/servlet entry) is not confidently
 # labeled an engine, while runnable systems (which never trigger the gate) are untouched.
 _LIBRARY_ENGINE_DISCOUNT = 0.55
+# The same shape for the opposite mistake (C3-43): a repository whose primary
+# surface is HTTP owns storage/kernel packages and hub topology *because* it
+# serves requests with them, so that evidence is not discriminating for "engine"
+# either. Scaled by the measured density, so a repository with a marginal surface
+# is barely discounted and one that is nothing but endpoints is discounted fully.
+# Below the library figure on purpose: an application CAN be built around an
+# engine, whereas a non-runnable artifact cannot be an engine product at all.
+_HTTP_ENGINE_DISCOUNT = 0.45
 # Entry kinds that make a repo a runnable system (an app/server/servlet/daemon) rather
 # than a consumable library. Used to gate the library de-bias (P0).
 _RUNNABLE_ENTRY_KINDS = frozenset({
@@ -409,6 +417,42 @@ class ArchetypeClassifier:
             add("application", "runtime_entry_plus_services",
                 f"bootstrap/server entry + service mass ≈ {cm['service']:.0%}",
                 2.0, 1.0, max(cm["service"], 0.2))
+        # C3-43. The strongest MEASURED discriminator this build has for what a
+        # repository is — its HTTP surface — was read by `_primary_interface` and
+        # ignored here, so a 3 574-endpoint CRUD REST monolith was labelled
+        # `engine` on package-name masses and a fan-in Gini, while the legacy
+        # classifier it supersedes answered `api` and was right. Same ramp as the
+        # interface dimension, from the same features: one authority for the
+        # figure, two dimensions reading it.
+        if f.endpoint_share is None:
+            # Unmeasured is not zero: contribute nothing, and say so.
+            add("application", "http_surface_density",
+                "endpoint surface not measured — this dimension is scored without it",
+                2.5, 0.0, 0.0)
+        else:
+            density = f.endpoint_share * 100
+            http_strength = (
+                max(0.0, min(1.0, (density - 0.5) / 1.5)) if f.endpoint_total else 0.0
+            )
+            add("application", "http_surface_density",
+                f"{f.endpoint_total} endpoints over {f.total_files} files "
+                f"({density:.2f}/100 files) — the repository's primary surface is HTTP",
+                2.5, http_strength, 1.0)
+            # And the same fact as NAMED negative evidence for `engine`, the way
+            # the library de-bias below does it: a codebase whose product is a
+            # request surface owns storage and kernel packages *because* it is an
+            # application, so that mass is not, alone, evidence of an engine.
+            engine_positive_http = sum(max(0.0, e.contribution) for e in cands["engine"])
+            if http_strength > 0 and engine_positive_http > 0:
+                cands["engine"].append(Evidence(
+                    "http_surface_discount",
+                    f"{f.endpoint_total} endpoints over {f.total_files} files: the "
+                    "primary surface of this repository is HTTP, and the storage/"
+                    "kernel mass and hub topology an application needs to serve it "
+                    "are not, alone, evidence of an engine product",
+                    1.0, 1.0, 1.0,
+                    -engine_positive_http * _HTTP_ENGINE_DISCOUNT * http_strength,
+                ))
         # platform: many modules + multiple concept clusters + distribution
         multi = 1.0 if f.module_count >= 8 else f.module_count / 8
         add("platform", "many_modules_multi_concern",

{sourcecode-4.5.3 → sourcecode-4.6.0}/src/sourcecode/baseline_autocapture.py

@@ -210,7 +210,19 @@ def worktree_dirty(
         return None
     if out.returncode != 0:
         return None
-    lines = [ln for ln in out.stdout.splitlines() if ln.strip()]
+    # C1-27. "Dirty" means the tree *the analysis reads* differs from HEAD. A
+    # change under an editor or agent state directory is not one — no analyser
+    # opens those files — and counting it reported STALE on a snapshot that
+    # described the tree exactly, with `Delta: 0` beside it. One authority answers
+    # which paths qualify (`path_filters.analysis_reads`) and its default is
+    # inclusive: anything not named there still counts, because a missed change
+    # is a false *fresh* and a spurious one only costs a rebuild.
+    from sourcecode.path_filters import analysis_reads
+
+    lines = [
+        ln for ln in out.stdout.splitlines()
+        if ln.strip() and analysis_reads(_porcelain_path(ln))
+    ]
     if ignore is None:
         return bool(lines)
     candidates = [ignore] if isinstance(ignore, (str, Path)) else list(ignore)

{sourcecode-4.5.3 → sourcecode-4.6.0}/src/sourcecode/cache.py

@@ -213,6 +213,14 @@ def _untracked_tree_fingerprint(target: Path) -> str:
     return f"ff{h.hexdigest()[:14]}"
 
 
+def _porcelain_status_path(line: str) -> str:
+    """The path a `git status --porcelain` line refers to (destination on renames)."""
+    body = line[3:] if len(line) > 3 else ""
+    if " -> " in body:
+        body = body.split(" -> ", 1)[1]
+    return body.strip().strip('"')
+
+
 def worktree_signature(repo_root: Path, scope: Optional[Path] = None) -> str:
     """Hex fingerprint of the **exact tree state an analysis would read**.
 
@@ -252,12 +260,30 @@ def worktree_signature(repo_root: Path, scope: Optional[Path] = None) -> str:
         porcelain = st.stdout if st.returncode == 0 else ""
     except Exception:
         porcelain = ""
+    # C1-27. The signature describes *the tree an analysis would read*, so a line
+    # about a file no analyser opens does not belong in it — an editor settings
+    # file was invalidating every cached answer and publishing STALE beside
+    # `Delta: 0`. One authority decides which paths qualify, and its default is
+    # inclusive: anything not named there still invalidates.
+    from sourcecode.path_filters import analysis_reads
+
+    porcelain = "\n".join(
+        ln for ln in porcelain.splitlines()
+        if ln.strip() and analysis_reads(_porcelain_status_path(ln))
+    )
     if not porcelain.strip():
         return head  # clean — HEAD fully describes the tree
 
     try:
+        from sourcecode.path_filters import _TOOL_STATE_DIRS
+
+        # The same exclusion, expressed to git for the half that is not a list of
+        # paths. A tracked editor-state file must not enter the signature through
+        # the diff after being kept out of the status.
+        _diff_spec = list(pathspec) or ["--", "."]
+        _diff_spec += [f":(exclude,glob)**/{d}/**" for d in sorted(_TOOL_STATE_DIRS)]
         df = subprocess.run(
-            ["git", "-C", str(root), "diff", "HEAD", *pathspec],
+            ["git", "-C", str(root), "diff", "HEAD", *_diff_spec],
             capture_output=True, text=True, timeout=20,
         )
         diff_txt = df.stdout if df.returncode == 0 else ""

{sourcecode-4.5.3 → sourcecode-4.6.0}/src/sourcecode/cli.py

@@ -8853,7 +8853,9 @@ def explain_cmd(
         except Exception:
             cir = ContextGraph.build(file_list, target).cir  # fallback: never break explain
         model = SpringSemanticModel.build(cir)
-        explanation = explain_class(class_name, cir, model).capped(limit)
+        # `root` is what lets the security section read the request-chain rules a
+        # configuration class declares in a DSL rather than in annotations (C3-41).
+        explanation = explain_class(class_name, cir, model, root=target).capped(limit)
     finally:
         _prog.finish()
     if _cc_look is not None:

{sourcecode-4.5.3 → sourcecode-4.6.0}/src/sourcecode/explain.py

@@ -15,6 +15,7 @@ are transitional and tracked for a later phase.
 from __future__ import annotations
 
 from dataclasses import dataclass, field
+from pathlib import Path
 from typing import TYPE_CHECKING, Optional
 
 from sourcecode.caller_metrics import CALLER_METRIC_RECONCILIATION
@@ -397,7 +398,15 @@ def _structural_purpose(
 
 
 def _build_public_methods(class_fqn: str, raw_nodes: list[dict]) -> list[str]:
-    """Return public method names for class_fqn from raw_ir nodes."""
+    """Return public method names for class_fqn from raw_ir nodes.
+
+    C3-41: a factory member (`symbol_kind: "bean"`) is a method, and it is the
+    *whole* API of a configuration class. Excluding the kind reported
+    `public_methods: []` for every `@Configuration` in every repository. Its
+    Java visibility is not the test either — a factory member is called by the
+    container, not by a caller in this repository, so a package-private one is
+    still the class's published surface.
+    """
     prefix = class_fqn + "#"
     methods: list[str] = []
     for node in raw_nodes:
@@ -405,6 +414,11 @@ def _build_public_methods(class_fqn: str, raw_nodes: list[dict]) -> list[str]:
         if not fqn.startswith(prefix):
             continue
         kind = node.get("symbol_kind") or node.get("type") or ""
+        if kind == "bean":
+            name = fqn[len(prefix):]
+            if name and not name.startswith("<"):
+                methods.append(name)
+            continue
         if kind not in ("method", "endpoint", "constructor", ""):
             continue
         modifiers: list[str] = node.get("modifiers") or []
@@ -465,10 +479,45 @@ def _build_callers(
     return [_simple(f) if counts[_simple(f)] == 1 else f for f in fqns]
 
 
-def _build_deps(class_fqn: str, graph: "ContextGraph") -> list[str]:
-    """DI injected dependencies, simple names — via the ContextGraph public API."""
-    deps = graph.injected_dependencies_of(class_fqn)
-    return sorted({_simple(d) for d in deps})
+def _build_deps(
+    class_fqn: str, graph: "ContextGraph", raw_nodes: "Optional[list[dict]]" = None
+) -> list[str]:
+    """What this class depends on, simple names — via the ContextGraph public API.
+
+    Injected dependencies for a class that is wired; **plus what a factory
+    declares**, for a class that does the wiring (C3-41). A configuration class
+    injects nothing: it names its collaborators in the signatures of its factory
+    members and builds them in their bodies, so reading only the injection graph
+    reported `outgoing_deps: []` for a class with eight collaborators per method.
+
+    Evidence, not naming: the extra sources are read only when the class actually
+    declares factory members, and both are atoms the IR already publishes — the
+    class-scope type surface and the instantiation facts.
+    """
+    deps = set(graph.injected_dependencies_of(class_fqn))
+    if _factory_members(class_fqn, raw_nodes or []):
+        deps.update(
+            ref.type for ref in graph.class_type_references_in(class_fqn) if ref.type
+        )
+        for member in _factory_members(class_fqn, raw_nodes or []):
+            deps.update(inst.type for inst in graph.instantiations_in(member) if inst.type)
+    return sorted({_simple(d) for d in deps if d})
+
+
+def _factory_members(class_fqn: str, raw_nodes: list[dict]) -> list[str]:
+    """Members of this class that declare a bean rather than behaviour.
+
+    The IR already classifies them (`symbol_kind: "bean"`); this is the one place
+    that reads it, so both the method list and the dependency list agree about
+    what a configuration class contains.
+    """
+    prefix = class_fqn + "#"
+    return [
+        str(node.get("fqn"))
+        for node in raw_nodes
+        if str(node.get("fqn") or "").startswith(prefix)
+        and (node.get("symbol_kind") or "") == "bean"
+    ]
 
 
 def _build_events_published(class_fqn: str, model: "SpringSemanticModel") -> list[str]:
@@ -566,6 +615,44 @@ def _build_security(
     return result
 
 
+def _build_chain_rules(class_fqn: str, cir: "CanonicalRepositoryIR", root) -> list[str]:
+    """Access rules this class declares for the request chain.
+
+    C3-41. A class that configures the chain writes its constraints in a DSL, not
+    in annotations, so the annotation reader returned `[]` for the one class whose
+    entire job is deciding access — while `posture`, in the same build, read its
+    rules from that very file. One authority answers it (`chain_rules`), so the
+    two commands cannot describe the same file differently.
+    """
+    if root is None:
+        return []
+    from sourcecode.chain_rules import rules_from_source
+
+    rel = _source_file_of(class_fqn, cir)
+    if not rel:
+        return []
+    try:
+        source = (Path(root) / rel).read_text(encoding="utf-8", errors="replace")
+    except OSError:
+        return []
+    out: list[str] = []
+    for rule in rules_from_source(source, rel):
+        patterns = ", ".join(rule.patterns) if rule.patterns else "**"
+        out.append(
+            f"request chain: {patterns} → {rule.decision} "
+            f"({rel}:{rule.line})"
+        )
+    return out
+
+
+def _source_file_of(class_fqn: str, cir: "CanonicalRepositoryIR") -> str:
+    """The repo-relative file declaring this class, from the IR's own node."""
+    for node in _get_raw_nodes(cir):
+        if node.get("fqn") == class_fqn:
+            return str(node.get("file") or node.get("source_file") or "")
+    return ""
+
+
 def _build_endpoints(class_fqn: str, model: "SpringSemanticModel") -> list[str]:
     """REST endpoints declared on this controller class."""
     endpoints = model.endpoint_index.endpoints_for(class_fqn)
@@ -585,13 +672,27 @@ def explain_class(
     class_name: str,
     cir: "CanonicalRepositoryIR",
     model: "SpringSemanticModel",
+    root: "Optional[Path]" = None,
 ) -> ClassExplanation:
     """Build a ClassExplanation for class_name from existing CIR + model.
 
-    Never raises — wraps all derivation in try/except.
+    Never raises. C3-41: a section that *failed* now says so. Every section was
+    wrapped in a bare `except: []`, which made a crash and a measured emptiness
+    the same output — and an empty section reads as *"there are none"*.
     """
     warnings: list[str] = []
 
+    def _section_of(name: str, build, default):
+        """Run one section builder; a failure becomes a warning, not a silence."""
+        try:
+            return build()
+        except Exception as exc:  # pragma: no cover — defensive by contract
+            warnings.append(
+                f"{name} could not be derived ({type(exc).__name__}): this section "
+                f"is empty because the derivation failed, not because there is none"
+            )
+            return default
+
     try:
         class_fqn, all_matches = _resolve_fqn(class_name, cir)
     except Exception:
@@ -630,40 +731,32 @@ def explain_class(
     except Exception:
         purpose = f"{stereotype} class"
 
-    try:
-        public_methods = _build_public_methods(class_fqn, raw_nodes)
-    except Exception:
-        public_methods = []
-
-    try:
-        incoming_callers = _build_callers(class_fqn, cir, graph)
-    except Exception:
-        incoming_callers = []
-
-    try:
-        outgoing_deps = _build_deps(class_fqn, graph)
-    except Exception:
-        outgoing_deps = []
-
-    try:
-        events_published = _build_events_published(class_fqn, model)
-    except Exception:
-        events_published = []
-
-    try:
-        events_consumed = _build_events_consumed(class_fqn, model)
-    except Exception:
-        events_consumed = []
-
-    try:
-        transactions = _build_transactions(class_fqn, model)
-    except Exception:
-        transactions = []
-
-    try:
-        security_constraints = _build_security(class_fqn, raw_nodes, cir)
-    except Exception:
-        security_constraints = []
+    public_methods = _section_of(
+        "public_methods", lambda: _build_public_methods(class_fqn, raw_nodes), []
+    )
+    incoming_callers = _section_of(
+        "incoming_callers", lambda: _build_callers(class_fqn, cir, graph), []
+    )
+    outgoing_deps = _section_of(
+        "outgoing_deps", lambda: _build_deps(class_fqn, graph, raw_nodes), []
+    )
+    events_published = _section_of(
+        "events_published", lambda: _build_events_published(class_fqn, model), []
+    )
+    events_consumed = _section_of(
+        "events_consumed", lambda: _build_events_consumed(class_fqn, model), []
+    )
+    transactions = _section_of(
+        "transactions", lambda: _build_transactions(class_fqn, model), []
+    )
+    security_constraints = _section_of(
+        "security_constraints", lambda: _build_security(class_fqn, raw_nodes, cir), []
+    )
+    # A class that configures the request chain declares its constraints in a
+    # DSL. Same section, one authority (`chain_rules`) — never a second parser.
+    security_constraints = security_constraints + _section_of(
+        "security_constraints", lambda: _build_chain_rules(class_fqn, cir, root), []
+    )
 
     try:
         rest_endpoints = _build_endpoints(class_fqn, model)