Yuu6798 · Yuu6798 · Jun 12, 2026 · Jun 12, 2026 · Jun 12, 2026 · Jun 12, 2026
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -175,7 +175,7 @@ Status legend: **ACTIVE** (current spec/contract, AI agents should read first) /
 | `docs/exit_codes.md` | ACTIVE | Stable CLI exit code policy (0 / 1 / 2 / 3 / 4) for CI integration, including `--strict-repair`, `severity: info` Advisor channel, and per-subcommand notes |
 | `docs/json_schema.md` | ACTIVE | CLI JSON envelopes — verdict / compile at `schema_version="6"`, compile-repair at independent `schema_version="1"`, validate-plan at independent `schema_version="2"`. Includes compatibility policy and v2→v3 through v5→v6 diffs (plus validate-plan v1→v2) |
 | `docs/cli_test_inventory.md` | ACTIVE | CLI test coverage inventory, runtime notes, and conservative reduction candidates |
-| `docs/target_yaml_guide.md` | ACTIVE | `target.yaml` authoring guide. Practical companion to `design.md §4` / `cli_usage.md`. Centralises authoring hazards D1 (`--package-root` scope vs `tests/` visibility), D3 (template / user constraint duplication), D4 (config-only PR の vacuous PASS) — 2026-05-07 Session 4 dogfooding 由来 |
+| `docs/target_yaml_guide.md` | ACTIVE | `target.yaml` authoring guide. Practical companion to `design.md §4` / `cli_usage.md`. Centralises authoring hazards D1 (`--package-root` scope vs `tests/` visibility), D3 (template / user constraint duplication), D4 (config-only PR の vacuous PASS), D6 (nested-function complexity 盲点) — 2026-05-07 Session 4 / 2026-05-28 real-PR dogfooding 由来 |
 | `docs/target_authoring_surface.md` | ACTIVE | Authoring surface 設計契約 (Brief 8 / CSCI-41)。target.yaml は hand-written 必須でない / 生成経路 3 通り (recipe + sources / catalog 参照 / hand-written) / LLM 経路は Brief 8b 分離 / 全経路は verdict 前に declared intent として固定 / Authoring・Advisor・Provenance surface は evaluator 不可参照 / `candidate_code_used: false` 固定。§23.3.1 の実装側 catch-up |
 | `docs/ssp_protocol.md` | ACTIVE | SSP v0.1 normative spec: SensorOutput / Finding / SSPDelta / SSPVerdict definitions, 5-element SAST + 3-element SCA fingerprint, Python profile AST normalization, delta computation, verdict precedence (`unknown > fail > pass`), JSON Schema artifact, Sensor Provenance Invariant (§23.1 mirror), determinism requirements, core isolation contract |
 | `docs/ssp_usage_guide.md` | ACTIVE | SSP practical usage guide: quick start (Semgrep SAST / pip-audit SCA / fixture mode), output formats (JSON / human / SARIF), CI integration (GitHub Actions workflow / exit code routing / fixture-based CI), hand-built SensorOutput examples, delta computation overview, relationship to core Semantic CI |

diff --git a/docs/cli_usage.md b/docs/cli_usage.md
@@ -738,7 +738,7 @@ semantic-ci target-doctor [--target <yaml>] [--package-root <dir>]
                           [--format {human,json}] [--output <file>]
 ```
 
-Audits a `target.yaml` for seven authoring hazards and renders them as
+Audits a `target.yaml` for eight authoring hazards and renders them as
 advisories. Advisor surface — advisory presence does not change the
 verdict and does not change the exit code (`docs/exit_codes.md`).
 
@@ -747,6 +747,7 @@ verdict and does not change the exit code (`docs/exit_codes.md`).
 | `ADVISORY-D1` | `test_surface_delta.*` constraint exists, but no test files (`test_*.py` / `*_test.py` / `tests/`) are visible under `--package-root`. |
 | `ADVISORY-D3` | A user constraint duplicates a template-expanded constraint (same kind/target/operator/expected). |
 | `ADVISORY-D4` | The target is lock-only and the candidate diff (`--baseline-rev` ↔ `--candidate-rev`) touches no Python files; the verdict would be a vacuous PASS. Skipped silently when neither rev is given and git is unavailable. |
+| `ADVISORY-D6` | The target declares a `complexity_delta` constraint and the candidate diff (`--baseline-rev` ↔ `--candidate-rev`) grows the nested-function count in an in-scope Python file. The complexity extractor does not descend into nested defs, so a reported cyclomatic/cognitive decrease may be displacement, not simplification. Skipped silently when neither rev is given and git is unavailable. |
 | `ADVISORY-I1` | `intent` is the empty string. Repair adapters and `validate-plan` produce better guidance when intent describes the change purpose; use `init --intent` or edit `target.yaml`. |
 | `ADVISORY-P1` | `primary_kind: feature` has no positive addition constraint. |
 | `ADVISORY-P2` | `primary_kind: bugfix` has no `test_surface_delta.new_cases` expectation. |

diff --git a/docs/dogfooding_findings_tracker.md b/docs/dogfooding_findings_tracker.md
@@ -22,9 +22,9 @@ Status taxonomy:
 | D3 | 2026-05-07 Session 4 dogfood | template / user constraint duplication | user constraint shadows template default with identical semantics, doubles evidence | **解決** | `docs/target_yaml_guide.md` Hazard 2 + `ADVISORY-D3` detector |
 | D4 | 2026-05-07 Session 4 dogfood | vacuous PASS (out-of-scope diff) | config-only PR produces empty Python `CodeStateDelta`; lock-only template constraints pass without exercising the actual change | **解決** | `docs/target_yaml_guide.md` Hazard 3 + `ADVISORY-D4` detector |
 | D5 | 2026-05-07 Session 5 dogfood (FINDING-1) | set operator partial-match semantics | partial-dict `expected` items canonicalised as different elements from full extractor records — false positive on `includes_*` / `subset_of` / `superset_of`, **false negative (CI bypass) on `excludes_all`** | **解決** | PR #65 (CSCI-35c) — Match Schema partial-record matching + flat-projection aliases + `evidence.matched`; schema_version v4→v5 |
-| D6 | 2026-05-28 real-PR complexity dogfood (FINDING-F1) | vacuous PASS (extractor coverage gap) — **重複・関連 = sibling of D4** | nested function bodies are excluded from `ComplexityEntry` by `python_complexity_extractor` spec (`api_surface` parity); refactor that nests outer-function body into nested helpers reports large CC drop while real complexity is unchanged | **未解決** | Candidate paths: (a) `docs/target_yaml_guide.md` new Hazard 4 + `ADVISORY-D6` detector mirroring D4; (b) extractor spec change to emit nested-function entries (long-term, schema-impacting). Reproduction: langgraph PR #3700 (8/1 vacuous PASS in real-PR pass) |
+| D6 | 2026-05-28 real-PR complexity dogfood (FINDING-F1) | vacuous PASS (extractor coverage gap) — **重複・関連 = sibling of D4** | nested function bodies are excluded from `ComplexityEntry` by `python_complexity_extractor` spec (`api_surface` parity); refactor that nests outer-function body into nested helpers reports large CC drop while real complexity is unchanged | **解決** | `docs/target_yaml_guide.md` Hazard 4 + `ADVISORY-D6` detector (`authoring/hazards.py::detect_d6`, candidate path (a), mirrors D4's diff-aware contract) — fires when a verdict-participating `complexity_delta` constraint meets nested-def growth in the in-scope candidate diff (`--baseline-rev` ↔ `--candidate-rev`); growth signal computed per file via `authoring/nested_defs.py`, syntax-error sides skipped fail-silent. Path (b) (extractor emits nested-function entries) remains a long-term, schema-impacting option, not pursued. Reproduction: langgraph PR #3700 (8/1 vacuous PASS in real-PR pass) |
 | D7 | 2026-05-28 real-PR complexity dogfood (FINDING-F2) | authoring mismatch (operator / constraint pairing) | `extract-method` refactor is mathematically guaranteed to **micro-increase cyclomatic** (each extracted function adds base 1), even with `_` prefix discipline and api_surface preserved. Cognitive is the metric that drops. Authors declaring `complexity_delta.cyclomatic ≤ 0` for extract-method refactors hit a structural false-FAIL | **未解決** | Candidate paths: (a) authoring guide section "Choosing complexity metric per refactor pattern" recommending `cognitive_delta` for extract-method; (b) future `ADVISORY-D7` detector emitted when a `change.primary_kind=refactor` target uses `cyclomatic_delta ≤ 0` and the diff matches extract-method shape. Low priority: this is authoring advice, not a CI integrity hazard |
-| D8 | 2026-06-07 scale + security dogfood (SCA gap) | SCA sensor dependency-source discovery gap | SSP SCA auto-discovery (`_requirements_file` in `src/semantic_ci_code/cli/commands/ssp.py`) only found `requirements.txt` at repo root; the `--locked` fallback only accepted `pylock.toml` / requirements lockfiles. PEP 621 pyproject-only projects (litellm) and `pdm.lock` projects (pdm) declared deps in unrecognised formats → `pip-audit --locked .` errors "no lockfiles found" → empty JSON → adapter degraded to `unknown` (exit 3). Correct graceful degradation (no silent false PASS, honours `unknown > fail > pass`) but a real usability gap that blocked SCA on most modern Python projects | **解決** | CSCI-55 — dependency source discovery now recognises `requirements.txt`, `pylock.toml` / `pylock.*.toml`, `uv.lock`, `pdm.lock`, `poetry.lock`, and static PEP 621 `[project].dependencies`; lock sources are converted deterministically to pinned temp requirements, optional/non-default-group/marker-inactive packages are filtered, and malformed recognized sources fail closed to SSP `unknown` |
+| D8 | 2026-06-07 scale + security dogfood (SCA gap) | SCA sensor dependency-source discovery gap | SSP SCA auto-discovery (`_requirements_file` in `src/semantic_ci_code/cli/commands/ssp.py`) only found `requirements.txt` at repo root; the `--locked` fallback only accepted `pylock.toml` / requirements lockfiles. PEP 621 pyproject-only projects (litellm) and `pdm.lock` projects (pdm) declared deps in unrecognised formats → `pip-audit --locked .` errors "no lockfiles found" → empty JSON → adapter degraded to `unknown` (exit 3). Correct graceful degradation (no silent false PASS, honours `unknown > fail > pass`) but a real usability gap that blocked SCA on most modern Python projects | **解決** | CSCI-55 / PR #151 — dependency source discovery now recognises `requirements.txt`, `pylock.toml` / `pylock.*.toml`, `uv.lock`, `pdm.lock`, `poetry.lock`, and static PEP 621 `[project].dependencies`; lock sources are converted deterministically to pinned temp requirements, optional/non-default-group/marker-inactive packages are filtered, and malformed recognized sources fail closed to SSP `unknown` |
 
 ## Reading order
 
@@ -37,8 +37,8 @@ Status taxonomy:
 ## Classification at a glance
 
 - **重複・関連 pairs**: D4 ↔ D6 (both are "vacuous PASS" via extractor coverage gap, distinct mechanism — D4 is "diff outside Python scope", D6 is "diff inside scope but inside nested function")
-- **解決 (6 of 8)**: D1, D2, D3, D4, D5, D8
-- **未解決 (2 of 8)**: D6 (mitigation path open), D7 (authoring advice, low priority)
+- **解決 (7 of 8)**: D1, D2, D3, D4, D5, D6, D8
+- **未解決 (1 of 8)**: D7 (authoring advice, low priority)
 - **observation-only (not a D#)**: F6 (pattern-SAST logic-vuln blindspot) — **UNTESTED HYPOTHESIS, not a demonstrated observation in the 2026-06-07 pass**: the Semgrep registry rulesets returned HTTP 403, so Semgrep ran with 0 loaded rules over 0 paths and produced no valid SAST measurement. F6 records the *a-priori* expectation that deterministic SAST misses semantic / business-logic vulns, cross-linked to Phase H (`docs/llm_sensor_adapter_planning.md`) as **motivation** — it is **not** empirically validated by this pass. Recorded in `docs/dogfooding_scale_and_security.md` (which now carries a validity warning + repro note for redoing the SAST sub-pass under a network policy allowing `semgrep.dev`). Distinct from the demonstrated observations of the same pass: real vulns merged-then-fixed (git evidence) and SCA clean-on-litellm (pip-audit positive-controlled with `jinja2==2.11.2` → 5 CVEs)
 
 ## Source pass index

diff --git a/docs/exit_codes.md b/docs/exit_codes.md
@@ -78,8 +78,8 @@ git errors (`CompileError` on `target.yaml`, git revision resolution
 failure when `--baseline-rev` / `--candidate-rev` is given, git
 unavailable when explicitly required) exit 3. Internal bugs exit 4. When
 neither `--baseline-rev` nor `--candidate-rev` is given and git is
-unavailable or no baseline can be resolved, ADVISORY-D4 is silently
-skipped rather than failing. There is no `--strict-advice` flag; CI that
+unavailable or no baseline can be resolved, ADVISORY-D4 and ADVISORY-D6
+are silently skipped rather than failing. There is no `--strict-advice` flag; CI that
 wants to gate on advisory presence should consume `--format json` and
 apply a workflow-level policy. Silent success on bad input is forbidden —
 the advisor surface only suppresses the verdict step, not the input

diff --git a/docs/json_schema.md b/docs/json_schema.md
@@ -318,12 +318,12 @@ compile-repair, or validate-plan envelopes. The shape is pinned by
 |---|---|
 | `schema_version` | Always `"advisory-1"`. |
 | `subcommand` | Always `"target-doctor"`. |
-| `advisories[].code` | One of `ADVISORY-D1`, `ADVISORY-D3`, `ADVISORY-D4`, `ADVISORY-I1`, `ADVISORY-P1`, `ADVISORY-P2`, `ADVISORY-S1`. |
+| `advisories[].code` | One of `ADVISORY-D1`, `ADVISORY-D3`, `ADVISORY-D4`, `ADVISORY-D6`, `ADVISORY-I1`, `ADVISORY-P1`, `ADVISORY-P2`, `ADVISORY-S1`. |
 | `advisories[].severity` | Always `"info"` — the Advisor surface never participates in the verdict (`docs/code_semantic_ci_design.md §23.3.1`). |
 | `advisories[].message` | Human-readable explanation of the hazard. |
 | `advisories[].evidence` | Per-advisory diagnostic fields (e.g. `constraint_id`, `target`, `package_root`, `files_touched_count`). |
 
-Advisories are emitted in canonical order (D1 → D3 → D4 → I1 → P1 → P2 → S1)
+Advisories are emitted in canonical order (D1 → D3 → D4 → D6 → I1 → P1 → P2 → S1)
 with `constraint_id` as the within-code tiebreak so output is byte-identical
 across runs. Advisory presence does not change the exit code — see
 `docs/exit_codes.md`.

diff --git a/docs/target_yaml_guide.md b/docs/target_yaml_guide.md
@@ -281,6 +281,41 @@ non-Python artifacts (see `design.md §13` on out-of-scope items). It is
 `semantic-ci target-doctor` (Brief 8 / CSCI-43) detects this hazard as
 `ADVISORY-D4` once it lands.
 
+## Hazard 4 — Nested functions are invisible to complexity numbers (D6)
+
+`python_complexity_extractor` only emits entries for the `api_surface`
+parity subset: module-level functions and direct methods of module-level
+classes. When it walks a function body it **stops descent at nested
+`def` boundaries** — a nested helper contributes 0 to the enclosing
+function's cyclomatic/cognitive number, and the helper itself is never
+emitted as its own entry.
+
+The consequence for `complexity_delta` constraints is the D6 sibling of
+Hazard 3's vacuous PASS (observed on a real PR in
+`docs/dogfooding_real_pr_complexity.md` FINDING-F1): a refactor that moves
+an outer function's body into function-nested helpers reports a large
+complexity *drop* while the real complexity is unchanged — it has only
+been displaced below the extractor's horizon. A lock like
+`complexity_delta.cyclomatic less_than_or_equal 0` passes, and the verdict
+silently endorses the displacement.
+
+**What this means for authors:**
+
+- A green complexity verdict on a refactor that introduced nested helpers
+  is weaker evidence than the number suggests. Check whether the helpers
+  should be **module-level functions** (possibly `_`-prefixed) — those are
+  extracted, complexity-counted, and constrained.
+- This is the extractor behaving as specified (`api_surface` emission
+  parity, `design.md` CSCI-8), not a counting bug. The blind spot is the
+  declared trade-off for a deterministic, owned formula.
+
+`semantic-ci target-doctor` detects this hazard as `ADVISORY-D6` when
+`--baseline-rev` / `--candidate-rev` are given: it fires when the target
+declares a verdict-participating `complexity_delta` constraint and the
+candidate diff grows the nested-def count in an in-scope Python file.
+Growth is a heuristic displacement signal, not proof — review the diff
+rather than treating the advisory as a verdict.
+
 ## Constraint Authoring Tips
 
 ### Pick `kind` deliberately

diff --git a/src/semantic_ci_code/authoring/advisory.py b/src/semantic_ci_code/authoring/advisory.py
@@ -7,6 +7,7 @@
     "ADVISORY-D1",
     "ADVISORY-D3",
     "ADVISORY-D4",
+    "ADVISORY-D6",
     "ADVISORY-I1",
     "ADVISORY-P1",
     "ADVISORY-P2",