refine: collection-time test skip markers

- Replace runtime pytest.skip() calls inside test/fixture bodies with module-level capability probes gated by pytest.mark.skipif, across benchmarks/, cc_develop, cc_foundry, cc_oss, cc_research, codemap-py, and codex-rig; codify the rule in AGENTS.md and CLAUDE.md ("Test skips are collection-time decorators only ... never call pytest.skip() from a test or fixture body")
- Remove now-redundant portability micro-tests (test_no_tmp_literal_in_source, test_stdout_reconfigure_present, test_shebang_env_python) after widening .pre-commit-config.yaml's no-tmp-literal check to also cover bin/ scripts (previously hooks/ only), and swap the resulting /tmp/... doctest literals for /opt/... in cc_oss, cc_research, and codemap-py bin scripts
- cc_oss: consolidate 13 near-identical per-script --help-exits-0 tests into one parametrized tests/test_bin_help_contract.py that walks bin/ dynamically
- cc_foundry: merge the hyphenated tests/test_find-polluter.py into tests/test_find_polluter.py, keeping one canonical CLI/security test module
- codex-rig: sync_codex.py now reconciles the managed plugin set from a configured local (non-Git) marketplace instead of forcing a Git-only marketplace upgrade; extract shared tests/_platform.py capability probes (POSIX bash, symlinks, FIFO, mode retention) replacing duplicated per-file probes
- Rewrite README.md, docs/index.md, and every plugin README (foundry, oss, develop, research, codemap-py, codex-rig) for accuracy and discoverability: correct stale agent/skill counts and command examples, remove benchmark-result duplication in favor of the canonical benchmarks/README.md, and expand docs/llms.txt and docs/robots.txt for AI-crawler discoverability
- Patch-bump plugin.json for develop, foundry, oss, research, and codemap-py (both manifests), and codex-rig (0.7.5 to 0.7.6), syncing descriptions to the current skill/agent rosters; sync marketplace.json and docs/hooks.py structured data to match

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Author: Borda

.claude-plugin/marketplace.json

@@ -1,31 +1,32 @@
 {
+  "description": "Independently installable Claude Code plugins for configuration, Python development, open-source maintenance, ML research, and static Python structure.",
   "name": "borda-ai-rig",
   "owner": {
     "name": "Borda"
   },
   "plugins": [
     {
-      "description": "Production-grade Claude Code config: specialist agents, slash-command skills, and event-driven hooks for Python/ML OSS development",
+      "description": "Claude Code configuration and workflow maintenance: 11 skills, 10 specialist agents, rules, hooks, audit, calibration, and session handover",
       "name": "foundry",
       "source": "./plugins/cc_foundry"
     },
     {
-      "description": "OSS maintenance skills: analyse, resolve, review, release",
+      "description": "Open-source maintenance: issue and repository analysis, PR review, feedback resolution, release-readiness assessment, setup, and 4 specialist agents",
       "name": "oss",
       "source": "./plugins/cc_oss"
     },
     {
-      "description": "Implementation workflow skills: feature, fix, refactor, plan, debug",
+      "description": "Validate-first Python development: plan, feature, fix, refactor, debug, review, and setup workflows",
       "name": "develop",
       "source": "./plugins/cc_develop"
     },
     {
-      "description": "ML/perf investigation skills: topic, plan, judge, run, sweep",
+      "description": "Reviewable ML research: literature, planning, methodology review, bounded runs, sweeps, implementation verification, ablations, retrospectives, Kaggle, and setup",
       "name": "research",
       "source": "./plugins/cc_research"
     },
     {
-      "description": "Codebase structural index \u2014 scan Python projects, query the import graph for blast-radius and coupling",
+      "description": "Static Python structure for Claude Code: indexing, queries, test impact, reference renames, integration, and session telemetry debriefs",
       "name": "codemap-py",
       "source": "./plugins/codemap-py"
     }

.claude/README.md

@@ -304,7 +304,10 @@ repos:
         entry: '["''`]/tmp/'
         # Replaces tests/hooks/tmp-paths.test.js, which checked a hardcoded 6-file cc_foundry list
         # and missed the other 16 plugin hooks; this glob picks up new hooks automatically.
-        files: ^plugins/[^/]+/hooks/.*\.(js|py)$|^\.claude/hooks/.*\.(js|py)$|^\.codex/hooks/.*\.(js|py)$
+        # bin/ included for the same reason: the per-script test_no_tmp_literal_in_source copies
+        # covered 8 of 124 bin/ scripts. Doctests must avoid quoted /tmp/ too — a doctest's
+        # expected-output line is indistinguishable from a real path constant to a line matcher.
+        files: ^plugins/[^/]+/(hooks|bin)/.*\.(js|py)$|^\.claude/hooks/.*\.(js|py)$|^\.codex/hooks/.*\.(js|py)$
 
       - id: check-benchmark-portable-paths
         name: 🧳 No hardcoded machine paths in benchmark source

.codex/README.md

@@ -29,6 +29,7 @@ Scripts, hooks, `bin/` entry points, and CI steps all run on Linux, macOS, and n
 - A workflow `run:` step invoking `.sh` needs explicit `shell: bash` — the Windows default shell dot-sources it and exits zero, a false green.
 - Symlinks, file modes, and uid checks are capabilities: degrade in production code first.
 - Skips are the last resort: never a blanket `skipif(sys.platform == "win32")`, always a capability probe skipping on `OSError`, with each surviving skip documented and re-audited.
+- Test skips must be collection-time decorators (`pytest.mark.skipif`, `pytest.mark.skip`, or parametrized marks); never call `pytest.skip()` from a test or fixture body.
 - Green macOS is absence of regression, not Windows support: prove Windows semantics with `PureWindowsPath` or `ntpath`, since monkeypatching `os.name` does not change `pathlib`.
 
 ## Benchmark Isolation

.pre-commit-config.yaml

@@ -53,6 +53,7 @@ Scripts, hooks, `bin/`, CI steps all run Linux + macOS + native Windows. Fix at
 - CI `run:` calling `.sh` needs explicit `shell: bash` — Windows pwsh dot-sources it, exits 0, runs nothing (false green)
 - Symlink/mode/uid = capabilities: degrade in production code first
 - Skip last resort: never blanket `skipif(sys.platform == "win32")` — probe capability, skip on `OSError`; document + re-audit each surviving skip
+- Test skips are collection-time decorators only (`pytest.mark.skipif`, `pytest.mark.skip`, parametrized marks); never call `pytest.skip()` from a test or fixture body
 - Green macOS ≠ Windows support: prove with `PureWindowsPath`/`ntpath`; monkeypatching `os.name` does not change `pathlib`
 
 ## Benchmark Isolation

AGENTS.md

@@ -858,7 +858,7 @@ Omitting `--tasks` runs all 73 locked tasks (219 A/B/C cells); use an explicit f
 
 Rows below use the canonical arm name with the retired one in parentheses: `C_strict` was called `C_required` when this run was frozen, and `benchmarks/results/code-2026-08-04.json` still records the old string — search that artifact for `C_required`, not `C_strict`. The [agentic measurement caveats](#agentic-measurement-caveats) apply to this table: the arms ran in fixed A→B→C order against a warm provider cache, and the `erec`/`rrec`/`aqs` columns carry the substring-credit and unweighted-mean inflation described there.
 
-<!-- result-sync: duplicated/summarized in ../plugins/codemap-py/README.md#claude-agentic-2026-08-04; update both files or record an explicit divergence note. -->
+<!-- result-sync: ../plugins/codemap-py/README.md#claude-agentic-2026-08-04 is an anchor stub only — the plugin README intentionally no longer duplicates run-specific tables (§Benchmark Isolation). This table is canonical; do not re-add values there. -->
 
 | Model  | Arm                       |   n |     in tok |  out tok |    cost $ | elapsed s |     erec |     rrec |      aqs |  correct |
 | ------ | ------------------------- | --: | ---------: | -------: | --------: | --------: | -------: | -------: | -------: | -------: |
@@ -1088,7 +1088,7 @@ Full summary + per-task reading: [`results/bench-summary-2026-07-29.md`](results
 
 **Two value axes — read separately, never blended.** (1) **Reliability/quality**: safety-grade + structural recall — codemap **13/13 safety-grade every tier** vs plain 8/13 → 12/13 → 13/13; the primary proposition, holds up-tier. (2) **Economy (cost/tokens/time)**: read at *matched caller fan-in* — the win grows with fan-in (cost 0.35× haiku / 0.54× opus on high-fan-in tasks); raw median token ratio is a *secondary, caveated* number that → 1 as models get terser. Accuracy Δ is a saturation-sensitive tie-breaker, not a headline.
 
-<!-- result-sync: this is the canonical July 29 table; ../plugins/codemap-py/README.md#three-model-comparison contains the distinct June 22 legacy run, so do not synchronize their values. -->
+<!-- result-sync: this is the canonical July 29 table. ../plugins/codemap-py/README.md#three-model-comparison is now an anchor stub only — the June 22 legacy run it used to hold was removed from the plugin README (§Benchmark Isolation). The June 22 table below remains the canonical copy; the two runs are still distinct and must not be blended. -->
 
 | Tier      | Plain accuracy    | Codemap accuracy  | Δ accuracy | Safety-grade plain | Safety-grade codemap | Token× med / mean | Cost× med / mean |
 | --------- | ----------------- | ----------------- | ---------- | ------------------ | -------------------- | ----------------- | ---------------- |
@@ -1113,7 +1113,7 @@ Per-workflow codemap accuracy: query (n=28) 92.0 / 95.8 / 84.0%; debug (n=6) 100
 
 Results — June 22 2026 — 44 tasks × 2 arms × 3 models, pytorch-lightning-master. **Models** `claude-haiku-4-5` / `claude-sonnet-4-6` / `claude-opus-4-6` · **codemap version not recorded in these result lines** (predates the v0.13.2 agentic run; ~mid-June 2026) · **codemap-favoring steering harness** (not comparable with the July run above — see comparability note).
 
-<!-- result-sync: duplicated/summarized in ../plugins/codemap-py/README.md#three-model-comparison; update both files or record an explicit divergence note. All result tables in this historical block are canonical sources for downstream summaries. -->
+<!-- result-sync: ../plugins/codemap-py/README.md#three-model-comparison is an anchor stub only — the plugin README no longer duplicates run-specific tables (§Benchmark Isolation). All result tables in this historical block are the canonical sources for downstream summaries; do not re-add values to the plugin README. -->
 
 | Model      | Plain accuracy    | Codemap accuracy  | Accuracy lift | Safety-grade plain | Safety-grade codemap | Token ratio (median) | Token ratio range |
 | ---------- | ----------------- | ----------------- | ------------- | ------------------ | -------------------- | -------------------- | ----------------- |

CLAUDE.md

@@ -0,0 +1,19 @@
+"""Capability checks for benchmark launcher artifacts."""
+
+from __future__ import annotations
+
+import subprocess
+from pathlib import Path
+
+
+def raw_codemap_launchers_are_runnable(repo_root: Path) -> bool:
+    """Return whether both shipped raw Codemap launchers can run on this host."""
+    for name in ("scan-index", "scan-query"):
+        launcher = repo_root / "plugins" / "codemap-py" / "bin" / name
+        try:
+            result = subprocess.run([str(launcher), "--help"], capture_output=True, text=True, timeout=30, check=False)
+        except (OSError, subprocess.SubprocessError):
+            return False
+        if result.returncode != 0:
+            return False
+    return True

README.md

@@ -17,6 +17,13 @@
 
 BENCHMARKS_DIR = Path(__file__).parent.parent
 REPO_ROOT = BENCHMARKS_DIR.parent
+TESTS_DIR = Path(__file__).parent
+if str(TESTS_DIR) not in sys.path:
+    sys.path.insert(0, str(TESTS_DIR))
+PYTORCH_LIGHTNING_REPO = Path(os.environ.get("PL_REPO_PATH", str(REPO_ROOT / ".sandbox" / "pytorch-lightning")))
+PYTORCH_LIGHTNING_INDEXES = (
+    tuple(sorted(PYTORCH_LIGHTNING_REPO.rglob(".cache/codemap/*.json"))) if PYTORCH_LIGHTNING_REPO.exists() else ()
+)
 
 
 def _load_module(module_name: str, filename: str):
@@ -98,18 +105,13 @@ def script_gen_real_issues():
 def scan_query_binary() -> Path:
     """Path to scan-query binary; fails on POSIX when the tracked binary is absent.
 
-    Windows skips (the harness exercises the POSIX launcher only). On POSIX the
-    binary is a tracked file, so a missing path means a broken checkout, not an
-    unavailable dependency — fail loudly instead of yielding a false-green skip.
+    The enclosing tests use a collection-time launchability marker. Once selected,
+    a missing tracked binary means a broken checkout, so fail loudly rather than
+    yielding a false-green skip.
 
     Returns:
         Absolute path to the scan-query executable.
     """
-    if sys.platform == "win32":
-        pytest.skip(
-            "benchmark harness exercises the POSIX launcher only; "
-            "codemap-py Windows coverage lives in plugins/codemap-py tests"
-        )
     binary = REPO_ROOT / "plugins" / "codemap-py" / "bin" / "scan-query"
     if not binary.exists():
         pytest.fail(f"tracked scan-query binary missing at {binary} — broken checkout")
@@ -120,18 +122,13 @@ def scan_query_binary() -> Path:
 def scan_index_binary() -> Path:
     """Path to scan-index binary; fails on POSIX when the tracked binary is absent.
 
-    Windows skips (the harness exercises the POSIX launcher only). On POSIX the
-    binary is a tracked file, so a missing path means a broken checkout, not an
-    unavailable dependency — fail loudly instead of yielding a false-green skip.
+    The enclosing tests use a collection-time launchability marker. Once selected,
+    a missing tracked binary means a broken checkout, so fail loudly rather than
+    yielding a false-green skip.
 
     Returns:
         Absolute path to the scan-index executable.
     """
-    if sys.platform == "win32":
-        pytest.skip(
-            "benchmark harness exercises the POSIX launcher only; "
-            "codemap-py Windows coverage lives in plugins/codemap-py tests"
-        )
     binary = REPO_ROOT / "plugins" / "codemap-py" / "bin" / "scan-index"
     if not binary.exists():
         pytest.fail(f"tracked scan-index binary missing at {binary} — broken checkout")
@@ -142,7 +139,7 @@ def scan_index_binary() -> Path:
 def sample_repo(tmp_path_factory: pytest.TempPathFactory, scan_index_binary: Path) -> tuple[Path, Path]:
     """Clone psf/requests (shallow) and build a codemap index.
 
-    Skips if git clone fails (no network access).
+    Fails if clone or indexing fails after collection.
 
     Args:
         tmp_path_factory: pytest factory for session-scoped temp directories.
@@ -160,7 +157,7 @@ def sample_repo(tmp_path_factory: pytest.TempPathFactory, scan_index_binary: Pat
             timeout=120,
         )
     except (subprocess.CalledProcessError, subprocess.TimeoutExpired) as exc:
-        pytest.skip(f"git clone failed (no network?): {exc}")
+        pytest.fail(f"git clone failed (no network?): {exc}")
 
     try:
         subprocess.run(
@@ -171,35 +168,32 @@ def sample_repo(tmp_path_factory: pytest.TempPathFactory, scan_index_binary: Pat
             cwd=str(clone_dir),
         )
     except subprocess.CalledProcessError as exc:
-        pytest.skip(f"scan-index failed: {exc.stderr.decode()}")
+        pytest.fail(f"scan-index failed: {exc.stderr.decode()}")
 
     index_candidates = sorted(clone_dir.rglob(".cache/codemap/*.json"))
     if not index_candidates:
-        pytest.skip("scan-index ran but produced no index file")
+        pytest.fail("scan-index ran but produced no index file")
 
     return clone_dir, index_candidates[0]
 
 
 @pytest.fixture(scope="session")
 def pytorch_lightning_repo() -> Path:
-    """Path to a pytorch-lightning checkout; skips if not found.
+    """Path to a decorator-validated pytorch-lightning checkout.
 
     Checks ``PL_REPO_PATH`` env var first, then the pinned in-project clone
     ``.sandbox/pytorch-lightning`` (created by ``run-all.sh``).
 
     Returns:
         Absolute path to the pytorch-lightning repository root.
     """
-    default = Path(__file__).resolve().parents[2] / ".sandbox" / "pytorch-lightning"
-    repo = Path(os.environ.get("PL_REPO_PATH", str(default)))
-    if not repo.exists():
-        pytest.skip(f"pytorch-lightning repo not found at {repo}; set PL_REPO_PATH env var")
-    return repo
+    assert PYTORCH_LIGHTNING_REPO.is_dir(), "pytorch-lightning test lacks checkout marker"
+    return PYTORCH_LIGHTNING_REPO
 
 
 @pytest.fixture(scope="session")
 def pytorch_lightning_index(pytorch_lightning_repo: Path, scan_query_binary: Path) -> Path:
-    """Pre-built codemap index for pytorch-lightning; skips if not found.
+    """Decorator-validated pre-built Codemap index for pytorch-lightning.
 
     The index is expected at ``.cache/codemap/<repo-name>.json`` inside the
     repo.  Run ``scan-index --root <repo>`` once to create it.
@@ -212,9 +206,5 @@ def pytorch_lightning_index(pytorch_lightning_repo: Path, scan_query_binary: Pat
         Absolute path to the JSON index file.
     """
     _ = scan_query_binary  # ensure binary fixture resolved
-    candidates = sorted(pytorch_lightning_repo.rglob(".cache/codemap/*.json"))
-    if not candidates:
-        pytest.skip(
-            f"No codemap index found under {pytorch_lightning_repo}/.cache/codemap/; run: scan-index --root <repo>"
-        )
-    return candidates[0]
+    assert PYTORCH_LIGHTNING_INDEXES, "pytorch-lightning index test lacks index marker"
+    return PYTORCH_LIGHTNING_INDEXES[0]