refactor(doctor): T3 — 工具/运行时体检面进 Rust doctor bootstrap (#48)

.github/workflows/ci.yml

@@ -200,7 +200,9 @@ jobs:
 
       - name: Doctor
         shell: pwsh
-        run: .\archive\check-code-intel-tools.ps1 -RepoPath . -RequireRepowise:$false
+        # T3 (#48): the tool/runtime probe is native now; the PowerShell entry
+        # point is a thin forwarder onto this same subcommand.
+        run: code-intel doctor bootstrap --repo-path . --no-require-repowise
 
       - name: GitHub research contract tests
         shell: pwsh
@@ -530,7 +532,8 @@ jobs:
 
       - name: Doctor
         shell: pwsh
-        run: ./archive/check-code-intel-tools.ps1 -RepoPath . -RequireRepowise:$false -Json
+        # T3 (#48): native probe; the PowerShell entry point now forwards here.
+        run: code-intel doctor bootstrap --repo-path . --no-require-repowise --json
 
       - name: Hospital fail-closed contract tests
         shell: pwsh

.github/workflows/release.yml

@@ -133,8 +133,9 @@ jobs:
       - name: Doctor
         shell: pwsh
         # Runners don't install repowise; smoke runs on the lite fallback
-        # (same as ci.yml). RequireRepowise defaults to $true since 0.2.0.
-        run: .\archive\check-code-intel-tools.ps1 -RepoPath . -RequireRepowise:$false
+        # (same as ci.yml). The probe requires repowise by default, hence the
+        # explicit opt-out. Native since T3 (#48).
+        run: code-intel doctor bootstrap --repo-path . --no-require-repowise
 
       - name: GitHub research contract tests
         shell: pwsh

.sentrux/baseline.json

@@ -5,18 +5,18 @@
   },
   "metrics": {
     "complex_fn_count": 12,
-    "coupling_score": 45.26,
-    "cross_module_edges": 1041,
+    "coupling_score": 45.79,
+    "cross_module_edges": 1076,
     "cycle_count": 0,
-    "files": 230,
-    "functions": 3117,
+    "files": 235,
+    "functions": 3181,
     "god_file_count": 32,
     "max_complexity": 162,
-    "quality_signal": 3607,
-    "total_import_edges": 1041
+    "quality_signal": 3603,
+    "total_import_edges": 1076
   },
-  "savedAt": 1785242084,
+  "savedAt": 1785396251,
   "schema": "code-intel-sentrux-baseline.v2",
   "scope": ".",
-  "sourceCommit": "1512c9d5777cf683daa6c95c4c8415164af93cf3"
+  "sourceCommit": "4be192665d5339ae47a2c80306ac2cc097cb56d1"
 }

.sentrux/rules.toml

@@ -7,15 +7,30 @@
 # no god files, coupling grade B) is tracked modernization debt
 # (issue #14, "Remaining tracked debt").
 #
-# Measured ratchet evidence (sentrux-native 2.0.0, 2026-07-27, security-004
+# Measured ratchet evidence (sentrux-native 2.0.0, 2026-07-30, T3 doctor
 # tree, recorded together with .sentrux/baseline.json in the same commit):
-#   max_complexity 162 (archive/scripts/tests/test-code-intel-pipeline.ps1),
-#   god_file_count 29, coupling_score 45.33, cycle_count 0.
-# The 45.13 -> 45.33 coupling and 3968 -> 3967 quality movement is the cost
-# of the shared tool_path resolver module (security-004, #33): one new module
-# imported by six call sites replaces bare Command::new("rg"/"git") names, so
-# the added import edges are the fix working as intended, not drift. God files
-# and cycles did not move.
+#   coupling_score 45.26 -> 45.79, quality_signal 3607 -> 3603,
+#   god_file_count 32 (unchanged), cycle_count 0 (unchanged).
+# Cause: T3 (#48) replaced 409 lines of archive/check-code-intel-tools.ps1
+# with a native Rust probe split across four cohesive modules
+# (doctor_bootstrap/{mod,config,paths,probe}.rs) plus one integration test
+# file. Import edges went 1041 -> 1076 over 230 -> 235 files.
+#
+# Why the movement is accepted rather than engineered away: coupling_score is
+# import_lines / file_count * 10, and this tree's 45.26 average is held down by
+# large PowerShell files that carry almost no import lines. Any idiomatic Rust
+# module — roughly four to six `use`/`mod` lines — sits above that average, so
+# adding well-factored Rust always pushes the ratio up. Reaching 45.26 here
+# would have required dropping ~13 `use` lines in favour of inline fully
+# qualified paths: worse code, chosen only to move a line-counting metric.
+# What was NOT accepted: the first cut of this work landed the probe as one
+# 1124-line file, which tripped god_file_count 32 -> 33 (quality -120). That
+# is a real regression and was fixed by the module split above, not baselined.
+# Prior entry (2026-07-27, security-004): 45.13 -> 45.33 coupling and
+# 3968 -> 3967 quality was the cost of the shared tool_path resolver module
+# (#33): one new module imported by six call sites replaces bare
+# Command::new("rg"/"git") names, so the added import edges are the fix
+# working as intended, not drift. God files and cycles did not move.
 # Prior entry (2026-07-26): 45.07 -> 45.13 / 3969 -> 3968 was the ast-grep
 # internalization conformance suite cost.
 # The previous B / 70 / no-god thresholds were never green under any engine:

CHANGELOG.md

@@ -7,6 +7,34 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- **The doctor bootstrap probe is native Rust** (issue #48, T3 of the PS1
+  retirement campaign). `code-intel doctor bootstrap` now computes the
+  tool/runtime health inventory that `archive/check-code-intel-tools.ps1` used
+  to implement in 409 lines of PowerShell; the script is a ~35-line thin
+  forwarder retained for the installer and rollback paths. The observation it
+  emits keeps the `code-intel-doctor-bootstrap-observation.v1` schema, the
+  `observation_only` authority, and the same `ok`/`missing` pair, so installer
+  and CI consumers are unchanged.
+
+  ```bash
+  code-intel doctor bootstrap --repo-path . --no-require-repowise --json
+  ```
+
+  CI and release workflows call the binary directly. Running the forwarder on a
+  machine without the binary is now reported as a `code-intel binary` entry in
+  `missing` (exit 1) rather than as a crash.
+
+### Fixed
+
+- **The doctor capability no longer answers from a stub when `pwsh` is
+  absent.** The adapter previously shelled out to the PowerShell probe and, on
+  failure to launch it, fell back to an in-process approximation that reported
+  `graphProvider` presence as hardcoded `true` — masking exactly the drift the
+  doctor exists to surface. With one native probe there is no fallback path and
+  no `pwsh` dependency on the kernel path.
+
 ## [0.7.0-beta.2] — 2026-07-30
 
 ### Changed — action required