feat: topic-aware classification (#47), type-specific thresholds (#48)

#47 — Body content classification:
- Unmatched sections now get a secondary scan of body text (first 2000 chars)
- Register maps, address offsets, block diagrams, milestones, roadmaps
  route to verification.md (technical reference)
- Windows path setup, per-machine config, NTFS workarounds route to
  drift-metrics.md (environment)
- Reduces misclassification where heading is generic but content is specific

#48 — Type-specific governance thresholds:
- fpga-rtl: rules=1000, workflow=500, verification=600
- yocto-bsp: rules=1000, workflow=500, verification=500
- embedded-hardware: rules=1000, verification=500
- pcb-hardware: rules=900, verification=500
- Reads project type from scaffold.yml, falls back to defaults
- Software projects keep lower defaults (rules=800, etc.)

Closes #47, closes #48

Co-Authored-By: Oz <oz-agent@warp.dev>

Author: tbitcs

src/specsmith/auditor.py

@@ -283,20 +283,62 @@ def check_ledger_health(root: Path) -> list[AuditResult]:
 # ---------------------------------------------------------------------------
 
 
+# Default thresholds (used when no project type is detected)
+_DEFAULT_THRESHOLDS: dict[str, int] = {
+    "AGENTS.md": 200,
+    "docs/governance/rules.md": 800,
+    "docs/governance/workflow.md": 400,
+    "docs/governance/roles.md": 300,
+    "docs/governance/context-budget.md": 300,
+    "docs/governance/verification.md": 400,
+    "docs/governance/drift-metrics.md": 300,
+}
+
+# Type-specific overrides — hardware/embedded projects have denser rules.
+_TYPE_THRESHOLD_OVERRIDES: dict[str, dict[str, int]] = {
+    "fpga-rtl": {
+        "docs/governance/rules.md": 1000,
+        "docs/governance/workflow.md": 500,
+        "docs/governance/verification.md": 600,
+    },
+    "yocto-bsp": {
+        "docs/governance/rules.md": 1000,
+        "docs/governance/workflow.md": 500,
+        "docs/governance/verification.md": 500,
+    },
+    "embedded-hardware": {
+        "docs/governance/rules.md": 1000,
+        "docs/governance/verification.md": 500,
+    },
+    "pcb-hardware": {
+        "docs/governance/rules.md": 900,
+        "docs/governance/verification.md": 500,
+    },
+}
+
+
+def _get_thresholds(root: Path) -> dict[str, int]:
+    """Get governance size thresholds, scaled by project type if available."""
+    thresholds = dict(_DEFAULT_THRESHOLDS)
+    scaffold_path = root / "scaffold.yml"
+    if scaffold_path.exists():
+        try:
+            import yaml
+
+            with open(scaffold_path) as f:
+                raw = yaml.safe_load(f) or {}
+            ptype = raw.get("type", "")
+            overrides = _TYPE_THRESHOLD_OVERRIDES.get(ptype, {})
+            thresholds.update(overrides)
+        except Exception:  # noqa: BLE001
+            pass  # Use defaults on any error
+    return thresholds
+
+
 def check_context_size(root: Path) -> list[AuditResult]:
-    """Check governance file sizes against thresholds."""
+    """Check governance file sizes against type-aware thresholds."""
     results: list[AuditResult] = []
-    # AGENTS.md is strict (loaded every session). Modular governance files
-    # are lazily loaded per task type so can be larger.
-    thresholds = {
-        "AGENTS.md": 200,
-        "docs/governance/rules.md": 800,
-        "docs/governance/workflow.md": 400,
-        "docs/governance/roles.md": 300,
-        "docs/governance/context-budget.md": 300,
-        "docs/governance/verification.md": 400,
-        "docs/governance/drift-metrics.md": 300,
-    }
+    thresholds = _get_thresholds(root)
 
     for rel_path, max_lines in thresholds.items():
         path = root / rel_path

src/specsmith/importer.py

@@ -786,6 +786,20 @@ def _extract_governance_sections(root: Path) -> dict[str, str]:
     }
     unmatched: list[tuple[str, str]] = []
 
+    # Body-level content keywords for secondary classification.
+    # Used when heading doesn't match — scan body text for strong signals.
+    _BODY_ARCHITECTURE_KW = [
+        "register map", "address offset", "0x0", "register name",
+        "block diagram", "data flow", "interface spec",
+        "directory layout", "src/", "repository structure",
+        "milestone", "roadmap", "completion", "phase 2 target",
+    ]
+    _BODY_DRIFT_KW = [
+        "subst v:", "path-length", "one-time setup", "per-machine",
+        "environment variable", "install once", "bootstrap",
+        "windows path", "ntfs",
+    ]
+
     for heading, body in sections.items():
         key_lower = heading.lower()
         matched = False
@@ -795,7 +809,14 @@ def _extract_governance_sections(root: Path) -> dict[str, str]:
                 matched = True
                 break  # First match wins
         if not matched:
-            unmatched.append((heading, body))
+            # Secondary pass: scan body content for strong topic signals
+            body_lower = body[:2000].lower()  # Cap scan for performance
+            if any(kw in body_lower for kw in _BODY_ARCHITECTURE_KW):
+                buckets["verification"].append((heading, body))  # technical reference
+            elif any(kw in body_lower for kw in _BODY_DRIFT_KW):
+                buckets["drift-metrics"].append((heading, body))
+            else:
+                unmatched.append((heading, body))
 
     # Unmatched sections go to rules.md as project-specific rules
     if unmatched: