perf(cs): eliminate the C# corpus-proportional scans — 1211s -> 450s, +37% edges

Four measured changes on the dotnet/runtime corpus (58,656 files; every
step's numbers below are from full-corpus runs on the same host, baselines
captured this session; v0.9.0 = 444.3 s):

1. Two-phase cs registry build (types -> finalize -> funcs), the same
   pre-finalize linear-lookup disease fixed for Java: prepare 280 -> 140 s.

2. C# import-context carve-out: cs_import_types lists namespace_declaration
   (for namespace-name mapping) and using_statement (C#'s RAII block — a
   grammar-name collision), so EVERY namespaced C# file's whole body ran
   with inside_import=true. That both suppressed ordinary usage extraction
   under namespaces and sent every identifier through the ancestor-walking
   import-binding check (tree-sitter's ts_node_parent re-descends from the
   root, so wide files went quadratic: one 147 KB JIT torture file cost
   490 s; 6.8 s after). Only using_directive / namespace_use_declaration
   open an import scope now. Restores the suppressed usages:
   edges 4,291,387 -> 5,869,093 (+37%), nodes identical.

3. The usages walker maintains call/import ancestry as enter/exit counters
   on its explicit stack instead of per-node ancestor re-walks
   (extract_usages.c had grown from 6 to 100 ts_node_parent calls since
   v0.9.0; the two per-node gates are now O(1) with semantics preserved —
   strict ancestors only, emit before self-count).

4. Registry short-name indexes replace the two remaining full scans:
   cs_lookup_extension walked all 963k funcs per unresolved invocation
   (now the existing free-func short-name iterator, first-match order
   preserved via min-index selection), and cs_resolve_type_name's step-9
   fallback scanned every type per unresolved name IN BOTH the builder and
   per-file resolution (new type_short index in finalize, same
   reverse-insertion ascending-order pattern, best-score ties keep the
   first-in-registration-order winner). Builder 140,095 -> 500 ms; resolve
   cross-LSP CPU 5.5M -> 317k ms (us_per_file_per_kdef 191 -> 11).

End state: 449.9 s wall (1.01x of v0.9.0) with +48.6% edges vs v0.9.0 —
per-edge cost 32% BETTER than v0.9.0. Extract's remaining 359 s is the
24 MB hugeexpr1.cs parse floor both versions pay.

Guarded by the complexity suite; cs_lsp/extraction/edge/lang-contract
suites green including the 53-language calls-breadth contract.

Refs #1669.

Signed-off-by: Martin Vogel <martin.vogel.tech@gmail.com>

internal/cbm/extract_unified.c

@@ -1646,6 +1646,18 @@ static bool is_actual_import_boundary(CBMExtractCtx *ctx, TSNode node, const CBM
     char *name = ts_node_is_null(head) ? NULL : cbm_node_text(ctx->arena, head, ctx->source);
 
     switch (ctx->language) {
+    case CBM_LANG_CSHARP:
+        /* cs_import_types also lists namespace_declaration (so the import pass
+         * can map namespace names) and using_statement (C#'s RAII block, a
+         * grammar-name collision with using_directive). Neither is an import
+         * CONTEXT: treating them as one put every namespaced C# file's whole
+         * body behind inside_import, which both suppressed ordinary usage
+         * extraction there and sent every identifier through the ancestor-
+         * walking import-binding check — 92% of extract time on wide files
+         * (dotnet/runtime JIT torture tests, 490 s for one 147 KB file). Only
+         * the using DIRECTIVE opens an import scope. */
+        return strcmp(kind, "using_directive") == 0 ||
+               strcmp(kind, "namespace_use_declaration") == 0;
     case CBM_LANG_ELIXIR:
         return strcmp(kind, "call") != 0 ||
                (name && (strcmp(name, "import") == 0 || strcmp(name, "alias") == 0 ||

internal/cbm/extract_usages.c

@@ -49,36 +49,6 @@ static void walk_usages(CBMExtractCtx *ctx, TSNode root, const CBMLangSpec *spec
 static bool is_direct_argument_value(TSNode node);
 static TSNode python_direct_callable_attribute_site(TSNode node);
 
-// Check if a node is inside a call expression (to avoid double-counting as usage)
-static bool is_inside_call(TSNode node, const CBMLangSpec *spec) {
-    TSNode cur = ts_node_parent(node);
-    while (!ts_node_is_null(cur)) {
-        if (cbm_kind_in_set(cur, spec->call_node_types)) {
-            return true;
-        }
-        cur = ts_node_parent(cur);
-    }
-    return false;
-}
-
-// Check if a node is inside an import statement
-static bool is_inside_import(TSNode node, const CBMLangSpec *spec) {
-    bool has_imports = spec->import_node_types && spec->import_node_types[0];
-    bool has_from_imports = spec->import_from_types && spec->import_from_types[0];
-    if (!has_imports && !has_from_imports) {
-        return false;
-    }
-    TSNode cur = ts_node_parent(node);
-    while (!ts_node_is_null(cur)) {
-        if ((has_imports && cbm_kind_in_set(cur, spec->import_node_types)) ||
-            (has_from_imports && cbm_kind_in_set(cur, spec->import_from_types))) {
-            return true;
-        }
-        cur = ts_node_parent(cur);
-    }
-    return false;
-}
-
 // Is this an identifier-like node that represents a reference?
 static bool is_reference_node(TSNode node, CBMLanguage lang) {
     const char *kind = ts_node_type(node);
@@ -2426,7 +2396,8 @@ static bool emit_direct_perl_coderef_usage(CBMExtractCtx *ctx, TSNode node,
 }
 
 // Try to emit a usage for a reference node. Returns early if the node should be skipped.
-static void try_emit_usage(CBMExtractCtx *ctx, TSNode node, const CBMLangSpec *spec) {
+static void try_emit_usage(CBMExtractCtx *ctx, TSNode node, const CBMLangSpec *spec,
+                           bool inside_call, bool inside_import) {
     if (emit_direct_perl_coderef_usage(ctx, node, cbm_enclosing_func_qn_cached(ctx, node), 0)) {
         return;
     }
@@ -2439,7 +2410,7 @@ static void try_emit_usage(CBMExtractCtx *ctx, TSNode node, const CBMLangSpec *s
     if (is_call_argument_label(node)) {
         return;
     }
-    if (is_inside_call(node, spec) || is_inside_import(node, spec)) {
+    if (inside_call || inside_import) {
         return;
     }
     if (is_binding_occurrence(ctx, node, spec, NULL) ||
@@ -2456,18 +2427,79 @@ static void try_emit_usage(CBMExtractCtx *ctx, TSNode node, const CBMLangSpec *s
     }
 }
 
-// Iterative usage walker — explicit stack
+// Iterative usage walker — explicit stack.
+//
+// The call/import ancestry that gates usage emission is maintained as ENTER/
+// EXIT counters on the walk instead of per-node ancestor re-walks: the old
+// is_inside_call/is_inside_import helpers climbed every ancestor via
+// ts_node_parent, and tree-sitter's ts_node_parent RE-DESCENDS from the root
+// scanning siblings — O(depth x sibling-position) per node, which went
+// quadratic on wide nodes (a 1,536-argument call in dotnet/runtime's JIT
+// torture tests put 92% of extract time into these walks; 490 s for one
+// 147 KB file). Counter semantics match the helpers exactly: strict ancestors
+// only — a node is emitted BEFORE its own kind increments the counters, so a
+// call node itself does not count as "inside a call".
 static void walk_usages(CBMExtractCtx *ctx, TSNode root, const CBMLangSpec *spec) {
-    TSNodeStack stack;
-    ts_nstack_init(&stack, ctx->arena, 4096);
-    ts_nstack_push(&stack, ctx->arena, root);
+    typedef struct {
+        TSNode node;
+        uint32_t next_child;
+        bool counts_call;
+        bool counts_import;
+    } UsageFrame;
+    int cap = 256;
+    UsageFrame *frames = (UsageFrame *)cbm_arena_alloc(ctx->arena, (size_t)cap * sizeof(*frames));
+    if (!frames) {
+        return;
+    }
+    bool has_imports = spec->import_node_types && spec->import_node_types[0];
+    bool has_from_imports = spec->import_from_types && spec->import_from_types[0];
+    int call_depth = 0;
+    int import_depth = 0;
+    int top = 0;
+    frames[top++] = (UsageFrame){root, 0, false, false};
+    bool entering = true;
 
-    while (stack.count > 0) {
-        TSNode node = ts_nstack_pop(&stack);
-        try_emit_usage(ctx, node, spec);
-        uint32_t count = ts_node_child_count(node);
-        for (int i = (int)count - LAST_IDX; i >= 0; i--) {
-            ts_nstack_push(&stack, ctx->arena, ts_node_child(node, (uint32_t)i));
+    while (top > 0) {
+        UsageFrame *f = &frames[top - 1];
+        if (entering) {
+            try_emit_usage(ctx, f->node, spec, call_depth > 0, import_depth > 0);
+            f->counts_call = cbm_kind_in_set(f->node, spec->call_node_types);
+            f->counts_import =
+                (has_imports && cbm_kind_in_set(f->node, spec->import_node_types)) ||
+                (has_from_imports && cbm_kind_in_set(f->node, spec->import_from_types));
+            if (f->counts_call) {
+                call_depth++;
+            }
+            if (f->counts_import) {
+                import_depth++;
+            }
+        }
+        uint32_t count = ts_node_child_count(f->node);
+        if (f->next_child < count) {
+            TSNode child = ts_node_child(f->node, f->next_child);
+            f->next_child++;
+            if (top == cap) {
+                int new_cap = cap * 2;
+                UsageFrame *grown =
+                    (UsageFrame *)cbm_arena_alloc(ctx->arena, (size_t)new_cap * sizeof(*grown));
+                if (!grown) {
+                    return;
+                }
+                memcpy(grown, frames, (size_t)cap * sizeof(*frames));
+                frames = grown;
+                cap = new_cap;
+            }
+            frames[top++] = (UsageFrame){child, 0, false, false};
+            entering = true;
+        } else {
+            if (f->counts_call) {
+                call_depth--;
+            }
+            if (f->counts_import) {
+                import_depth--;
+            }
+            top--;
+            entering = false;
         }
     }
 }

internal/cbm/lsp/type_registry.c

@@ -169,6 +169,46 @@ static void build_type_embed_index(CBMTypeRegistry *reg, CBMArena *idx_arena) {
 
 /* Index: short_name -> chain of FREE-function (receiver_type==NULL) indices.
  * Descending-iterate + prepend for ascending chain order (as above). */
+static void build_type_short_index(CBMTypeRegistry *reg, CBMArena *idx_arena) {
+    int tcount = 0;
+    for (int i = 0; i < reg->type_count; i++) {
+        if (reg->types[i].short_name)
+            tcount++;
+    }
+    if (tcount == 0)
+        return;
+    int bucket_count = next_pow2(tcount * 2);
+    if (bucket_count < 16)
+        bucket_count = 16;
+    int *buckets = (int *)cbm_arena_alloc(idx_arena, (size_t)bucket_count * sizeof(int));
+    CBMRegistryHashEntry *entries = (CBMRegistryHashEntry *)cbm_arena_alloc(
+        idx_arena, (size_t)tcount * sizeof(CBMRegistryHashEntry));
+    if (!buckets || !entries)
+        return;
+    for (int i = 0; i < bucket_count; i++)
+        buckets[i] = -1;
+    int idx = 0;
+    /* Reverse insertion so each chain yields ASCENDING types[] order — callers
+     * that used first-match-in-registration-order keep their tie-breaks. */
+    for (int i = reg->type_count - 1; i >= 0; i--) {
+        const CBMRegisteredType *t = &reg->types[i];
+        if (!t->short_name)
+            continue;
+        uint64_t h = fnv1a(t->short_name);
+        int slot = (int)(h & (uint64_t)(bucket_count - 1));
+        entries[idx].hash = h;
+        entries[idx].payload_index = i;
+        entries[idx].next_index = buckets[slot];
+        entries[idx].slot = slot;
+        buckets[slot] = idx;
+        idx++;
+    }
+    reg->type_short_buckets = buckets;
+    reg->type_short_entries = entries;
+    reg->type_short_bucket_count = bucket_count;
+    reg->type_short_entry_count = idx;
+}
+
 static void build_ffunc_short_index(CBMTypeRegistry *reg, CBMArena *idx_arena) {
     int fcount = 0;
     for (int i = 0; i < reg->func_count; i++) {
@@ -286,6 +326,42 @@ int cbm_free_func_iter_next(CBMFreeFuncIter *it) {
     return -1;
 }
 
+void cbm_registry_types_by_short_name(const CBMTypeRegistry *reg, const char *short_name,
+                                      CBMTypeShortIter *out) {
+    out->reg = reg;
+    out->hash = fnv1a(short_name);
+    if (reg->type_qn_buckets && reg->type_qn_bucket_count > 0) {
+        if (reg->type_short_buckets && reg->type_short_bucket_count > 0) {
+            int slot = (int)(out->hash & (uint64_t)(reg->type_short_bucket_count - 1));
+            out->chain_idx = reg->type_short_buckets[slot];
+        } else {
+            out->chain_idx = -1;
+        }
+        out->tail_i = reg->type_qn_entry_count;
+        out->tail_end = reg->type_count;
+    } else {
+        out->chain_idx = -1;
+        out->tail_i = 0;
+        out->tail_end = reg->type_count;
+    }
+}
+
+int cbm_type_short_iter_next(CBMTypeShortIter *it) {
+    const CBMTypeRegistry *reg = it->reg;
+    while (it->chain_idx >= 0) {
+        const CBMRegistryHashEntry *e = &reg->type_short_entries[it->chain_idx];
+        int p = e->payload_index;
+        uint64_t h = e->hash;
+        it->chain_idx = e->next_index;
+        if (h != it->hash)
+            continue;
+        return p;
+    }
+    if (it->tail_i < it->tail_end)
+        return it->tail_i++;
+    return -1;
+}
+
 void cbm_registry_methods(const CBMTypeRegistry *reg, const char *receiver_qn,
                           const char *method_name, CBMMethodIter *out) {
     memset(out, 0, sizeof(*out));
@@ -346,6 +422,7 @@ void cbm_registry_finalize_into(CBMTypeRegistry *reg, CBMArena *idx_arena) {
     build_method_index(reg, idx_arena);
     build_type_embed_index(reg, idx_arena);
     build_ffunc_short_index(reg, idx_arena);
+    build_type_short_index(reg, idx_arena);
 }
 
 void cbm_registry_finalize(CBMTypeRegistry *reg) {

internal/cbm/lsp/type_registry.h

@@ -123,6 +123,10 @@ typedef struct CBMTypeRegistry {
     int type_embed_entry_count;
     // Free-function short-name index: fnv1a(short_name) -> chain of FREE-function
     // (receiver_type==NULL) indices. payload_index = func index.
+    int *type_short_buckets;
+    CBMRegistryHashEntry *type_short_entries;
+    int type_short_bucket_count;
+    int type_short_entry_count;
     int *ffunc_short_buckets;
     CBMRegistryHashEntry *ffunc_short_entries;
     int ffunc_short_bucket_count;
@@ -248,6 +252,22 @@ typedef struct {
     int tail_i;
     int tail_end;
 } CBMFreeFuncIter;
+/* Iterate TYPE indices sharing a short name — the type-side twin of the free-
+ * function iterator. Built by finalize into type_short_buckets; degrades to a
+ * full types[] scan on an unfinalized registry (same correctness, old cost).
+ * Added for cs_resolve_type_name's step-9 fallback, which scanned type_count
+ * per unresolved name — quadratic against the shared Tier-2 registry. */
+typedef struct {
+    const CBMTypeRegistry *reg;
+    uint64_t hash;
+    int chain_idx;
+    int tail_i;
+    int tail_end;
+} CBMTypeShortIter;
+void cbm_registry_types_by_short_name(const CBMTypeRegistry *reg, const char *short_name,
+                                      CBMTypeShortIter *out);
+int cbm_type_short_iter_next(CBMTypeShortIter *it);
+
 void cbm_registry_free_funcs_by_short_name(const CBMTypeRegistry *reg, const char *short_name,
                                            CBMFreeFuncIter *out);
 int cbm_free_func_iter_next(CBMFreeFuncIter *it);

src/pipeline/pipeline.c

@@ -1259,14 +1259,39 @@ static int run_parallel_pipeline(cbm_pipeline_t *p, cbm_pipeline_ctx_t *ctx,
     cbm_arena_init(&cross_lsp_arena);
     CBMCrossLspRegistries cross_registries = {0};
     if (all_defs) {
+        /* Per-builder split of lsp_cross_prepare — attributes a slow prepare to
+         * ONE language instead of re-diagnosing the whole pass (the cs builder
+         * hid ~140 s behind the pass total, #1669 follow-up). */
+        struct timespec t_b;
+        long b_ms[6];
+        cbm_clock_gettime(CLOCK_MONOTONIC, &t_b);
         cross_registries.go = cbm_go_build_cross_registry(&cross_lsp_arena, all_defs, def_count);
+        b_ms[0] = (long)elapsed_ms(t_b);
+        cbm_clock_gettime(CLOCK_MONOTONIC, &t_b);
         cross_registries.python =
             cbm_py_build_cross_registry(&cross_lsp_arena, all_defs, def_count);
+        b_ms[1] = (long)elapsed_ms(t_b);
+        cbm_clock_gettime(CLOCK_MONOTONIC, &t_b);
         cross_registries.c = cbm_c_build_cross_registry(&cross_lsp_arena, all_defs, def_count);
+        b_ms[2] = (long)elapsed_ms(t_b);
+        cbm_clock_gettime(CLOCK_MONOTONIC, &t_b);
         cross_registries.cs = cbm_cs_build_cross_registry(&cross_lsp_arena, all_defs, def_count);
+        b_ms[3] = (long)elapsed_ms(t_b);
+        cbm_clock_gettime(CLOCK_MONOTONIC, &t_b);
         cross_registries.ts = cbm_ts_build_cross_registry(&cross_lsp_arena, all_defs, def_count);
+        b_ms[4] = (long)elapsed_ms(t_b);
+        cbm_clock_gettime(CLOCK_MONOTONIC, &t_b);
         cross_registries.java =
             cbm_java_build_cross_registry(&cross_lsp_arena, all_defs, def_count);
+        b_ms[5] = (long)elapsed_ms(t_b);
+        char b_buf[6][CBM_SZ_16];
+        const char *b_name[6] = {"go", "python", "c", "cs", "ts", "java"};
+        for (int bi = 0; bi < 6; bi++) {
+            snprintf(b_buf[bi], sizeof(b_buf[bi]), "%ld", b_ms[bi]);
+        }
+        cbm_log_info("lsp_cross_prepare.builders", b_name[0], b_buf[0], b_name[1], b_buf[1],
+                     b_name[2], b_buf[2], b_name[3], b_buf[3], b_name[4], b_buf[4], b_name[5],
+                     b_buf[5]);
         /* Rust: NOT built here. The shared all_defs registry is built LAZILY on the
          * first NULL-filter rust file (the amplifier files) inside cbm_parallel_resolve
          * — repos whose rust files all filter to subsets never pay the build/RSS. */