@ruso-0/nreki 10.18.0 → 10.18.1

package/CHANGELOG.md

@@ -1,176 +1,249 @@
-# Changelog
-
-All notable changes to NREKI will be documented in this file.
-
-## [10.18.0] - 2026-05-02
-
-# v10.18.0 — Schema compatibility + CSS parser + UX deadlock fix
-
-## Fixed
-
-### Critical: Schema break in strict MCP clients (Bloque A)
-SDK MCP `@modelcontextprotocol/sdk` was silently resolving from `^1.12.1` 
-to `1.27.1` via `npx -y` distribution. Versions 1.24.0+ inject an 
-`execution.taskSupport: 'forbidden'` field in tool registrations that 
-strict MCP clients (Gemini CLI, Codex CLI) reject as non-spec, causing 
-NREKI tools to be silently dropped. Now pinned to `1.23.1` (last clean 
-version, verified empirically).
-
-All other runtime dependencies pinned to current lockfile resolutions to 
-prevent recurrence:
-- `@typescript-eslint/parser`, `chokidar`, `eslint`, `eslint-plugin-jsx-a11y`, 
-  `eslint-plugin-react`, `eslint-plugin-react-hooks`, `picomatch`, 
-  `web-tree-sitter`, `zod`, `@xenova/transformers`
-- TypeScript peer dependency capped at `>=5.5.0 <6.0.0`
-
-### Critical: P0 deadlock in Chronos + Cognitive Enforcer (Bug D reclassified)
-Files marked HIGH friction by Chronos triggered an infinite loop:
-1. Layer 2 (edit gate) blocks edit and instructs LLM to `read compress:false`
-2. LLM obeys
-3. Layer 1 (Cognitive Enforcer) blocks the raw read on >100 line files
-4. Layer 2 keeps demanding raw read, Layer 1 keeps rejecting
-5. Context budget exhausted
-
-Fix: Layer 2 now emits the JSON-RPC tool call including 
-`_nreki_bypass:"chronos_recovery"` token that Layer 1 honors. Both 
-single-edit and batch-edit gates updated.
-
-### CSS parser captures at-rules (Bloque C2)
-Tree-sitter-css query previously only captured `rule_set` selectors, 
-making `@keyframes`, `@media`, `@import`, and `@font-face` blocks 
-invisible to the chunks index. `fast_grep` for at-rule strings returned 
-"No matches" even when present in source files.
-
-Now captures 5 node types with `@symbol_name` binding:
-- `rule_set` (regular CSS rules)
-- `keyframes_statement` (`@keyframes`)
-- `media_statement` (`@media`)
-- `import_statement` (`@import`)
-- `at_rule` (catch-all for `@font-face`, `@supports`, etc.)
-
-### CSS focus normalization (Hallazgo Colateral)
-Parser strips punctuation from selectors at extraction (`.foo` → `foo`), 
-but `compressor-foveal.ts` kept the user's focus raw. TFC-Pro silently 
-failed to find `.home-sidebar` against the indexed `home-sidebar`. Now 
-applies the same normalization to focus when the file is CSS or HTML.
-
-### TFC-Pro returns discriminated union (Bloque C1)
-Previously returned `TfcResult | null` and emitted misleading 
-"NOT FOUND or TOO LARGE (Density Shield)" which conflated two distinct 
-cases. New `TfcResultPayload` type:
-- `{ kind: "success", data: TfcResult }`
-- `{ kind: "not_found" }` → "symbol X not found, run outline first"
-- `{ kind: "shield_tripped", ratio, ... }` → "focus spans >85% of file, 
-  use compress:false or omit focus"
-
-### Multi-focus validation with domain shield (Bloque C3.A)
-Agents passing `focus="func1 func2 func3"` (space-separated) previously 
-became a single token that never matched, returning generic "not_found". 
-Now emits an educational error in non-web file types. CSS/HTML preserve 
-multi-word selectors (e.g., `body .foo`).
-
-## Migration notes
-
-If upgrading from earlier versions and you observe that:
-- `fast_grep` returns no matches for CSS at-rules in projects you 
-  previously indexed
-- Indexed paths still point to old project locations (e.g., after 
-  moving the project to a different directory)
-
-**Workaround:** delete the `.nreki/` directory and `.nreki.db` file from 
-your project root to force reindex with the new parser.
-
-```bash
-# Linux/Mac
-rm -rf .nreki/ .nreki.db
-
-# Windows PowerShell  
-Remove-Item -Recurse -Force .nreki, .nreki.db
-```
-
-A permanent migration mechanism (paths relative to projectRoot, 
-schema-versioned cache invalidation) is shipping in v10.18.1.
-
-## Known issues for v10.18.1
-
-- Path stale: paths in `.nreki.db` and `.nreki/backups/` are stored 
-  absolute, breaking when projects are moved between directories.
-- 8 npm audit vulnerabilities in transitive dependencies (review case 
-  by case in v10.18.1).
-
-## Dependencies
-
-- `@modelcontextprotocol/sdk`: `^1.12.1` → `1.23.1` (pinned, downgrade)
-- All other runtime deps pinned to current lockfile resolutions
-- 9 commits in this release including 3 internal smoke tests as 
-  permanent regression tooling
-
-## [10.17.0] - 2026-04-25
-
-### Reliability pass: 6 critical bugs fixed
-
-Pre-publish audit revealed 6 bugs in v10.16.x tags (never published to
-npm). All fixed in this release.
-
-- **fix(parser)**: `.tsx` files now use dedicated tree-sitter-tsx
-  grammar. Previously JSX caused 0 chunks, making React/Next/Remix
-  components invisible to NREKI. `da8275e`
-- **fix(engine)**: File watcher auto-starts on boot. `fast_grep`
-  bootstraps index on first call. chokidar config uses
-  `ignoreInitial: true` and root-directory watching (required by
-  chokidar@4, which dropped glob support) to prevent race with
-  handler-side indexing. `77afbed`
-- **fix(parser)**: Top-level `const`/`let` declarations now indexed.
-  Previously only arrow-function/function-expression constants
-  captured. Config objects, lookup tables, hardcoded values now
-  visible to fast_grep. `25287b2`
-- **fix(semantic-edit)**: `batch_edit` path keys normalized via
-  `safePath()` for both stores and lookups, including hologram branch.
-  Relative paths no longer silently skip blast radius detection.
-  `907b7d2` + follow-up commit.
-- **fix(path-jail)**: Windows drive letter comparison case-insensitive.
-  Lowercase `d:/` no longer blocked as traversal. Affects VSCode URIs,
-  WSL symlinks, shell copy-paste. `9c346d9`
-- **fix(signature)**: `extractSignature` strips body from one-liner
-  functions, cutting at `=>` or `{` at zero paren/angle depth. No more
-  false positive blast radius on body-only edits. `76d778e`
-
-### Performance improvements
-
-Zero-allocation rewrite + FFI diet from abandoned v10.16.0 branch now
-published in this release:
-
-- **perf(navigate)**: fast_grep O(L) pointer walk via `indexOf`
-  (V8 memchr SIMD internally). Replaces per-hit allocations.
-- **perf(fast_grep)**: Lean SELECT (4 cols) reduces WASM-to-JS FFI
-  decoding overhead. Legacy 10-col method removed (breaking for
-  any hypothetical external consumer).
-
-Combined benchmark delta on NREKI src/ (10 iter, median):
-  - export (156 hits):   11.24ms -> 5.22ms  (-54%)
-  - findDefinition:       2.98ms -> 2.26ms  (-24%)
-  - chronosMemory:        2.16ms -> 1.73ms  (-20%)
-  - [OK]:                 1.81ms -> 1.52ms  (-16%)
-
-### Multi-agent manifest
-
-`nreki init` now generates CLAUDE.md + AGENTS.md + SKILL.md. All three
-share the IMMUNE SYSTEM operational rules: 80L Guillotine, Multi-patching
-ACID, Anti-Sweep Shield, TTRD Shield with CFI, Engrams with ASSERT prefix.
-Compatible with Claude Code, Codex, Gemini, and any MCP-capable agent.
-
-### Tests
-
-877 passing, 1 skipped (Linux-specific case-sensitivity test skipped on
-win32). +26 tests added across 6 bug fixes.
-
-### Note on unreleased versions
-
-v10.16.0 (`c08e61a`) and v10.16.1 (`ae9a526`) were git-tagged but never
-reached npm registry due to the 6 bugs above discovered in pre-publish
-audit. v10.17.0 is the next version reaching npm after v10.15.1.
-
-## [10.16.1] - 2026-04-24
+# Changelog
+
+All notable changes to NREKI will be documented in this file.
+
+## [10.18.1] - 2026-05-04
+
+# v10.18.1 — Cache schema versioning + web-symbol normalization + sprint hygiene
+
+Sprint focus: bug fixes, cache invalidation correctness, test isolation, and
+documentation precision. Closes the migration TODO from v10.18.0
+("schema-versioned cache invalidation shipping in v10.18.1") and the web-symbol
+asymmetry between parser storage and edit lookup.
+
+## Fixed
+
+### `batch_edit` / `edit` symbol normalization on web files (CSS / HTML / JSON)
+Asymmetric normalization: the parser stripped prefixes at chunk storage time
+(`.` for CSS classes, `#` for IDs, quotes for JSON keys), but `semanticEdit`
+and `batchSemanticEdit` matched against the raw caller-supplied symbol. LLM
+callers sending `.className` or `"key"` now correctly match parser-stored
+chunks. Also closes a key-leak in blast-radius detection where `oldRawCodes` /
+`newRawCodes` were keyed off the normalized symbol while downstream lookups
+used the raw input. (`59c3934`)
+
+### Test isolation leak in router and backward-compat tests
+Tests no longer pollute global state; previously-flaky tests now pass
+deterministically. (`77ae56c`)
+
+## Added
+
+### Parser schema version gate for cache invalidation (`69e5acd`, `bb10bb4`)
+Bumped `PARSER_SCHEMA_VERSION` to 2. Existing `.nreki.db` caches with v1 schema
+are auto-wiped on next run, eliminating the manual `rm -rf .nreki .nreki.db`
+step that v10.18.0 documented as a workaround. Includes characterization tests
+covering version-gate transitions.
+
+### `READ_ONLY_ACTIONS` whitelist in circuit breaker (`9717ce7`)
+20 read-only actions skip the `containsError()` heuristic to prevent
+false-positive doom-loop trips during normal navigation. `filter_output` is
+deliberately excluded — it remains the primary gateway for detecting
+test/build error loops. `response.isError === true` continues to trip the
+breaker even on read-only actions (hardware signal vs heuristic).
+
+### NREKI artifact exclusion from `DEFAULT_IGNORE` (`a50df2d`)
+`**/.nreki/**` and `**/.nreki.db*` patterns now ignored by default. Critical
+for upcoming corpus work — without this, NREKI re-indexed its own artifacts on
+every repo it scanned.
+
+## Changed
+
+### Removed `nreki-enforcer` hook integration (`fe2c0d0`)
+Cleanup of unused Claude hook. No functional impact on package consumers.
+
+## Documentation
+
+### Clarified v7.x spectral gap language in CHANGELOG (`2efa26a`)
+"for predictive analysis" → "for spectral gap analysis" in the v7.x λ₃ entry.
+AHI is a deterministic descriptive composite (percentile bucketing P20/P50/P80
+over peer-group distribution), not a supervised classifier — phrasing aligned
+with that design.
+
+## Tests
+
+- 909 passed | 1 skipped (893 base + 16 new from
+  `tests/web-symbol-normalization.test.ts`).
+- `tsc --noEmit` clean.
+
+## Compatibility
+
+- **Recommended runtime: Claude Sonnet 4.6+** for 1M context window support.
+  The 1M context beta header was retired for Sonnet 4.5/4 on 2026-04-30;
+  NREKI's compression remains effective on 200k contexts but benefits more
+  from 1M.
+- `@modelcontextprotocol/sdk` remains pinned at `1.23.1` (1.24.0+ injects
+  `execution: { taskSupport: 'forbidden' }` which breaks NREKI's tool
+  registration).
+
+## [10.18.0] - 2026-05-02
+
+# v10.18.0 — Schema compatibility + CSS parser + UX deadlock fix
+
+## Fixed
+
+### Critical: Schema break in strict MCP clients (Bloque A)
+SDK MCP `@modelcontextprotocol/sdk` was silently resolving from `^1.12.1` 
+to `1.27.1` via `npx -y` distribution. Versions 1.24.0+ inject an 
+`execution.taskSupport: 'forbidden'` field in tool registrations that 
+strict MCP clients (Gemini CLI, Codex CLI) reject as non-spec, causing 
+NREKI tools to be silently dropped. Now pinned to `1.23.1` (last clean 
+version, verified empirically).
+
+All other runtime dependencies pinned to current lockfile resolutions to 
+prevent recurrence:
+- `@typescript-eslint/parser`, `chokidar`, `eslint`, `eslint-plugin-jsx-a11y`, 
+  `eslint-plugin-react`, `eslint-plugin-react-hooks`, `picomatch`, 
+  `web-tree-sitter`, `zod`, `@xenova/transformers`
+- TypeScript peer dependency capped at `>=5.5.0 <6.0.0`
+
+### Critical: P0 deadlock in Chronos + Cognitive Enforcer (Bug D reclassified)
+Files marked HIGH friction by Chronos triggered an infinite loop:
+1. Layer 2 (edit gate) blocks edit and instructs LLM to `read compress:false`
+2. LLM obeys
+3. Layer 1 (Cognitive Enforcer) blocks the raw read on >100 line files
+4. Layer 2 keeps demanding raw read, Layer 1 keeps rejecting
+5. Context budget exhausted
+
+Fix: Layer 2 now emits the JSON-RPC tool call including 
+`_nreki_bypass:"chronos_recovery"` token that Layer 1 honors. Both 
+single-edit and batch-edit gates updated.
+
+### CSS parser captures at-rules (Bloque C2)
+Tree-sitter-css query previously only captured `rule_set` selectors, 
+making `@keyframes`, `@media`, `@import`, and `@font-face` blocks 
+invisible to the chunks index. `fast_grep` for at-rule strings returned 
+"No matches" even when present in source files.
+
+Now captures 5 node types with `@symbol_name` binding:
+- `rule_set` (regular CSS rules)
+- `keyframes_statement` (`@keyframes`)
+- `media_statement` (`@media`)
+- `import_statement` (`@import`)
+- `at_rule` (catch-all for `@font-face`, `@supports`, etc.)
+
+### CSS focus normalization (Hallazgo Colateral)
+Parser strips punctuation from selectors at extraction (`.foo` → `foo`), 
+but `compressor-foveal.ts` kept the user's focus raw. TFC-Pro silently 
+failed to find `.home-sidebar` against the indexed `home-sidebar`. Now 
+applies the same normalization to focus when the file is CSS or HTML.
+
+### TFC-Pro returns discriminated union (Bloque C1)
+Previously returned `TfcResult | null` and emitted misleading 
+"NOT FOUND or TOO LARGE (Density Shield)" which conflated two distinct 
+cases. New `TfcResultPayload` type:
+- `{ kind: "success", data: TfcResult }`
+- `{ kind: "not_found" }` → "symbol X not found, run outline first"
+- `{ kind: "shield_tripped", ratio, ... }` → "focus spans >85% of file, 
+  use compress:false or omit focus"
+
+### Multi-focus validation with domain shield (Bloque C3.A)
+Agents passing `focus="func1 func2 func3"` (space-separated) previously 
+became a single token that never matched, returning generic "not_found". 
+Now emits an educational error in non-web file types. CSS/HTML preserve 
+multi-word selectors (e.g., `body .foo`).
+
+## Migration notes
+
+If upgrading from earlier versions and you observe that:
+- `fast_grep` returns no matches for CSS at-rules in projects you 
+  previously indexed
+- Indexed paths still point to old project locations (e.g., after 
+  moving the project to a different directory)
+
+**Workaround:** delete the `.nreki/` directory and `.nreki.db` file from 
+your project root to force reindex with the new parser.
+
+```bash
+# Linux/Mac
+rm -rf .nreki/ .nreki.db
+
+# Windows PowerShell  
+Remove-Item -Recurse -Force .nreki, .nreki.db
+```
+
+A permanent migration mechanism (paths relative to projectRoot, 
+schema-versioned cache invalidation) is shipping in v10.18.1.
+
+## Known issues for v10.18.1
+
+- Path stale: paths in `.nreki.db` and `.nreki/backups/` are stored 
+  absolute, breaking when projects are moved between directories.
+- 8 npm audit vulnerabilities in transitive dependencies (review case 
+  by case in v10.18.1).
+
+## Dependencies
+
+- `@modelcontextprotocol/sdk`: `^1.12.1` → `1.23.1` (pinned, downgrade)
+- All other runtime deps pinned to current lockfile resolutions
+- 9 commits in this release including 3 internal smoke tests as 
+  permanent regression tooling
+
+## [10.17.0] - 2026-04-25
+
+### Reliability pass: 6 critical bugs fixed
+
+Pre-publish audit revealed 6 bugs in v10.16.x tags (never published to
+npm). All fixed in this release.
+
+- **fix(parser)**: `.tsx` files now use dedicated tree-sitter-tsx
+  grammar. Previously JSX caused 0 chunks, making React/Next/Remix
+  components invisible to NREKI. `da8275e`
+- **fix(engine)**: File watcher auto-starts on boot. `fast_grep`
+  bootstraps index on first call. chokidar config uses
+  `ignoreInitial: true` and root-directory watching (required by
+  chokidar@4, which dropped glob support) to prevent race with
+  handler-side indexing. `77afbed`
+- **fix(parser)**: Top-level `const`/`let` declarations now indexed.
+  Previously only arrow-function/function-expression constants
+  captured. Config objects, lookup tables, hardcoded values now
+  visible to fast_grep. `25287b2`
+- **fix(semantic-edit)**: `batch_edit` path keys normalized via
+  `safePath()` for both stores and lookups, including hologram branch.
+  Relative paths no longer silently skip blast radius detection.
+  `907b7d2` + follow-up commit.
+- **fix(path-jail)**: Windows drive letter comparison case-insensitive.
+  Lowercase `d:/` no longer blocked as traversal. Affects VSCode URIs,
+  WSL symlinks, shell copy-paste. `9c346d9`
+- **fix(signature)**: `extractSignature` strips body from one-liner
+  functions, cutting at `=>` or `{` at zero paren/angle depth. No more
+  false positive blast radius on body-only edits. `76d778e`
+
+### Performance improvements
+
+Zero-allocation rewrite + FFI diet from abandoned v10.16.0 branch now
+published in this release:
+
+- **perf(navigate)**: fast_grep O(L) pointer walk via `indexOf`
+  (V8 memchr SIMD internally). Replaces per-hit allocations.
+- **perf(fast_grep)**: Lean SELECT (4 cols) reduces WASM-to-JS FFI
+  decoding overhead. Legacy 10-col method removed (breaking for
+  any hypothetical external consumer).
+
+Combined benchmark delta on NREKI src/ (10 iter, median):
+  - export (156 hits):   11.24ms -> 5.22ms  (-54%)
+  - findDefinition:       2.98ms -> 2.26ms  (-24%)
+  - chronosMemory:        2.16ms -> 1.73ms  (-20%)
+  - [OK]:                 1.81ms -> 1.52ms  (-16%)
+
+### Multi-agent manifest
+
+`nreki init` now generates CLAUDE.md + AGENTS.md + SKILL.md. All three
+share the IMMUNE SYSTEM operational rules: 80L Guillotine, Multi-patching
+ACID, Anti-Sweep Shield, TTRD Shield with CFI, Engrams with ASSERT prefix.
+Compatible with Claude Code, Codex, Gemini, and any MCP-capable agent.
+
+### Tests
+
+877 passing, 1 skipped (Linux-specific case-sensitivity test skipped on
+win32). +26 tests added across 6 bug fixes.
+
+### Note on unreleased versions
+
+v10.16.0 (`c08e61a`) and v10.16.1 (`ae9a526`) were git-tagged but never
+reached npm registry due to the 6 bugs above discovered in pre-publish
+audit. v10.17.0 is the next version reaching npm after v10.15.1.
+
+## [10.16.1] - 2026-04-24
 
 ### Multi-agent manifest completion
 
@@ -1036,7 +1109,7 @@ Three localized security/correctness fixes shipped as separate bisectable commit
 
 ### Added
 - **Fiedler Vector extraction**: `analyzeTopology` now returns the full eigenvector `v2` (bridge fragility map)
-- **Third eigenvalue (λ₃)**: Enables spectral gap computation ∇(λ₃ - λ₂) for predictive analysis
+- **Third eigenvalue (λ₃)**: Enables spectral gap computation ∇(λ₃ - λ₂) for spectral gap analysis
 - **Third eigenvector (v3)**: Topological stress coordinates per node
 - **Gauge Fixing**: Deterministic phase canonicalization prevents sign ambiguity across commits (critical for ML pipelines)
 - **Gram-Schmidt deflation**: Reusable `powerIteration()` function extracts arbitrary eigenvectors

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ruso-0/nreki",
-  "version": "10.18.0",
+  "version": "10.18.1",
   "description": "MCP plugin that validates AI agent edits in RAM before they touch disk. Spectral clustering, architecture diffs, bridge detection, dead code oracle, and cross-file semantic checks for TypeScript, Go (gopls), and Python (pyright). Zero cloud dependencies.",
   "main": "dist/index.js",
   "type": "module",

package/src/compressor-foveal.ts

@@ -22,7 +22,7 @@
 
 import path from "path";
 import crypto from "crypto";
-import { type ParsedChunk, type ParseResult } from "./parser.js";
+import { type ParsedChunk, type ParseResult, normalizeWebSymbol } from "./parser.js";
 import { extractDependencies, cleanSignature } from "./utils/imports.js";
 import { Embedder } from "./embedder.js";
 import type { NrekiEngine } from "./engine.js";
@@ -109,15 +109,13 @@ export async function tfcCompress(
     if (parseResult.chunks.length === 0) return { kind: "not_found" };
 
     // 1. MULTI-FOCAL TARGETS + OVERLOAD FIX
+    // v10.18.1: shared normalizeWebSymbol covers CSS+HTML+JSON uniformly.
+    // Previous inline regex covered only CSS (despite isCss naming HTML too,
+    // it ran identical CSS rules). JSON foci were never normalized.
     const ext = path.extname(filePath).toLowerCase();
-    const isCss = ext === ".css" || ext === ".html";
 
     const foci = focusInput.split(",").map(s => {
-        let clean = s.trim().toLowerCase();
-        if (isCss) {
-            clean = clean.replace(/[.#,]/g, " ").replace(/\s+/g, " ").trim();
-        }
-        return clean;
+        return normalizeWebSymbol(s.trim(), ext).toLowerCase();
     }).filter(Boolean);
     const foveas = new Set<ParsedChunk>();
 

package/src/database.ts

@@ -126,6 +126,23 @@ export class NrekiDB {
         // Setup schema first (creates metadata table needed for dimension lookup)
         this.setupSchema();
 
+        // ─── Schema Version Gate ──────────────────────────────────────
+        // Force full reindex when parser/index format changes between
+        // versions. Users on older schema get a clean slate without
+        // manual .nreki.db deletion. Bumped in v10.18.1.
+        const PARSER_SCHEMA_VERSION = 2;
+        const storedSchema = parseInt(this.getMetadata("parser_schema_version") ?? "0", 10);
+        if (storedSchema < PARSER_SCHEMA_VERSION) {
+            if (storedSchema > 0) {
+                logger.warn(
+                    `[NREKI] Parser schema upgrade (v${storedSchema} -> v${PARSER_SCHEMA_VERSION}). ` +
+                    `Forcing full AST reindex.`
+                );
+            }
+            this.wipeAllIndexedData();
+            this.setMetadata("parser_schema_version", String(PARSER_SCHEMA_VERSION));
+        }
+
         // Load vector index using stored dimension (default 512)
         const storedDim = parseInt(this.getMetadata("embedding_dim") ?? "512", 10);
         if (fs.existsSync(this.vecPath)) {
@@ -277,9 +294,42 @@ export class NrekiDB {
         );
     }
 
+    /**
+     * Wipes all indexed data from RAM and disk.
+     * Prevents ghost vectors and ID drift by cleaning up sequence counters
+     * and orphaned .vec files. Used during cache invalidation.
+     */
+    private wipeAllIndexedData(): void {
+        // 1. Purge disk artifacts to prevent ghost data on next boot
+        if (fs.existsSync(this.vecPath)) {
+            try {
+                fs.unlinkSync(this.vecPath);
+            } catch (err) {
+                logger.warn(`[NREKI] Could not delete stale vector index ${this.vecPath}: ${err}`);
+            }
+        }
+
+        // 2. Wipe SQL tables
+        this.db.run("DELETE FROM chunks");
+        this.db.run("DELETE FROM files");
+
+        // 3. Reset AUTOINCREMENT sequences (ONLY for wiped tables)
+        try {
+            this.db.run("DELETE FROM sqlite_sequence WHERE name = 'chunks'");
+        } catch {
+            // Ignored: sqlite_sequence is created automatically by SQLite on first INSERT
+        }
+
+        // 4. Reset RAM state
+        this.vecIndex = new VectorIndex();
+        this.kwIndex = new KeywordIndex();
+        this.rawIdentsByFile.clear();
+        this.rawIdentsLoaded = false;
+    }
+
     /**
      * Check if the active embedding dimension matches what was stored.
-     * If they differ, clear all vectors and update the stored dimension.
+     * If they differ, wipe all indexed data and update the stored dimension.
      * Returns true if a re-index is needed.
      */
     checkEmbeddingDimension(activeDim: number): boolean {
@@ -287,14 +337,7 @@ export class NrekiDB {
 
         if (storedDim && parseInt(storedDim, 10) !== activeDim) {
             logger.warn(`Embedding dimension changed (${storedDim} -> ${activeDim}). Clearing index.`);
-            // Clear all vectors
-            this.vecIndex = new VectorIndex();
-            // Clear all chunks and files so they get re-indexed
-            this.db.run("DELETE FROM chunks");
-            this.db.run("DELETE FROM files");
-            this.kwIndex = new KeywordIndex();
-            this.rawIdentsByFile.clear();
-            this.rawIdentsLoaded = false;
+            this.wipeAllIndexedData();
             this.setMetadata("embedding_dim", String(activeDim));
             return true;
         }

package/src/engine.ts

@@ -49,6 +49,10 @@ const DEFAULT_IGNORE = [
     "**/__pycache__/**",
     "**/*.min.js",
     "**/*.d.ts",
+    // NREKI's own artifacts (prevents auto-indexing of self in dogfooding
+    // and contamination of corpus benchmarks against external repos)
+    "**/.nreki/**",
+    "**/.nreki.db*",
 ];
 
 // ─── Session Tracker ─────────────────────────────────────────────────
@@ -112,15 +116,15 @@ export class NrekiEngine {
     private embedder: Embedder;
     private parser: ASTParser;
     private compressor: Compressor;
-    private advancedCompressor: AdvancedCompressor;
+    private advancedCompressor: AdvancedCompressor;
 
     // Sub-pipelines (injected after initialize)
     private indexer!: IndexPipeline;
     private searcher!: SearchEngine;
 
-    private watcher: FSWatcher | null = null;
-    private watcherReady: Promise<void> | null = null;
-    private config: Required<EngineConfig>;
+    private watcher: FSWatcher | null = null;
+    private watcherReady: Promise<void> | null = null;
+    private config: Required<EngineConfig>;
     private indexingQueue = new Set<string>();
     private isIndexing = false;
     private initialized = false;
@@ -454,33 +458,33 @@ export class NrekiEngine {
 
     // ─── File Watching ────────────────────────────────────────────
 
-    startWatcher(): Promise<void> {
-        if (this.watcher) return this.watcherReady ?? Promise.resolve();
-
-        const shouldWatchFile = (fp: string) => this.config.extensions.includes(path.extname(fp).toLowerCase());
-
-        this.watcher = chokidar.watch(this.config.watchPaths, {
-            ignored: this.config.ignorePaths,
-            persistent: true,
-            ignoreInitial: true, // prevent race with handler bootstrap
-        });
-
-        this.watcherReady = new Promise((resolve, reject) => {
-            this.watcher!.once("ready", resolve);
-            this.watcher!.once("error", reject);
-        });
-
-        this.watcher
-            .on("add", (fp: string) => { if (shouldWatchFile(fp)) this.queueIndexing(fp); })
-            .on("change", (fp: string) => { if (shouldWatchFile(fp)) this.queueIndexing(fp); })
-            .on("unlink", (fp: string) => {
-                if (!shouldWatchFile(fp)) return;
-                this.db.clearChunks(fp);
-                this.scheduleSave();
-            });
-
-        return this.watcherReady;
-    }
+    startWatcher(): Promise<void> {
+        if (this.watcher) return this.watcherReady ?? Promise.resolve();
+
+        const shouldWatchFile = (fp: string) => this.config.extensions.includes(path.extname(fp).toLowerCase());
+
+        this.watcher = chokidar.watch(this.config.watchPaths, {
+            ignored: this.config.ignorePaths,
+            persistent: true,
+            ignoreInitial: true, // prevent race with handler bootstrap
+        });
+
+        this.watcherReady = new Promise((resolve, reject) => {
+            this.watcher!.once("ready", resolve);
+            this.watcher!.once("error", reject);
+        });
+
+        this.watcher
+            .on("add", (fp: string) => { if (shouldWatchFile(fp)) this.queueIndexing(fp); })
+            .on("change", (fp: string) => { if (shouldWatchFile(fp)) this.queueIndexing(fp); })
+            .on("unlink", (fp: string) => {
+                if (!shouldWatchFile(fp)) return;
+                this.db.clearChunks(fp);
+                this.scheduleSave();
+            });
+
+        return this.watcherReady;
+    }
 
     /** Queue a file for indexing (debounced). */
     private queueIndexing(filePath: string): void {
@@ -544,12 +548,12 @@ export class NrekiEngine {
 
     /** Stop the file watcher. */
     stopWatcher(): void {
-        if (this.watcher) {
-            this.watcher.close();
-            this.watcher = null;
-            this.watcherReady = null;
-        }
-    }
+        if (this.watcher) {
+            this.watcher.close();
+            this.watcher = null;
+            this.watcherReady = null;
+        }
+    }
 
     // ─── DB Delegation Wrappers (Stats & Tracking) ─────────────────
 

package/src/middleware/circuit-breaker.ts

@@ -17,6 +17,33 @@ import type { McpToolResponse } from "../router.js";
 /** Inactivity timeout for auto-reset (60 seconds). */
 const INACTIVITY_TIMEOUT_MS = 60_000;
 
+/**
+ * Read-only actions exempt from heuristic text-matching of error patterns
+ * via containsError(). These actions can legitimately return responses
+ * containing words like "error", "failed", "exception" (search hits on
+ * try/catch blocks, fast_grep over real source code, audit reports with
+ * "Type errors" frequency, agent's own memorize/engram annotations).
+ *
+ * Important: this skip applies ONLY to containsError() text heuristic.
+ * Explicit response.isError === true (hardware signal: path jail
+ * violation, security error, etc.) STILL trips the breaker for these
+ * actions, because that is a contract violation, not a false-positive.
+ *
+ * Excluded from this set: filter_output (it is the gateway for the
+ * agent to read Bash/test/build output - text matching of "FAIL"/"error"
+ * there is exactly the doom-loop signal the breaker must catch).
+ */
+const READ_ONLY_ACTIONS = new Set<string>([
+    // nreki_navigate
+    "search", "definition", "references", "outline", "map",
+    "fast_grep", "prepare_refactor", "orphan_oracle", "type_shape",
+    // nreki_code (read-only sub-actions; filter_output deliberately excluded)
+    "read", "compress",
+    // nreki_guard (state inspection + agent's own memory annotations)
+    "status", "report", "audit",
+    "pin", "unpin", "reset", "set_plan", "memorize", "engram",
+]);
+
 
 // ─── Level-Specific Payloads ─────────────────────────────────────────
 
@@ -182,7 +209,15 @@ export function wrapWithCircuitBreaker(
         // Record the result for pattern detection
         // Truncate to 2KB before hashing — terminal output can be megabytes
         const responseText = response.content.map(c => c.text ?? "").filter(Boolean).join("\n").slice(0, 2048);
-        const hasError = response.isError === true || containsError(responseText);
+        // Read-only actions skip the containsError() text heuristic to avoid
+        // false-positive trips on legitimate error mentions in source code,
+        // search results, audit reports, or agent's own memory annotations.
+        // Explicit response.isError === true (hardware signal) still trips
+        // the breaker for these actions: it is a contract violation, not a
+        // heuristic match.
+        const isReadOnly = READ_ONLY_ACTIONS.has(action);
+        const hasError = response.isError === true ||
+            (!isReadOnly && containsError(responseText));
 
         if (hasError) {
             const loopCheck = cb.recordToolCall(

package/src/parser.ts

@@ -398,13 +398,9 @@ export class ASTParser {
                     }
                 }
 
-                // Web symbol name normalization (D1)
+                // Web symbol name normalization (D1) — extracted to shared utility (v10.18.1)
                 if (symbolName) {
-                    if (ext === ".css") {
-                        symbolName = symbolName.replace(/[.#,]/g, " ").replace(/\s+/g, " ").trim();
-                    } else if (ext === ".json" || ext === ".html") {
-                        symbolName = symbolName.replace(/^["']|["']$/g, "").trim();
-                    }
+                    symbolName = normalizeWebSymbol(symbolName, ext);
                 }
 
                 const node = mainCapture.node;
@@ -574,3 +570,29 @@ export class ASTParser {
         return typeMap[captureName] ?? captureName;
     }
 }
+
+/**
+ * Normalize a web symbol (CSS selector, HTML attribute value, JSON key)
+ * to match the canonical form stored in ParsedChunk.symbolName.
+ *
+ * CSS: strips ".", "#", "," prefixes; collapses whitespace.
+ *   ".foo, #bar" → "foo bar"
+ * HTML/JSON: strips surrounding quotes.
+ *   '"key"' → "key"
+ * Other extensions: no-op (returns input unchanged).
+ *
+ * Used at the API boundary of semantic-edit + compressor-foveal to
+ * accept LLM-supplied symbols with web prefixes that the parser stripped
+ * during chunk creation. Without this normalization, ".foo" sent by an
+ * LLM caller fails to match a parser-stored chunk named "foo".
+ */
+export function normalizeWebSymbol(symbolName: string, ext: string): string {
+    if (!symbolName) return symbolName;
+    if (ext === ".css") {
+        return symbolName.replace(/[.#,]/g, " ").replace(/\s+/g, " ").trim();
+    }
+    if (ext === ".json" || ext === ".html") {
+        return symbolName.replace(/^["']|["']$/g, "").trim();
+    }
+    return symbolName;
+}

package/src/semantic-edit.ts

@@ -14,7 +14,7 @@
 
 import fs from "fs";
 import path from "path";
-import { ASTParser, type ParsedChunk } from "./parser.js";
+import { ASTParser, type ParsedChunk, normalizeWebSymbol } from "./parser.js";
 import { AstSandbox } from "./ast-sandbox.js";
 import { Embedder } from "./embedder.js";
 import { readSource } from "./utils/read-source.js";
@@ -458,11 +458,20 @@ export async function batchSemanticEdit(
     for (const [filePath, fileEdits] of editsByFile.entries()) {
         let virtualCode = vfs.get(filePath)!;
         const parseResult = await parser.parse(filePath, virtualCode);
+        const ext = path.extname(filePath).toLowerCase();
 
         // ─── PHASE A: Cluster edits by AST node identity (startIndex) ───
         const chunkGroups = new Map<number, { chunk: ParsedChunk; edits: BatchEditOp[] }>();
 
         for (const edit of fileEdits) {
+            // v10.18.1: normalize web symbols IN-PLACE so downstream
+            // oldRawCodes/newRawCodes keys (built from chunk.symbolName)
+            // align with handleBatchEdit's blast-radius lookup
+            // (built from edit.symbol). Without in-place mutation,
+            // signature changes on .css/.json/.html files would evade
+            // detection silently.
+            edit.symbol = normalizeWebSymbol(edit.symbol, ext);
+
             // AMBIGUITY CHECK: Reject if multiple symbols share the same name.
             const allMatches = parseResult.chunks.filter(c => {
                 const name = c.symbolName || extractName(c);
@@ -781,12 +790,12 @@ export async function batchSemanticEdit(
         }
     }
     if (!topologyChanged) {
-        for (const edit of edits) {
-            if (edit.mode && edit.mode !== "replace" && edit.mode !== "patch") continue;
-            const key = `${safePath(projectRoot, edit.path)}::${edit.symbol}`;
-            const oldRaw = oldRawCodes.get(key);
-            const newRaw = newRawCodes.get(key);
-            if (oldRaw && newRaw && detectSignatureChange(oldRaw, newRaw)) {
+        for (const edit of edits) {
+            if (edit.mode && edit.mode !== "replace" && edit.mode !== "patch") continue;
+            const key = `${safePath(projectRoot, edit.path)}::${edit.symbol}`;
+            const oldRaw = oldRawCodes.get(key);
+            const newRaw = newRawCodes.get(key);
+            if (oldRaw && newRaw && detectSignatureChange(oldRaw, newRaw)) {
                 topologyChanged = true;
                 break;
             }
@@ -862,6 +871,11 @@ export async function semanticEdit(
         };
     }
 
+    // v10.18.1: normalize web symbols at API boundary so LLM-supplied
+    // ".foo" / '"key"' match parser-normalized chunk.symbolName.
+    const ext = path.extname(filePath).toLowerCase();
+    symbolName = normalizeWebSymbol(symbolName, ext);
+
     // Find matching chunks by name (prefer AST symbolName, fallback to regex)
     const matches: Array<{ chunk: ParsedChunk; name: string }> = [];
     const allNames: string[] = [];
@@ -1123,7 +1137,6 @@ export async function semanticEdit(
     const newCodeTokens = Embedder.estimateTokens(newCode ?? spliceRes.newRawCode);
     const tokensAvoided = Math.max(0, fullFileTokens + symbolTokens - newCodeTokens);
 
-    const ext = path.extname(filePath).toLowerCase();
     let topologyChanged = false;
     if (rawCode && spliceRes.newRawCode && detectSignatureChange(rawCode, spliceRes.newRawCode)) {
         topologyChanged = true;

package/src/utils/to-posix.ts

@@ -1,13 +1,13 @@
-/**
- * to-posix.ts — Cross-platform path normalization
- *
- * Converts Windows backslash paths to forward-slash POSIX format.
- * Single source of truth — eliminates 4 duplicated copies across
- * the kernel, backends, and hologram modules.
- */
-
-import * as path from "path";
-
-export function toPosix(p: string): string {
-    return path.normalize(p).replace(/\\/g, "/");
-}
+/**
+ * to-posix.ts — Cross-platform path normalization
+ *
+ * Converts Windows backslash paths to forward-slash POSIX format.
+ * Single source of truth — eliminates 4 duplicated copies across
+ * the kernel, backends, and hologram modules.
+ */
+
+import * as path from "path";
+
+export function toPosix(p: string): string {
+    return path.normalize(p).replace(/\\/g, "/");
+}